Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes

Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes Mayara Montilha Cotrim, Cristiane Nunes Francisco, Pedro José Farias Fernandes This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9593796/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study aims to evaluate the relative contribution of spectral and geomorphometric attributes in the supervised classification of forest physiognomies in urban remnants of the Atlantic Forest, using machine learning algorithms. The performance of three algorithms—Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM)—was compared. The dataset consisted of twelve attributes, including spectral variables (bands and vegetation index) and non-spectral variables (slope and vegetation height), derived from very high spatial resolution data of forests located in the municipality of Niterói, part of the Rio de Janeiro Metropolitan Region, which encompasses approximately 50% of its territory. Three classification tests were conducted, and the Random Forest (RF) algorithm achieved the best performance, with an overall accuracy of 96% and F-score values above 90%. In comparison, the Support Vector Machine (SVM) obtained an overall accuracy of 88% and F-score values exceeding 74%. Non-spectral attributes, particularly elevation and vegetation height, showed the greatest importance in classification, with permutation indices of 0.48 and 0.12, respectively, while the remaining attributes scored below 0.04. The supervised classification distinguished three physiognomic classes: tall forest, medium forest, and low forest. Beyond demonstrating the superior performance of RF in classifying forest remnants, this study highlights the relevance of geomorphometric variables—especially elevation—in characterizing vegetation physiognomy, whether due to anthropogenic influence or the role of topography in plant physiology. Atlantic Forest Forest remnants Supervised classification Algorithms Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Remote sensing (RS) imagery constitutes an essential resource for vegetation mapping and monitoring. In addition to enabling the identification of vegetation structure [ 7 ] and species [ 53 ], RS facilitates the detection of anthropogenic impacts on vegetation physiognomy, as observed in forest remnants within urban areas [ 56 ]. Its main advantages include rapid and efficient analyses, reduced processing costs, and the ability to assess changes and dynamics of the studied objects [ 14 , 72 ]. Among the applications of RS in vegetation cover studies, classification stands out. Currently, this process has been increasingly performed using machine learning algorithms to automate mapping [ 46 ], generating pattern-recognition models based on spectral and non-spectral data [ 27 ]. Machine learning algorithms for land cover classification can be grouped into supervised, unsupervised, and semi-supervised techniques [ 46 ]. Supervised approaches rely on samples of predefined classes selected from imagery; unsupervised techniques employ clustering algorithms based on pixel value similarity [ 58 ]; while semi-supervised classification combines labeled and unlabeled data [ 44 , 27 ]. Supervised machine learning algorithms such as Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM) have demonstrated excellent performance in the analysis of complex remote sensing (RS) datasets [ 3 , 8 , 42 ]. A key advantage of these methods is that they are non-parametric, meaning they do not assume normality of input data nor make assumptions regarding data frequency distributions, which makes them widely applied in RS studies [ 8 ]. Nevertheless, few studies have explored the application of these models to very high spatial resolution data in forests located within urban areas, with the specific aim of comparing and analyzing their performance. High-resolution imagery in urban environments poses challenges for classification due to spatial and spectral diversity, the complexity of urban surfaces, and the presence of building shadows [ 32 , 52 ]. Research has shown that non-parametric machine learning algorithms are highly efficient for classification tasks. In a comprehensive review of supervised classification, Alshammari [ 4 ] reinforces this perspective by highlighting that non-parametric methods are particularly effective in capturing complex non-linear relationships, often achieving overall accuracy levels above 90% across multiple domains, thereby substantially reducing classification errors. Despite these results, it is important to emphasize that the main challenge in applying supervised classification lies in the unequal distribution of patterns among groups [ 25 , 20 ]. According to these authors, this occurs because, in certain classes, algorithms struggle to extract specific information—for instance, in distinguishing between different forest physiognomies. To improve classification performance, it is essential to apply predictive variables (attributes) that describe the characteristics of the data samples used in classification, thereby maximizing the probability of correct decision-making by the algorithm [ 76 , 48 ]. Among the spectral attributes that can positively influence the classification of forest remnants are vegetation indices, derived from differences in spectral responses between red and near-infrared bands. The Normalized Difference Vegetation Index (NDVI) is one of the most widely used [ 45 ], enabling the analysis of forest dynamics as well as monitoring vegetation density and health [ 5 ]. In addition to spectral imagery, remote sensing also provides Digital Elevation Models (DEMs), from which geomorphometric variables are derived. These variables are used to describe terrain shape and topography [ 65 ] and have contributed to successful outcomes [ 73 , 31 ] by incorporating relief-related attributes into image classification datasets. Slope, considered a fundamental geomorphometric attribute for terrain analysis, is one of the variables derived from DEMs. These attributes have become increasingly available due to advances in geotechnologies for collecting altimetric data, particularly the Light Detection and Ranging (LiDAR) system, which generates Digital Elevation Models (DEMs) through the emission of laser pulses by airborne and orbital sensors [ 19 ]. Geomorphometric attributes stand out for their importance in understanding the spatial distribution of forests, as they influence vegetation development by conditioning water distribution, solar incidence, and soil characteristics on slopes [ 1 , 9 ]. Their inclusion in classification datasets is therefore essential for vegetation cover studies. In addition to geomorphometric attributes, DEMs also provide key variables for vegetation characterization, such as the Digital Height Model (DHM). This represents the height of terrain features and, in the case of vegetation cover, the height of the vegetation itself, obtained by subtracting surface elevation data from ground elevation data [ 37 , 24 , 12 ]. In vegetation cover classification, the Digital Height Model (DHM) and geomorphometric attributes assist in discriminating different physiognomies, such as rupestrian vegetation located on steep slopes, grasslands originating from anthropogenic activities, and forests at different stages of regeneration. According to Hartling et al. [ 33 ], this approach corresponds to the fusion of optical and structural data and can improve classification in areas with vegetation of varying heights. Recent literature has shown successful applications involving LiDAR data and spectral imagery combined with machine learning algorithms for land cover classification [ 40 , 6 , 44 ]. This methodology, which integrates machine learning models with a database containing a vast number of attributes—referred to as Geospatial Big Data [ 43 ]—enhances classification and spatial analysis of forest remnants, particularly in urban areas, which are complex environments composed of diverse constructions and affected by shadows in high-resolution imagery [ 52 ]. Therefore, the present study aims to evaluate the relative contribution of spectral and geomorphometric attributes in the supervised classification of forest physiognomies in urban remnants of the Atlantic Forest, using machine learning algorithms and comparing the performance of different approaches. The study was conducted in the municipality of Niterói, part of the Rio de Janeiro Metropolitan Region (RMRJ), which has approximately half of its territory covered by Atlantic Forest remnants [ 51 ]. 2. Study area The forest remnants of the municipality of Niterói (Fig. 1 ), as in other cities within the Atlantic Forest biome, predominantly occur on slopes reaching up to 400 m in altitude. In areas with steeper slopes, near rocky outcrops, rupestrian vegetation is found, characterized by a herbaceous physiognomy. On the plains, coastal ecosystems are present, with diverse structural and floristic composition, such as restingas, floodplains, and mangroves [ 61 ]. The Atlantic Forest is associated with a humid tropical climate, characterized by average temperatures of 25°C [ 75 ] and high annual rainfall, exceeding 1,500 mm [ 64 ], without significant water deficit for plants, even during the drier winter months. As a result, tree species without drought protection predominate, forming a continuous, dense, and stratified canopy composed of the crowns of medium- and large-sized trees [ 54 , 55 ]. 3. Data and methods 3.1. Workflow of the Study The supervised classification was carried out over the area covered by Atlantic Forest remnants and their surroundings, with a 200 m buffer, resulting in a total area of 8,800 ha, equivalent to 65% of the municipality of Niterói. The remnants were delineated through unsupervised classification of a Planet orbital image, acquired in October 2021, using the K-means algorithm, which achieved the best performance among 48 experiments conducted by Cotrim et al. [ 21 ]. Among the freely available options, Planet images offer the highest spatial resolution and also enhanced spectral resolution in the visible range, which favors vegetation mapping and spatial analyses. Acquired by the constellation of 136 Dove satellites, they provide eight spectral bands (coastal blue, blue, green I, green, yellow, red edge (RE), red, and near-infrared (NIR)), a radiometric resolution of 12 bits, and a spatial resolution of 3 m. The images were downloaded from the Planet Explore platform, which provides free orthorectified images with atmospheric correction. This platform is part of Norway’s International Climate and Forests mission, which selects the best individual images to create a complete, continuous, and updated mosaic [ 60 ]. Images acquired near the southern hemisphere spring equinox were evaluated due to the reduced influence of topographic shadows, resulting from solar incidence on slopes, which facilitates target identification in rugged terrain. This period also coincides with the onset of the rainy season, aiding in the discrimination of vegetation physiognomy by capturing water stress still present in plants. Accordingly, the image selected was from October 23, 2021 [ 60 ]. To compose the geomorphometric dataset, the Digital Surface Model (DSM) and the Digital Terrain Model (DTM - Fig. 2 c) generated by LiDAR between October and November 2019 in Niterói were acquired. The DSM consists of data representing terrain altitude plus the height of objects on the Earth’s surface, corresponding to all laser return pulses, while the DTM is generated through the filtering process of the DSM using the last return signal of the laser pulse, which corresponds to the terrain surface [ 35 ]. The second stage consisted of processing the data collected previously, with the objective of generating additional attributes to compose the database used in the image classification experiments: the Digital Height Model (DHM), the Normalized Difference Vegetation Index (NDVI), and slope. The DHM (Fig. 2 a) was derived from LiDAR data by subtracting the DSM, which represents the top of terrain features, from the DTM, which represents the ground surface. The DHM was classified into four vegetation height classes: 0–0.5 m (grassland), 0.5–4 m (low vegetation), 4–12 m (medium vegetation), and > 12 m (tall vegetation) [ 39 , 16 ]. Based on the DTM, slope (Fig. 2 b) was generated, calculated from the planimetric distance and altimetric amplitude between two points, corresponding to the zenithal inclination angle of the terrain surface relative to the horizontal plane, with values ranging from 0° to 90° [ 74 ]. Six slope classes were established according to the IBGE classification [ 38 ]: 0–3% (flat); 3–8% (gently undulating); 8–20% (undulating); 20–45% (strongly undulating); 45–75% (mountainous); and > 75% (escarpment). Finally, NDVI (Fig. 2 d) was calculated as the normalized difference between the reflectance of the near-infrared and red bands [ 22 ]. The index ranges from − 1 to 1, with positive values corresponding to rocks, bare soil, and vegetation. Values closer to 1 indicate dense and healthy vegetation, while values near 0 represent terrain without vegetation cover. Negative values occur when targets correspond to water, clouds, or snow [ 22 ]. In the third stage, training samples were generated to train the supervised classification algorithm, according to user-defined classes, in the stage understood as machine learning [ 26 ]. A total of 74,723 samples (Fig. 3 ) were collected for eight land cover classes, randomly divided into two subsets by the classification algorithms, with 70% of the samples used for training and 30% for accuracy testing. Eight classes were defined, three corresponding to forests, one to grassland, one to wetlands, and the remaining to non-vegetation classes (Table 1 ). The distribution of training samples was carried out through the construction of polygons, which were subsequently converted into multiple points, using Planet imagery, Google Earth, the Digital Height Model (DHM), and field knowledge as the basis. Table 1 Land use and land cover classes defined for supervised classification. The table provides the description of each class and the number of reference samples used. Adapted from: MapBiomas [ 47 ], Jennings et al. [ 39 ], Junk et al. [ 41 ], and CONAMA Resolution No. 004/1994 [ 16 ]. Class Description N. of samples Rock Outcrop (RO) Exposed rock with or without rupestrian vegetation, located in shallow soils such as hilltops or steep slopes. 1,961 Tall Forest (TF) Predominantly arboreal vegetation with lower structural and floristic alteration. Height above 12 m, with emergent individuals exceeding 30 m, characterizing an advanced stage of regeneration. 20,253 Medium Forest (MF) Predominantly arboreal and shrub vegetation with variable floristic composition. Height between 4 and 12 m, characterizing an intermediate stage of regeneration. 16,189 Low Forest (LF) Herbaceous-shrub vegetation up to 4 m in height with an open canopy, allowing greater light penetration. Includes rupestrian vegetation—rupicolous plants on litholic soils or fixed on rocks.. 6,012 Grassland/Pasture (G/P) Herbaceous cover, predominantly grasses, for agricultural, recreational, or ornamental use, with height up to 50 cm. 3,792 Built-up Area (BU) Urban fabric composed of houses, buildings, multipurpose structures, roads, and other urban infrastructure 3,286 Water Body (WB) Natural water masses corresponding to coastal lagoons. 258 Wetland (WL) Herbs and shrubs—mangroves and marshes—adapted to waterlogged or poorly drained soils, originating from sedimentary deposits. Occur along lagoon margins or drainage channels in coastal areas.. 22,972 The database for classification was composed of twelve attributes, corresponding to the eight spectral bands of Planet imagery (seven visible bands and one near-infrared), two geomorphometric attributes (slope and DTM), and two vegetation attributes (DHM and NDVI). For such a large dataset, cloud-based data processing using programming languages is an important resource for vegetation cover classification. Thus, supervised classification was performed in the Google Colab environment, which includes the Python Scikit-Learn library, providing machine learning classification algorithms. The classification script was created by Fernandes and Sami [ 28 ] and is available at: https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9 . Random Forest (RF), developed by Breiman [ 10 ], is an ensemble learning algorithm in which multiple classifiers are combined to improve classification [ 66 , 10 ]. This algorithm creates decision trees from the training samples collected for data classification. The results generated by each tree are combined to obtain the final classification, increasing accuracy [ 71 , 10 ]. RF has shown excellent results for remote sensing data classification and is increasingly cited in specialized literature [ 2 , 57 , 8 ]. The K-Nearest Neighbors (KNN) algorithm, developed by Fix and Hodges [ 30 ], is one of the most widely used [ 27 ]. It predicts data based on a simple majority vote of the n nearest neighbors previously classified in an attribute space, following the principle that similar points belong to the same class. The number of neighbors and the distance measure between points around the pixel are the main parameters, along with similarity, to make predictions and assign the unclassified point to the attribute space of the class corresponding to the majority of its nearest pixels [ 67 , 27 ]. The Support Vector Machine (SVM) algorithm is a computational learning technique for pattern recognition problems introduced by Cortes and Vapnik [ 18 ]. It is a classifier that defines linear boundaries based on support vectors that maximize the distance between classes. However, since many classification problems are non-linear, kernel functions are used to transform the data into a higher-dimensional attribute space [ 27 ]. In this way, SVM maximizes the decision surface for class separation [ 50 , 18 ], showing good performance with remote sensing data [ 49 ]. For the configuration of hyperparameters related to the three machine learning algorithms used in this study, the default values of the Python Scikit-Learn library were adopted, as defined in Scikit Learn [ 67 , 68 , 69 ] and presented in Table 2 . Table 2 Hyperparameters of the algorithms used in supervised classification. Were: n_estimators – number of trees in the model, min_samples_leaf – minimum number of samples a leaf node must contain,min_samples_split – minimum number of samples required to split an internal node, random_state – seed for reproducibility, max_depth – maximum depth allowed for the trees, n_neighbors – number of neighbors considered for classification or regression, kernel – function used to map data into higher dimensions, C – regularization parameter controlling the penalty for misclassification Source: Python Scikit-Learn library. Model Hyperparameters Random Forest - RF n_estimators = 100 min_samples_leaf = 5 min_samples_split = 2 random_state = 42 max_depth = 80 K-nearest neighbour - KNN n_neighbors = 3 Support Vector Machine - SVM kernel = 'linear' C = 1.0 random_state = 42 The final stage consisted of evaluating the accuracy of the three classifications based on 30% of the collected samples, i.e., 22,417 points, using cross-validation [ 70 ]. In this way, the minimum requirement of 50 samples per class for maps with fewer than 12 classes, established by Congalton and Green [ 17 ], was exceeded. Error matrices were constructed with the following performance indicators: (1) overall accuracy – number of correctly classified samples relative to the total reference samples; (2) producer’s accuracy (precision) – number of correctly classified samples of class k relative to the total samples of that class; (3) user’s accuracy (recall) – number of correctly classified samples of class k relative to the total samples classified as that class; (4) F-score (f1-score) – harmonic mean between precision and recall [ 62 , 17 ]. Additionally, a ranking of permutation importance of the attributes used in the three classification experiments was generated. This metric evaluates the importance of attributes in a supervised learning model by measuring the impact each attribute has on model performance when its values are randomly perturbed while keeping the others fixed [ 10 , 29 ]. Finally, class balancing was applied to the classification with the best performance to assess its accuracy. For this, the Synthetic Minority Oversampling Technique (SMOTE) was used, which aims to create synthetic examples in the minority class of interest [ 13 ]. 4. Results and Discussion The results obtained in this study highlight, above all, the importance of geomorphometric attributes in the discrimination of vegetation classes, particularly the Digital Terrain Model (DTM), which showed the highest explanatory power among the variables analyzed, surpassing spectral attributes. Complementarily, the analysis of the performance of supervised classification algorithms demonstrated that Random Forest (RF) presented the best accuracy and consistency indicators across classes, while SVM and KNN showed limitations. In addition, the effects of data balancing and the individual contribution of attributes to classification were evaluated. Based on these findings, this section is organized into: (i) comparison of algorithm performance; (ii) influence of sample balancing; and (iii) importance of attributes in classification. 4.1 Comparison of Algorithm Performance In the assessment of the final results of the three experiments, the classification performed by the KNN algorithm presented a high level of “salt-and-pepper” noise; therefore, its accuracy was not analyzed. The land cover classification generated by the SVM algorithm achieved an overall accuracy of 87.67%. However, the values of producer’s accuracy (precision), user’s accuracy (recall), and F-score obtained for each class allow for a more precise and detailed assessment of the model’s performance. Producer’s accuracy indicates that classes such as WL, G/P, and BU performed exceptionally well, with values between 99% and 100%, whereas forest classes scored below 85%, suggesting greater confusion. User’s accuracy ranged from 82% to 100%, highlighting excellent performance in well-defined classes but lower performance in forest categories. Similarly, the F-score varied between 74% and 100%, reinforcing that, although the model is robust across several classes, it remains fragile in forest classes, where overlapping features lead to a higher number of errors. In summary, the three indicators together demonstrate that the model is globally effective but requires specific adjustments for classes with lower separability (Table 3 ). Table 3 Confusion matrix and derived accuracy metrics for land cover classification using the Support Vector Machine (SVM) model. The table presents reference samples, correctly classified instances, and precision, recall, and F-score for each class. Classes Reference samples AR G/P BU TF MF LF WB WL Tot.class C L A S S I F I E D AR 547 1 9 0 1 1 0 0 559 G/P 1 1,136 1 0 0 7 0 0 1,145 BU 16 6 920 0 0 1 2 0 946 TF 0 0 0 5,059 1,014 70 0 0 6,143 MF 1 0 0 990 3,709 179 0 0 4,879 LF 1 4 1 83 374 1,344 0 0 1,807 WB 0 0 0 0 0 0 66 0 66 WL 0 0 0 0 0 0 1 6,872 6,873 Total collected 566 1,147 931 6,132 5,098 1,602 69 6,872 22,417 Correctly classified 547 1,136 920 5,059 3,709 1,344 66 6,872 19,653 Precision 97% 99% 99% 83% 73% 84% 96% 100% 88% Recall 98% 99% 97% 82% 76% 74% 100% 100% F-score 97% 99% 98% 82% 74% 79% 98% 100% The land cover classification generated by the Random Forest (RF) algorithm (Table 4 ) achieved an overall accuracy of 95.67%. The indices of producer’s accuracy (precision), user’s accuracy (recall), and F-score demonstrate a significantly superior performance of RF compared to the previous matrix. Producer’s accuracy ranged from 89% to 100%, with classes such as G/P, WB, and WL reaching 100%, indicating that nearly all samples in these categories were correctly classified. User’s accuracy varied between 92% and 100%, confirming the reliability of the classifications. Likewise, the F-score, with values between 91% and 100%, highlighted the balance between producer’s and user’s accuracy, reinforcing that the model not only correctly identifies most instances but also adequately retrieves the relevant ones. Altogether, the three indicators reveal that RF achieved robust and consistent performance, with overall accuracy above 90%, substantially reducing errors observed in more problematic classes in the previous matrix, such as forest categories, and consolidating itself as a highly effective classifier for the dataset analyzed, particularly for classes with similar characteristics, such as forests. Table 4 Confusion matrix and derived accuracy metrics for land cover classification using the Random Forest (RF) model. The table presents reference samples, correctly classified instances, and precision, recall, and F-score for each class. Classes Reference samples AR G/P BU TF MF LF WB WL Tot.class C L A S S I F I E D AR 555 0 3 0 0 1 0 0 559 G/P 3 1,138 0 0 0 4 0 0 1,145 BU 3 1 939 1 0 1 0 0 945 TF 0 0 0 5,692 431 20 0 0 6,143 MF 0 0 0 333 4,526 20 0 0 4,879 LF 0 1 1 42 102 1,661 0 0 1,807 WB 0 0 0 0 0 0 62 4 66 WL 0 0 0 0 0 0 0 6,873 6,873 Total collected 561 1,140 943 6,068 5,059 1,707 62 6,877 22,417 Correctly classified 555 1,138 939 5,692 4,526 1,661 62 6,873 21,446 Precision 99% 100% 99% 94% 89% 97% 100% 100% 96% Recall 99% 99% 99% 93% 93% 92% 94% 100% F-score 99% 100% 99% 93% 91% 95% 97% 100% 4.2 Influence of Sample Balancing The performance of the Random Forest (RF) classifier with class balancing proved to be similar to that obtained without balancing. The SMOTE method resampled the number of samples to 16,000 per class, which caused the overall accuracy of the RF to decrease to 95.38%, representing a difference of only 0.29%. The lower value can be explained by the fact that SMOTE also increases the number of samples in classes with reduced area, which may introduce noise into the data. However, these effects were not sufficient to cause significant changes in overall accuracy. Thus, the results indicated that the model already exhibited strong performance even with imbalanced data, reinforcing the excellent predictive capacity of the RF model. Douzas et al. [ 23 ] also observed similar results, where overall accuracy decreased from 58.7% (without balancing) to 55.7% with SMOTE. 4.3 Importance of Attributes in Classification Regarding the importance of the twelve attributes used in classification, differences among the tested algorithms were also observed (Fig. 4 ). In the RF classification, DTM presented the highest permutation index, with a value of 0.48—four times greater than DHM, which had 0.12 and ranked second in importance. The remaining attributes—NDVI, slope, and spectral bands—showed indices below 0.04. Among these, the visible bands, with the exception of blue, exhibited the lowest values, all below 0.02. Sanlang et al. [ 59 ] also employed very high spatial resolution imagery and LiDAR data with the aim of mapping land cover and urban functional zones using machine learning algorithms. The authors demonstrated that the classification model with the best performance was the one that integrated all variables, and further emphasized the role of the RF algorithm and LiDAR-derived attributes in improving accuracy. Similarly, Jiang et al. [ 47 ] integrated multitemporal Sentinel-2 imagery (10 m) with LiDAR data to identify dominant tree species in the Shanghai region, using a hierarchical RF classification model. Their study demonstrated the feasibility of large-scale mapping of urban trees and highlighted that the inclusion of LiDAR-derived metrics was crucial for increasing classification accuracy in an urban area characterized by high heterogeneity of land cover classes. In the SVM classification, NDVI obtained the highest permutation importance index, with a value of 0.40, followed by DTM with approximately 0.34. The green, red, and near-infrared bands ranked next, with values ranging from 0.22 to 0.17, while DHM occupied the sixth position with a value of 0.16. The remaining attributes—including the blue band, other Planet spectral ranges, and slope—presented values below 0.13. These results demonstrate the importance of non-pure spectral attributes in forest classification, particularly DTM and DHM in both algorithms, as well as the vegetation spectral index (NDVI) in the SVM algorithm. DHM presented similar importance values in both classification algorithms, though in different positions—second in RF and sixth in SVM. In contrast, the spectral bands, including non-traditional ones such as coastal blue, green 1, yellow, and red-edge, ranked among the lowest positions in both algorithms. The importance of DTM in discriminating forest classes may be related to anthropogenic pressures and topographic conditions that influence the degree of conservation and the structural development of forest remnants [ 22 ]. In densely populated urban areas of the Atlantic Forest biome, which have historically undergone Brazilian economic cycles, such as in Niterói, remnants tend to concentrate in areas of difficult access, generally in rugged terrain and at higher altitudes, where human pressure is less intense. From a topographic perspective, steep upper slopes near the crystalline massif summits are characterized by shallow soils, which limit the development of large trees but favor vegetation conservation due to reduced anthropogenic impact. Conversely, lower slopes accumulate colluvial material and present deeper, less inclined soils, conditions that allow vegetation development. However, their proximity to human activities subjects them to greater pressure and risk of degradation. Thus, a contrast emerges: at higher altitudes, conservation is favored by distance from human occupation, although vegetation size is limited; at lower altitudes, forest development potential is greater, but proximity to anthropogenic activities compromises conservation [ 11 , 1 , 9 ]. Hurskainen et al. [ 23 ] and Pittman and Hu [ 56 ] highlighted the importance of auxiliary variables in land cover classification, particularly attributes such as elevation, NDVI, and slope. In the present study, the inclusion of these variables contributed significantly to the differentiation among high, medium, and low forest classes, especially in areas where spectral responses were similar across classes. The variable importance analysis indicated that topographic attributes played a key role in discriminating structural classes, suggesting that vegetation organization is strongly associated with environmental gradients controlled by relief. This behavior is consistent with the findings of Pittman and Hu [ 56 ], who demonstrated that geomorphometric variables can exhibit high predictive power by capturing variations in moisture, drainage, and solar exposure. In this context, areas located in different positions within the landscape tend to present variations in vegetation height and density, which explains the improvement in classification accuracy when these variables are incorporated. Thus, even without species-level distinction, it was possible to separate structural forest classes, evidencing that geomorphometry acts as an important conditioning factor of vegetation cover heterogeneity. The RF classification is presented in Fig. 5 . The forest classes totaled 6,131 ha, corresponding to 45.8% of the municipality of Niterói. The largest area was classified as high-canopy forest, representing 44%, followed by medium-canopy forest with 37%, and low-canopy forest with 19%. Thus, nearly 80% of the remnants exhibit a medium to advanced successional stage, which is a remarkably high index for urban areas within the Atlantic Forest biome. The grassland class, in turn, comprises only 517 ha, the smallest area among the four vegetation classes present on the slopes. On the other hand, the edges of the forest remnants are almost entirely surrounded by buildings, highlighting both their high vulnerability and the critical importance of forest protection policies. 5. Conclusion This study clearly demonstrated that geomorphometric variables, particularly elevation (DTM) and vegetation height (DHM), play a central role in the classification of forest physiognomy in urban environments with rugged terrain, surpassing the contribution of traditional spectral attributes. DTM obtained the highest permutation importance value in RF and ranked among the top three variables in SVM. DHM also stood out in both classifications with similar importance values. Conversely, although NDVI was prominent in SVM, isolated spectral bands—especially those in the visible range—consistently ranked lowest in importance. Furthermore, it was observed that the classification produced by the RF algorithm yielded the highest accuracy indicators, with overall accuracy above 90% and, most notably, excellent results for producer’s accuracy (precision), user’s accuracy (recall), and F-score. This indicates superior performance compared to the SVM algorithm, in addition to presenting fewer salt-and-pepper type noise artifacts. The RF classification also achieved higher producer’s accuracy, user’s accuracy, and F-score values than SVM for forest classes, highlighting its improved ability to separate classes with similar characteristics. Taken together, these results emphasize the relevance of geomorphometric variables in explaining the structure and distribution of forest remnants in urban areas, influencing both their development and conservation status. The proposed methodology contributes to advancing vegetation mapping in complex environments and underscores the importance of integrating geomorphometric variables into remote sensing studies. This approach can be applied across different contexts to enhance understanding of spatial patterns and landscape transformation processes, providing valuable support for conservation strategies and sustainable management. CRediT authorship contribution statement M. M. Cotrim : Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Conceptualization, Investigation. C. N. Francisco : Writing – review & editing, Writing – original draft, Formal analysis, Data curation, Supervision, Project administration. P. J. F. Fernandes : Writing – review & editing, Supervision, Formal analysis. Declarations CRediT authorship contribution statement M. M. Cotrim: Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Conceptualization, Investigation. C. N. Francisco: Writing – review & editing, Writing – original draft, Formal analysis, Data curation, Supervision, Project administration. P. J. F. Fernandes: Writing – review & editing, Supervision, Formal analysis. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – Finance Code 001. The Article Processing Charge (APC) for the publication of this research was funded by the Coordination for the Improvement of Higher Education Personnel – CAPES (Research Organization Registry identifier ROR: 00x0ma614). For the purposes of open access, the authors have assigned the Creative Commons CC BY license to any accepted version of the article. Data availability Data and code necessary to replicate this research can be found in:https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9. References Abdalla Ldos, Oliveira S, de Volotão LC S (2015) Modelagem dinâmica da vegetação baseada em autômatos celulares: um estudo de caso da regeneração da Mata Atlântica no distrito de Aldeia Velha - RJ. Rev Bras Cartogr 67(3) Alba E et al (2022) Comparação entre algoritmos de aprendizado de máquina para a identificação de floresta tropical sazonalmente seca. Anuário Inst Geociências 45 Alimjan G, Sun T, Liang Y, Jumahun H, Guan Y (2018) A new technique for remote sensing image classification based on combinatorial algorithm of SVM and KNN. Int J Pattern Recognit Artif Intell 32(07):1859012. https://doi.org/10.1142/S0218001418590127 (doi.orgin Bing) Alshammari AH, Bencsik G, Ali AH (2026) A survey of six classical classifiers, including algorithms, methodological characteristics, foundational variants, and recent advances. Algorithms 19(1):37 Anwer HA, Mohamed T, Hassan A (2025) Assessing vegetation dynamics in Al Jazirah, Sudan using NDVI-based remote sensing techniques. J Saudi Soc Agric Sci 24(18). https://doi.org/10.1007/s44447-025-00011-0 (doi.orgin Bing) Aziz G, Minallah N, Saeed A, Khan MA, Tariq U, Khan MA (2024) Remote sensing based forest cover classification using machine learning. Sci Rep 14:69. https://doi.org/10.1038/s41598-023-50863-1 (doi.orgin Bing) Baines O, Wilkes P, Disney M (2020) Quantifying urban forest structure with open-access remote sensing data sets. Urban Urban Green 50:126653. https://doi.org/10.1016/j.ufug.2020.126653 (doi.orgin Bing) Belgiu M, Drăguţ L (2016) Random forest in remote sensing: a review of applications and future directions. ISPRS J Photogramm Remote Sens 114:24–31. https://doi.org/10.1016/j.isprsjprs.2016.01.011 (doi.orgin Bing) Bispo PC et al (2012) Effects of the geomorphometric characteristics of the local terrain on floristic composition in the central Brazilian Amazon. Austral Ecol 37:491–499 Breiman L (2001) Random forests. Mach Learn 45(1):5–32 Burak DL, Silveira TCD, Machado LV, Mendonça EDS (2023) Soil biological properties along a topographic gradient in Brazil’s Atlantic Forest biome. Diversity 15(10):1097. https://doi.org/10.3390/d15101097 Castro FC, Centeno TM (2005) Segmentação de imagens geradas por perfilamento a laser para delimitação de árvores individuais em uma área de reflorestamento de eucaliptos. XII Simpósio Brasileiro de Sensoriamento Remoto, Goiânia, Brasil, pp 16–21. abril 2005 Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: synthetic minority over-sampling technique. J Artif Intell Res 16:321–357 Chen S, Lei F, Zang Z, Zhang M (2023) Forest mapping using a VGG16-UNet++ & stacking model based on Google Earth Engine in the urban area. IEEE Geosci Remote Sens Lett 20:1–5 Art. 2502205. https://doi.org/10.1109/LGRS.2023.3274128 Cohen J (1960) A coefficient of agreement for nominal scales. Educ Psychol Meas 20(1):37–46 Conselho Nacional do Meio Ambiente (CONAMA) (1994) Resolução nº 006, de 4 de maio de 1994. Disposições gerais para a classificação de áreas de preservação ambiental e unidades de conservação, e diretrizes para a elaboração de programas de conservação e manejo dessas áreas. Brasília: CONAMA. https://www.ibama.gov.br/sophia/cnia/legislacao/MMA/RE0006-040594.PDF (ibama.gov.br in Bing) Congalton R, Green K (2009) Assessing the accuracy of remotely sensed data: principles and practices. CRC/Taylor & Francis, Boca Raton, p 183 Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297 Corte APD, Broadbent EN, Santos LHO (2022) Aplicações do LiDAR para o inventário florestal – enfoque unidade de área. São Paulo: AT. ISBN 978-65-258-0122-3. https://doi.org/10.22533/at.ed.223222904 (doi.orgin Bing) Cusworth S, Gkoutos GV, Acharjee A (2024) A novel generative adversarial networks modelling for the class imbalance problem in high dimensional omics data. BMC Med Inf Decis Mak 24 Cotrim MM, Francisco CN, Fernandes PJF (no prelo) Avaliação do potencial de algoritmos não supervisionados para discriminação de fisionomias florestais Deering DW, Rouse JW, Haas RH, Schell JA (1975) Measuring forage production of grazing units from Landsat MSS data. Int Symp Remote Sens Environ 10 Ann Arbor Proc Vol 2:1169–1178 Douzas G et al (2019) Imbalanced learning in land cover classification: improving minority classes’ prediction accuracy using the geometric SMOTE algorithm. Remote Sens 11(24):3040. https://doi.org/10.3390/rs11243040 Dubayah R, Knox R, Hofton M, Blair JB, Drake J (2000) Land surface characterization using LiDAR remote sensing. In: Hill M, Aspinall R (eds) Spatial Information for Land Use Management. International Publishers Direct, Singapore Duda RO, Hart PE, Stork DG (2000) Pattern classification, 2nd edn. Wiley-Interscience, New York Erlhal GJ et al (1991) Um sistema de segmentação e classificação de imagem de satélite: aplicações em meteorologia e geoprocessamento. Anais, São José dos Campos, SP Faceli K et al (2021) Inteligência artificial: uma abordagem de aprendizado de máquina. Blucher, São Paulo Fernandes PJF, Sami JA (2023) Script para classificação de remanescentes florestais. Disponível em: https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9 (acrobat.adobe.com in Bing) Fisher A, Rudin C, Dominici F (2019) All models are wrong, but many are useful: learning a variable’s importance by studying an entire class of prediction models simultaneously. arXiv. http://arxiv.org/abs/1801.01489 Fix E, Hodges JL Jr (1951) Discriminatory analysis, nonparametric discrimination. USAF School of Aviation Medicine, Randolph Field, Tex., Project 21-49-004, Rept 4, AF41(128)-31 Gezahagn R, Mengistu F (2025) Morphometric and land use land cover analysis for the management of water resources in Guder sub-basin. Ethiopia Appl Water Sci 15:18. https://doi.org/10.1007/s13201-024-02325-w (doi.orgin Bing) Hao X et al (2023) Urban land cover classification using high-resolution remote sensing imagery: challenges and methods. Remote Sens 15(6):1252 Hartling S, Sagan V, Sidike P, Maimaitijiang M, Carron J (2019) Urban tree species classification using WorldView-2/3 and LiDAR data fusion approach and deep learning. Sensors 19(6):1284. https://doi.org/10.3390/s19061284 Huang R, Yao W, Xu Z, Cao L, Shen X (2022) Information fusion approach for biomass estimation in a plateau mountainous forest using a synergistic system comprising UAS-based digital camera and LiDAR. Comput Electron Agric 198 HUB SIGeo (2023) Disponível em: https://www.sigeo.niteroi.rj.gov.br/ (sigeo.niteroi.rj.gov.br in Bing) Hurskainen P, Adhikari H, Siljander M, Pellinen J, Heiskanen J (2019) Auxiliary datasets improve accuracy of object-based land use/land cover classification in heterogeneous savanna landscapes. Remote Sens Environ 233:111354. https://doi.org/10.1016/j.rse.2019.111354 (doi.orgin Bing) Hyyppä J, Kelle O, Lehikoinen M, Inkinen M (2001) A segmentation-based method to retrieve stem volume estimates from 3-D tree height models produced by laser scanners. IEEE Trans Geosci Remote Sens 39(5):969–975 IBGE (2009) Manual técnico de geomorfologia. 2ª ed. IBGE, Rio de Janeiro Jennings MD et al (2006) Description, documentation, and evaluation of associations and alliances within the U.S. national vegetation classification, version 4.5. Ecological Society of America, Vegetation Classification Panel, Washington, DC, 119 p Jiang Y, Li X, Peng L, Li C, Song T (2025) Assessing the potential of multi-seasonal Sentinel-2 satellite imagery combined with airborne LiDAR for urban tree species identification. Sci Rep 15:25107. https://doi.org/10.1038/s41598-025-10971-6 (doi.orgin Bing) Junk WJ, Bayley PB, Sparks RE (1989) The flood pulse concept in river-floodplain. Oil Fat Ind 7(7):249–250 Li C, Jie W, Lei W, Luanyun H, Peng G (2014) Comparison of classification algorithms and training sample sizes in urban land classification with Landsat Thematic Mapper imagery. Remote Sens 6(2):964–983. https://doi.org/10.3390/rs6020964 Li S, Dragicevic S, Castro FA, Sester M, Winter S, Coltekin A, Pettit C, Jiang B, Haworth J, Stein A, Cheng T (2016) Geospatial big data handling theory and methods: a review and research challenges. ISPRS J Photogramm Remote Sens 115:119–133 Li Y, Li Z, Wang Z, Wang K, Yu Q (2023) Semi-supervised remote sensing image scene classification with prototype-based consistency. J Remote Sens 16:1023 Li M, Cao S, Zhu Z, Wang Z, Myneni RB, Piao S (2023b) Earth Syst Sci Data 15:4181–4203. https://doi.org/10.5194/essd-15-4181-2023 (doi.orgin Bing) Spatiotemporally consistent global dataset of the GIMMS Normalized Difference Vegetation Index (PKU GIMMS NDVI) from 1982 to 2022 Macarringue T et al (2022) Developments in land use and land cover classification techniques in remote sensing: a review. J Geogr Inf Syst 14(3):123–145 MapBiomas B (2024) Legenda da Coleção 9: descrição das classes. Disponível em: https://brasil.mapbiomas.org/wp-content/uploads/sites/4/2024/08/Legenda-Colecao-9-Descricao-Detalhada-PDF-PT.pdf (brasil.mapbiomas.org in Bing) Maulani Y, Surendro K (2025) Detailed land use classification model based on vegetation indices and texture features. Remote Sens Appl Soc Environ 40:101786. https://doi.org/10.1016/j.rsase.2025.101786 (doi.orgin Bing) Maulik U, Chakraborty D (2017) Remote sensing image classification: a survey of support-vector-machine-based advanced techniques. IEEE Geosci Remote Sens Mag 5(1):33–52. https://doi.org/10.1109/MGRS.2016.2641240 (doi.orgin Bing) Nascimento RFF, Alcântara EH, Kampel M, Stech JL, Novo EMLM, Fonseca LMG (2009) O algoritmo support vector machines (SVM): avaliação da separação ótima de classes em imagens CCD-CBERS-2. Anais, pp 2079–2086 Neto D, Seabra V, Correia M (2015) Análise da distribuição da ocupação urbana no município de Niterói a partir do mapeamento de uso e cobertura da terra. UERJ/FFP, São Gonçalo, RJ Neyns R, Canters F (2022) Mapping of urban vegetation with high-resolution remote sensing: a review. Remote Sens 14(4):1031–1059. https://doi.org/10.3390/rs14041031 Ocón JP, Stavros EN, Steinberg SJ, Robertson J, Gillespie TW (2024) Remote sensing approaches to identify trees to species-level in the urban forest: a review. Prog Phys Geogr Earth Environ 48(3):438–453. https://doi.org/10.1177/03091333241252520 (doi.orgin Bing) Oliveira-Filho AT, Fontes MAL (2000) Patterns of floristic differentiation among Atlantic forests in southeastern Brazil and the influence of climate. Biotropica 32(4):793–810. https://doi.org/10.1111/j.1744-7429.2000.tb00619.x (doi.orgin Bing) Oliveira-Filho AT, Tameirão Neto E, Carvalho WAC, Werneck M, Brina AE, Vidal CV, Rezende SC, Pereira JAA (2005) Análise florística do compartimento arbóreo de áreas de Floresta Atlântica sensu lato na região das Bacias do Leste (Bahia, Minas Gerais, Espírito Santo e Rio de Janeiro). Rodriguésia 56(87):185–235. https://doi.org/10.1590/2175-78602005568715 (doi.orgin Bing) Oppong J, Namwamba JB, Twumasi YA, Ning ZH, Asare-Ansah AB, Akinrinwoye C, Annan J (2023) Urbanization and urban forest loss: a spatial analysis of five metropolitan districts in Ghana. Geol Ecol Landsc 1–10. https://doi.org/10.1080/24749508.2023.2202439 (doi.orgin Bing) Phan TN, Kuch V, Lehnert LW (2020) Land cover classification using Google Earth Engine and random forest classifier—the role of image composition. Remote Sens 12(15):2411 Phiri D, Morgenroth J (2017) Developments in Landsat land cover classification methods: a review. Remote Sens 12(20):3340 Pittman RC, Hu B (2023) Contribution of topographic features and categorization uncertainty for a tree species classification in the boreal biome of Northern Ontario. GIScience Remote Sens 60(1):2214994. https://doi.org/10.1080/15481603.2023.2214994 (doi.orgin Bing) Planet (2026) Understanding PlanetScope instruments. Disponível em: https://developers.planet.com/docs/apis/data/sensors/ (developers.planet.com in Bing) PMN (2018) Atlas das unidades de conservação do município de Niterói. Coordenação técnica: Amanda Jevaux da Silva Sousa. Edição Pedro Bittencourt, Niterói, RJ Rossiter DG (2004) Technical note: statistical methods for accuracy assessment of classified thematic maps. Int Inst Geo-inf Sci Earth Obs (ITC), Enschede Sanlang S, Cao S, Du M, Mo Y, Chen Q, He W (2021) Integrating aerial LiDAR and very-high-resolution images for urban functional zone mapping. Remote Sens 13(13):2573. https://doi.org/10.3390/rs13132573 Sobral BS, Oliveira-Júnior JF, Gois G, Terassi PMB, Muniz-Júnior JGR (2018) Variabilidade espaço-temporal e interanual da chuva no estado do Rio de Janeiro. Rev Bras Climatol 22:281–308 Schmidt K, Dikau R (1999) Landform classification of the eastern Andes of Ecuador based on a digital elevation model. Geomorphology 28(1–2):1–15 Scikit-learn (2024a) Ensemble methods. Disponível em: https://scikit-learn.org/0.16/modules/ensemble.html (scikit-learn.org in Bing) Scikit-learn (2024b) KNeighborsClassifier. Disponível em: https:// scikit-learn.org/dev/modules/generated/sklearn.neighbors.KNeighborsClassifier.html#sklearn.neighbors.KNeighborsClassifier (scikit-learn.org in Bing) Scikit-learn (2024c) RandomForestClassifier. Disponível em: https:// scikit-learn.org/1.5/modules/generated/sklearn.ensemble.RandomForestClassifier.html (scikit-learn.org in Bing) Scikit-learn (2024d) SVC. Disponível em: https:// scikit-learn.org/1.5/modules/generated/sklearn.svm.SVC.html#sklearn.svm.SVC (scikit-learn.org in Bing) Scikit-learn (2024e) Cross-validation: evaluating estimator performance. Disponível em: https://scikit-learn.org/stable/modules/cross_validation.html (scikit-learn.org in Bing) Sega M, Xiao Y (2011) Multivariate random forests. Wiley Interdiscip Rev Data Min Knowl Discov 1(1):80–87 Seabra VS (2012) Análise da paisagem em apoio aos estudos de favorabilidade à recuperação florestal na bacia hidrográfica do rio São João. Tese de Doutorado, PPGG-UFRJ, Rio de Janeiro, p 189 Sovann C et al (2025) Land cover classification of tropical land cover mapping. Remote Sens 17(9):1551. https://doi.org/10.3390/rs17091551 Valeriano MM (2008) TOPODATA: guia para utilização de dados geomorfométricos locais. Relatório Técnico INPE-15318-RPQ/818, INPE, São José dos Campos Veloso HP et al (1991) Classificação da vegetação brasileira, adaptada a um sistema universal. Ministério da Economia, Fazenda e Planejamento; IBGE, Rio de Janeiro Valkenborg D et al (2023) Supervised learning. Am J Orthod Dentofac Orthop 164(1):146–149 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9593796","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":635053607,"identity":"516a0a6e-973c-4d0f-b0fe-8d67f9a6bd90","order_by":0,"name":"Mayara Montilha Cotrim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYHADxsbDDAw2PAwHGBg+VIBIgoCNsQGoJQ2khXHGGeK0MDAcBiNCWnTbzxh+rqjZJmcu39xwuKDmvAzfAR7DhgMMd/JxaTE7k2MseebYbWPLNqDDZhy7zSMJ0fLMsgGXlgM5BpINbLcTNxwDauFhu81jcIDH/PEHhsMGOG05/8b4Z8M/mJZ/50BaQLbg0XIjx0yysQ2qhbftADFanpVZNvbdNjY4lgjU0pfMI3mYrbDhgMEzPA5L3nyz4dttOYPDxx8+5vlmZ893vHljw4GKOzi1MDBwoMsxgwg8GhgY2B/gkx0Fo2AUjIJRwMAAABK4Z0NVYhb6AAAAAElFTkSuQmCC","orcid":"","institution":"Fluminense Federal University","correspondingAuthor":true,"prefix":"","firstName":"Mayara","middleName":"Montilha","lastName":"Cotrim","suffix":""},{"id":635053608,"identity":"ab23ff14-4d27-4945-b9b3-2f1721c781e2","order_by":1,"name":"Cristiane Nunes Francisco","email":"","orcid":"","institution":"Federal Fluminense University","correspondingAuthor":false,"prefix":"","firstName":"Cristiane","middleName":"Nunes","lastName":"Francisco","suffix":""},{"id":635053612,"identity":"81fab888-20df-4ffc-b7aa-2ffd5f412564","order_by":2,"name":"Pedro José Farias Fernandes","email":"","orcid":"","institution":"Federal Fluminense University","correspondingAuthor":false,"prefix":"","firstName":"Pedro","middleName":"José Farias","lastName":"Fernandes","suffix":""}],"badges":[],"createdAt":"2026-05-02 12:38:44","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9593796/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9593796/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108565219,"identity":"43024d4a-6546-4df2-ac1c-8cc2bba6aa75","added_by":"auto","created_at":"2026-05-06 04:25:04","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":598278,"visible":true,"origin":"","legend":"\u003cp\u003eMap of the study area: Niterói, Rio de Janeiro, Brazil. Sources for the basemap include IBGE (2022) and PlanetScope (2021)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/0770a599f81410add77d0368.png"},{"id":108565195,"identity":"ca629040-d9f0-43a1-aa1f-9801e73daedc","added_by":"auto","created_at":"2026-05-06 04:24:58","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":851284,"visible":true,"origin":"","legend":"\u003cp\u003eMaps of the study area attributes used in the classification of urban forest physiognomy: (a) vegetation height (m); (b) slope classes; (c) elevation (m); and (d) normalized difference vegetation index (NDVI). Data sources include HUB SiGeo (2019) and PlanetScope (2021)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/795ea263155c8115ccd09564.png"},{"id":108565194,"identity":"13ba5880-9b97-4535-a9ea-05be0c3e68da","added_by":"auto","created_at":"2026-05-06 04:24:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":874077,"visible":true,"origin":"","legend":"\u003cp\u003eMap of the location of training and validation samples. Source: PlanetScope (2021)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/54dd1ae6cecde98cf6790f28.png"},{"id":108565210,"identity":"d2311a1f-8593-4f6c-85a2-3855c5d98fec","added_by":"auto","created_at":"2026-05-06 04:25:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":281445,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of variable importance by \u0026nbsp;permutation between Random Forest (RF) and Support Vector Machine (SVM) \u0026nbsp;models. In the RF model, DTM shows the highest importance (≈0.43), followed by \u0026nbsp;DHM (≈0.10), while all other variables remain close to zero. In contrast, the \u0026nbsp;SVM model emphasizes NDVI (≈0.40) and DTM (≈0.35), with other spectral bands \u0026nbsp;and topographic variables contributing moderately (≈0.05–0.15). This \u0026nbsp;comparison illustrates how different algorithms prioritize environmental and \u0026nbsp;spectral predictors in land cover classification\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/ba2f714f3e6ed7ac0a4c5c9c.png"},{"id":108565209,"identity":"c3d53193-0d57-4d7b-b114-27b04343e958","added_by":"auto","created_at":"2026-05-06 04:25:01","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":680264,"visible":true,"origin":"","legend":"\u003cp\u003eLand use and land cover classification map of the municipality of Niterói (RJ), produced with the Random Forest algorithm using Planet spectral data (October 2021), DTM (2019), slope, DHM, and NDVI. The identified classes include rocky outcrops, grasslands, building areas, different forest types, water bodies, and wetlands\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/a819a1f38cdd19a7708dfa26.png"},{"id":109067565,"identity":"17dd90e7-f74f-4264-9d6d-14fb214e3250","added_by":"auto","created_at":"2026-05-12 09:56:17","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3838305,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9593796/v1/e4b1e0f4-259c-4129-8e58-69e8b0d5f7a8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eRemote sensing (RS) imagery constitutes an essential resource for vegetation mapping and monitoring. In addition to enabling the identification of vegetation structure [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] and species [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e], RS facilitates the detection of anthropogenic impacts on vegetation physiognomy, as observed in forest remnants within urban areas [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e]. Its main advantages include rapid and efficient analyses, reduced processing costs, and the ability to assess changes and dynamics of the studied objects [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAmong the applications of RS in vegetation cover studies, classification stands out. Currently, this process has been increasingly performed using machine learning algorithms to automate mapping [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e], generating pattern-recognition models based on spectral and non-spectral data [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning algorithms for land cover classification can be grouped into supervised, unsupervised, and semi-supervised techniques [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Supervised approaches rely on samples of predefined classes selected from imagery; unsupervised techniques employ clustering algorithms based on pixel value similarity [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e]; while semi-supervised classification combines labeled and unlabeled data [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSupervised machine learning algorithms such as Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM) have demonstrated excellent performance in the analysis of complex remote sensing (RS) datasets [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. A key advantage of these methods is that they are non-parametric, meaning they do not assume normality of input data nor make assumptions regarding data frequency distributions, which makes them widely applied in RS studies [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eNevertheless, few studies have explored the application of these models to very high spatial resolution data in forests located within urban areas, with the specific aim of comparing and analyzing their performance. High-resolution imagery in urban environments poses challenges for classification due to spatial and spectral diversity, the complexity of urban surfaces, and the presence of building shadows [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eResearch has shown that non-parametric machine learning algorithms are highly efficient for classification tasks. In a comprehensive review of supervised classification, Alshammari [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] reinforces this perspective by highlighting that non-parametric methods are particularly effective in capturing complex non-linear relationships, often achieving overall accuracy levels above 90% across multiple domains, thereby substantially reducing classification errors.\u003c/p\u003e \u003cp\u003eDespite these results, it is important to emphasize that the main challenge in applying supervised classification lies in the unequal distribution of patterns among groups [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. According to these authors, this occurs because, in certain classes, algorithms struggle to extract specific information\u0026mdash;for instance, in distinguishing between different forest physiognomies.\u003c/p\u003e \u003cp\u003eTo improve classification performance, it is essential to apply predictive variables (attributes) that describe the characteristics of the data samples used in classification, thereby maximizing the probability of correct decision-making by the algorithm [\u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e76\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. Among the spectral attributes that can positively influence the classification of forest remnants are vegetation indices, derived from differences in spectral responses between red and near-infrared bands. The Normalized Difference Vegetation Index (NDVI) is one of the most widely used [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e], enabling the analysis of forest dynamics as well as monitoring vegetation density and health [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn addition to spectral imagery, remote sensing also provides Digital Elevation Models (DEMs), from which geomorphometric variables are derived. These variables are used to describe terrain shape and topography [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e] and have contributed to successful outcomes [\u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e73\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] by incorporating relief-related attributes into image classification datasets. Slope, considered a fundamental geomorphometric attribute for terrain analysis, is one of the variables derived from DEMs.\u003c/p\u003e \u003cp\u003eThese attributes have become increasingly available due to advances in geotechnologies for collecting altimetric data, particularly the Light Detection and Ranging (LiDAR) system, which generates Digital Elevation Models (DEMs) through the emission of laser pulses by airborne and orbital sensors [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eGeomorphometric attributes stand out for their importance in understanding the spatial distribution of forests, as they influence vegetation development by conditioning water distribution, solar incidence, and soil characteristics on slopes [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Their inclusion in classification datasets is therefore essential for vegetation cover studies.\u003c/p\u003e \u003cp\u003eIn addition to geomorphometric attributes, DEMs also provide key variables for vegetation characterization, such as the Digital Height Model (DHM). This represents the height of terrain features and, in the case of vegetation cover, the height of the vegetation itself, obtained by subtracting surface elevation data from ground elevation data [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn vegetation cover classification, the Digital Height Model (DHM) and geomorphometric attributes assist in discriminating different physiognomies, such as rupestrian vegetation located on steep slopes, grasslands originating from anthropogenic activities, and forests at different stages of regeneration. According to Hartling et al. [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], this approach corresponds to the fusion of optical and structural data and can improve classification in areas with vegetation of varying heights.\u003c/p\u003e \u003cp\u003eRecent literature has shown successful applications involving LiDAR data and spectral imagery combined with machine learning algorithms for land cover classification [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. This methodology, which integrates machine learning models with a database containing a vast number of attributes\u0026mdash;referred to as Geospatial Big Data [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]\u0026mdash;enhances classification and spatial analysis of forest remnants, particularly in urban areas, which are complex environments composed of diverse constructions and affected by shadows in high-resolution imagery [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTherefore, the present study aims to evaluate the relative contribution of spectral and geomorphometric attributes in the supervised classification of forest physiognomies in urban remnants of the Atlantic Forest, using machine learning algorithms and comparing the performance of different approaches. The study was conducted in the municipality of Niter\u0026oacute;i, part of the Rio de Janeiro Metropolitan Region (RMRJ), which has approximately half of its territory covered by Atlantic Forest remnants [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e].\u003c/p\u003e"},{"header":"2. Study area","content":"\u003cp\u003eThe forest remnants of the municipality of Niter\u0026oacute;i (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), as in other cities within the Atlantic Forest biome, predominantly occur on slopes reaching up to 400 m in altitude. In areas with steeper slopes, near rocky outcrops, rupestrian vegetation is found, characterized by a herbaceous physiognomy. On the plains, coastal ecosystems are present, with diverse structural and floristic composition, such as restingas, floodplains, and mangroves [\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e]. The Atlantic Forest is associated with a humid tropical climate, characterized by average temperatures of 25\u0026deg;C [\u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e75\u003c/span\u003e] and high annual rainfall, exceeding 1,500 mm [\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e], without significant water deficit for plants, even during the drier winter months. As a result, tree species without drought protection predominate, forming a continuous, dense, and stratified canopy composed of the crowns of medium- and large-sized trees [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e, \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"3. Data and methods","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Workflow of the Study\u003c/h2\u003e \u003cp\u003eThe supervised classification was carried out over the area covered by Atlantic Forest remnants and their surroundings, with a 200 m buffer, resulting in a total area of 8,800 ha, equivalent to 65% of the municipality of Niter\u0026oacute;i. The remnants were delineated through unsupervised classification of a Planet orbital image, acquired in October 2021, using the K-means algorithm, which achieved the best performance among 48 experiments conducted by Cotrim et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAmong the freely available options, Planet images offer the highest spatial resolution and also enhanced spectral resolution in the visible range, which favors vegetation mapping and spatial analyses. Acquired by the constellation of 136 Dove satellites, they provide eight spectral bands (coastal blue, blue, green I, green, yellow, red edge (RE), red, and near-infrared (NIR)), a radiometric resolution of 12 bits, and a spatial resolution of 3 m. The images were downloaded from the Planet Explore platform, which provides free orthorectified images with atmospheric correction. This platform is part of Norway\u0026rsquo;s International Climate and Forests mission, which selects the best individual images to create a complete, continuous, and updated mosaic [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eImages acquired near the southern hemisphere spring equinox were evaluated due to the reduced influence of topographic shadows, resulting from solar incidence on slopes, which facilitates target identification in rugged terrain. This period also coincides with the onset of the rainy season, aiding in the discrimination of vegetation physiognomy by capturing water stress still present in plants. Accordingly, the image selected was from October 23, 2021 [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo compose the geomorphometric dataset, the Digital Surface Model (DSM) and the Digital Terrain Model (DTM - Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec) generated by LiDAR between October and November 2019 in Niter\u0026oacute;i were acquired. The DSM consists of data representing terrain altitude plus the height of objects on the Earth\u0026rsquo;s surface, corresponding to all laser return pulses, while the DTM is generated through the filtering process of the DSM using the last return signal of the laser pulse, which corresponds to the terrain surface [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe second stage consisted of processing the data collected previously, with the objective of generating additional attributes to compose the database used in the image classification experiments: the Digital Height Model (DHM), the Normalized Difference Vegetation Index (NDVI), and slope.\u003c/p\u003e \u003cp\u003eThe DHM (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea) was derived from LiDAR data by subtracting the DSM, which represents the top of terrain features, from the DTM, which represents the ground surface. The DHM was classified into four vegetation height classes: 0\u0026ndash;0.5 m (grassland), 0.5\u0026ndash;4 m (low vegetation), 4\u0026ndash;12 m (medium vegetation), and \u0026gt;\u0026thinsp;12 m (tall vegetation) [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBased on the DTM, slope (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) was generated, calculated from the planimetric distance and altimetric amplitude between two points, corresponding to the zenithal inclination angle of the terrain surface relative to the horizontal plane, with values ranging from 0\u0026deg; to 90\u0026deg; [\u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e74\u003c/span\u003e]. Six slope classes were established according to the IBGE classification [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]: 0\u0026ndash;3% (flat); 3\u0026ndash;8% (gently undulating); 8\u0026ndash;20% (undulating); 20\u0026ndash;45% (strongly undulating); 45\u0026ndash;75% (mountainous); and \u0026gt;\u0026thinsp;75% (escarpment).\u003c/p\u003e \u003cp\u003eFinally, NDVI (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed) was calculated as the normalized difference between the reflectance of the near-infrared and red bands [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. The index ranges from \u0026minus;\u0026thinsp;1 to 1, with positive values corresponding to rocks, bare soil, and vegetation. Values closer to 1 indicate dense and healthy vegetation, while values near 0 represent terrain without vegetation cover. Negative values occur when targets correspond to water, clouds, or snow [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the third stage, training samples were generated to train the supervised classification algorithm, according to user-defined classes, in the stage understood as machine learning [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. A total of 74,723 samples (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) were collected for eight land cover classes, randomly divided into two subsets by the classification algorithms, with 70% of the samples used for training and 30% for accuracy testing.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eEight classes were defined, three corresponding to forests, one to grassland, one to wetlands, and the remaining to non-vegetation classes (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The distribution of training samples was carried out through the construction of polygons, which were subsequently converted into multiple points, using Planet imagery, Google Earth, the Digital Height Model (DHM), and field knowledge as the basis.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eLand use and land cover classes defined for supervised classification. The table provides the description of each class and the number of reference samples used.\u003c/p\u003e \u003cdiv class=\"Credit\"\u003e\u003cp\u003eAdapted from: MapBiomas [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e], Jennings et al. [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], Junk et al. [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], and CONAMA Resolution No. 004/1994 [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN. of samples\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRock Outcrop (RO)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExposed rock with or without rupestrian vegetation, located in shallow soils such as hilltops or steep slopes.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,961\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTall Forest (TF)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePredominantly arboreal vegetation with lower structural and floristic alteration. Height above 12 m, with emergent individuals exceeding 30 m, characterizing an advanced stage of regeneration.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20,253\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedium Forest (MF)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePredominantly arboreal and shrub vegetation with variable floristic composition. Height between 4 and 12 m, characterizing an intermediate stage of regeneration.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16,189\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLow Forest (LF)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHerbaceous-shrub vegetation up to 4 m in height with an open canopy, allowing greater light penetration. Includes rupestrian vegetation\u0026mdash;rupicolous plants on litholic soils or fixed on rocks..\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6,012\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrassland/Pasture (G/P)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHerbaceous cover, predominantly grasses, for agricultural, recreational, or ornamental use, with height up to 50 cm.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,792\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuilt-up Area (BU)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUrban fabric composed of houses, buildings, multipurpose structures, roads, and other urban infrastructure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,286\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWater Body (WB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNatural water masses corresponding to coastal lagoons.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e258\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWetland (WL)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHerbs and shrubs\u0026mdash;mangroves and marshes\u0026mdash;adapted to waterlogged or poorly drained soils, originating from sedimentary deposits. Occur along lagoon margins or drainage channels in coastal areas..\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22,972\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe database for classification was composed of twelve attributes, corresponding to the eight spectral bands of Planet imagery (seven visible bands and one near-infrared), two geomorphometric attributes (slope and DTM), and two vegetation attributes (DHM and NDVI). For such a large dataset, cloud-based data processing using programming languages is an important resource for vegetation cover classification.\u003c/p\u003e \u003cp\u003eThus, supervised classification was performed in the Google Colab environment, which includes the Python Scikit-Learn library, providing machine learning classification algorithms. The classification script was created by Fernandes and Sami [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] and is available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9\u003c/span\u003e\u003cspan address=\"https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eRandom Forest (RF), developed by Breiman [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], is an ensemble learning algorithm in which multiple classifiers are combined to improve classification [\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. This algorithm creates decision trees from the training samples collected for data classification. The results generated by each tree are combined to obtain the final classification, increasing accuracy [\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. RF has shown excellent results for remote sensing data classification and is increasingly cited in specialized literature [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe K-Nearest Neighbors (KNN) algorithm, developed by Fix and Hodges [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], is one of the most widely used [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. It predicts data based on a simple majority vote of the n nearest neighbors previously classified in an attribute space, following the principle that similar points belong to the same class. The number of neighbors and the distance measure between points around the pixel are the main parameters, along with similarity, to make predictions and assign the unclassified point to the attribute space of the class corresponding to the majority of its nearest pixels [\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe Support Vector Machine (SVM) algorithm is a computational learning technique for pattern recognition problems introduced by Cortes and Vapnik [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. It is a classifier that defines linear boundaries based on support vectors that maximize the distance between classes. However, since many classification problems are non-linear, kernel functions are used to transform the data into a higher-dimensional attribute space [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. In this way, SVM maximizes the decision surface for class separation [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], showing good performance with remote sensing data [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor the configuration of hyperparameters related to the three machine learning algorithms used in this study, the default values of the Python Scikit-Learn library were adopted, as defined in Scikit Learn [\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e, \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e, \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e] and presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHyperparameters of the algorithms used in supervised classification. Were: n_estimators \u0026ndash; number of trees in the model, min_samples_leaf \u0026ndash; minimum number of samples a leaf node must contain,min_samples_split \u0026ndash; minimum number of samples required to split an internal node, random_state \u0026ndash; seed for reproducibility, max_depth \u0026ndash; maximum depth allowed for the trees, n_neighbors \u0026ndash; number of neighbors considered for classification or regression, kernel \u0026ndash; function used to map data into higher dimensions, C \u0026ndash; regularization parameter controlling the penalty for misclassification Source: Python Scikit-Learn library.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHyperparameters\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eRandom Forest - RF\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003en_estimators\u0026thinsp;=\u0026thinsp;100\u003c/p\u003e \u003cp\u003emin_samples_leaf\u0026thinsp;=\u0026thinsp;5\u003c/p\u003e \u003cp\u003emin_samples_split\u0026thinsp;=\u0026thinsp;2\u003c/p\u003e \u003cp\u003erandom_state\u0026thinsp;=\u0026thinsp;42\u003c/p\u003e \u003cp\u003emax_depth\u0026thinsp;=\u0026thinsp;80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eK-nearest neighbour - KNN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003en_neighbors\u0026thinsp;=\u0026thinsp;3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSupport Vector Machine - SVM\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ekernel = 'linear'\u003c/p\u003e \u003cp\u003eC\u0026thinsp;=\u0026thinsp;1.0\u003c/p\u003e \u003cp\u003erandom_state\u0026thinsp;=\u0026thinsp;42\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe final stage consisted of evaluating the accuracy of the three classifications based on 30% of the collected samples, i.e., 22,417 points, using cross-validation [\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e]. In this way, the minimum requirement of 50 samples per class for maps with fewer than 12 classes, established by Congalton and Green [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], was exceeded. Error matrices were constructed with the following performance indicators: (1) overall accuracy \u0026ndash; number of correctly classified samples relative to the total reference samples; (2) producer\u0026rsquo;s accuracy (precision) \u0026ndash; number of correctly classified samples of class k relative to the total samples of that class; (3) user\u0026rsquo;s accuracy (recall) \u0026ndash; number of correctly classified samples of class k relative to the total samples classified as that class; (4) F-score (f1-score) \u0026ndash; harmonic mean between precision and recall [\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAdditionally, a ranking of permutation importance of the attributes used in the three classification experiments was generated. This metric evaluates the importance of attributes in a supervised learning model by measuring the impact each attribute has on model performance when its values are randomly perturbed while keeping the others fixed [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFinally, class balancing was applied to the classification with the best performance to assess its accuracy. For this, the Synthetic Minority Oversampling Technique (SMOTE) was used, which aims to create synthetic examples in the minority class of interest [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results and Discussion","content":"\u003cp\u003eThe results obtained in this study highlight, above all, the importance of geomorphometric attributes in the discrimination of vegetation classes, particularly the Digital Terrain Model (DTM), which showed the highest explanatory power among the variables analyzed, surpassing spectral attributes. Complementarily, the analysis of the performance of supervised classification algorithms demonstrated that Random Forest (RF) presented the best accuracy and consistency indicators across classes, while SVM and KNN showed limitations. In addition, the effects of data balancing and the individual contribution of attributes to classification were evaluated. Based on these findings, this section is organized into: (i) comparison of algorithm performance; (ii) influence of sample balancing; and (iii) importance of attributes in classification.\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Comparison of Algorithm Performance\u003c/h2\u003e \u003cp\u003eIn the assessment of the final results of the three experiments, the classification performed by the KNN algorithm presented a high level of \u0026ldquo;salt-and-pepper\u0026rdquo; noise; therefore, its accuracy was not analyzed.\u003c/p\u003e \u003cp\u003eThe land cover classification generated by the SVM algorithm achieved an overall accuracy of 87.67%. However, the values of producer\u0026rsquo;s accuracy (precision), user\u0026rsquo;s accuracy (recall), and F-score obtained for each class allow for a more precise and detailed assessment of the model\u0026rsquo;s performance. Producer\u0026rsquo;s accuracy indicates that classes such as WL, G/P, and BU performed exceptionally well, with values between 99% and 100%, whereas forest classes scored below 85%, suggesting greater confusion. User\u0026rsquo;s accuracy ranged from 82% to 100%, highlighting excellent performance in well-defined classes but lower performance in forest categories. Similarly, the F-score varied between 74% and 100%, reinforcing that, although the model is robust across several classes, it remains fragile in forest classes, where overlapping features lead to a higher number of errors. In summary, the three indicators together demonstrate that the model is globally effective but requires specific adjustments for classes with lower separability (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConfusion matrix and derived accuracy metrics for land cover classification using the Support Vector Machine (SVM) model. The table presents reference samples, correctly classified instances, and precision, recall, and F-score for each class.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"11\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" morerows=\"1\" nameend=\"c2\" namest=\"c1\" rowspan=\"2\"\u003e \u003cp\u003eClasses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"9\" nameend=\"c11\" namest=\"c3\"\u003e \u003cp\u003eReference samples\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eG/P\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBU\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eLF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eWB\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eWL\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e \u003cp\u003eTot.class\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003e\u003cb\u003eC\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eL\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eA\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eS\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eS\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eI\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eF\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eI\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eE\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e547\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e559\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eG/P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1,136\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e1,145\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e920\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e946\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e5,059\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1,014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e6,143\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e990\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e3,709\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e179\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e4,879\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e374\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e1,344\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e1,807\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e66\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e\u003cb\u003e6,872\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e6,873\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eTotal collected\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e566\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,147\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6,132\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5,098\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1,602\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e6,872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e22,417\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eCorrectly classified\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e547\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,136\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e920\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5,059\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e3,709\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1,344\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e6,872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e19,653\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e83%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e73%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e84%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e96%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e88%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e98%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e82%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e76%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e74%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eF-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e98%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e82%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e74%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e79%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e98%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe land cover classification generated by the Random Forest (RF) algorithm (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) achieved an overall accuracy of 95.67%. The indices of producer\u0026rsquo;s accuracy (precision), user\u0026rsquo;s accuracy (recall), and F-score demonstrate a significantly superior performance of RF compared to the previous matrix. Producer\u0026rsquo;s accuracy ranged from 89% to 100%, with classes such as G/P, WB, and WL reaching 100%, indicating that nearly all samples in these categories were correctly classified. User\u0026rsquo;s accuracy varied between 92% and 100%, confirming the reliability of the classifications. Likewise, the F-score, with values between 91% and 100%, highlighted the balance between producer\u0026rsquo;s and user\u0026rsquo;s accuracy, reinforcing that the model not only correctly identifies most instances but also adequately retrieves the relevant ones. Altogether, the three indicators reveal that RF achieved robust and consistent performance, with overall accuracy above 90%, substantially reducing errors observed in more problematic classes in the previous matrix, such as forest categories, and consolidating itself as a highly effective classifier for the dataset analyzed, particularly for classes with similar characteristics, such as forests.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConfusion matrix and derived accuracy metrics for land cover classification using the Random Forest (RF) model. The table presents reference samples, correctly classified instances, and precision, recall, and F-score for each class.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"11\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" morerows=\"1\" nameend=\"c2\" namest=\"c1\" rowspan=\"2\"\u003e \u003cp\u003eClasses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"9\" nameend=\"c11\" namest=\"c3\"\u003e \u003cp\u003eReference samples\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eG/P\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBU\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eLF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eWB\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eWL\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e \u003cp\u003eTot.class\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003e\u003cb\u003eC\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eL\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eA\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eS\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eS\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eI\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eF\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eI\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eE\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e555\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e559\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eG/P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1,138\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e1,145\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e939\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e945\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e5,692\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e431\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e6,143\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e333\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e4,526\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e4,879\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e102\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e1,661\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e1,807\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e62\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e\u003cb\u003e6,873\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e6,873\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eTotal collected\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e561\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,140\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e943\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6,068\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5,059\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1,707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e6,877\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e22,417\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eCorrectly classified\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,138\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e939\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5,692\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4,526\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1,661\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e6,873\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e21,446\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e89%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e96%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e93%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e93%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e92%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e94%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eF-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e93%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e91%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Influence of Sample Balancing\u003c/h2\u003e \u003cp\u003eThe performance of the Random Forest (RF) classifier with class balancing proved to be similar to that obtained without balancing. The SMOTE method resampled the number of samples to 16,000 per class, which caused the overall accuracy of the RF to decrease to 95.38%, representing a difference of only 0.29%. The lower value can be explained by the fact that SMOTE also increases the number of samples in classes with reduced area, which may introduce noise into the data. However, these effects were not sufficient to cause significant changes in overall accuracy. Thus, the results indicated that the model already exhibited strong performance even with imbalanced data, reinforcing the excellent predictive capacity of the RF model. Douzas et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] also observed similar results, where overall accuracy decreased from 58.7% (without balancing) to 55.7% with SMOTE.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Importance of Attributes in Classification\u003c/h2\u003e \u003cp\u003eRegarding the importance of the twelve attributes used in classification, differences among the tested algorithms were also observed (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). In the RF classification, DTM presented the highest permutation index, with a value of 0.48\u0026mdash;four times greater than DHM, which had 0.12 and ranked second in importance. The remaining attributes\u0026mdash;NDVI, slope, and spectral bands\u0026mdash;showed indices below 0.04. Among these, the visible bands, with the exception of blue, exhibited the lowest values, all below 0.02.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSanlang et al. [\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e] also employed very high spatial resolution imagery and LiDAR data with the aim of mapping land cover and urban functional zones using machine learning algorithms. The authors demonstrated that the classification model with the best performance was the one that integrated all variables, and further emphasized the role of the RF algorithm and LiDAR-derived attributes in improving accuracy. Similarly, Jiang et al. [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e] integrated multitemporal Sentinel-2 imagery (10 m) with LiDAR data to identify dominant tree species in the Shanghai region, using a hierarchical RF classification model. Their study demonstrated the feasibility of large-scale mapping of urban trees and highlighted that the inclusion of LiDAR-derived metrics was crucial for increasing classification accuracy in an urban area characterized by high heterogeneity of land cover classes.\u003c/p\u003e \u003cp\u003eIn the SVM classification, NDVI obtained the highest permutation importance index, with a value of 0.40, followed by DTM with approximately 0.34. The green, red, and near-infrared bands ranked next, with values ranging from 0.22 to 0.17, while DHM occupied the sixth position with a value of 0.16. The remaining attributes\u0026mdash;including the blue band, other Planet spectral ranges, and slope\u0026mdash;presented values below 0.13.\u003c/p\u003e \u003cp\u003eThese results demonstrate the importance of non-pure spectral attributes in forest classification, particularly DTM and DHM in both algorithms, as well as the vegetation spectral index (NDVI) in the SVM algorithm. DHM presented similar importance values in both classification algorithms, though in different positions\u0026mdash;second in RF and sixth in SVM. In contrast, the spectral bands, including non-traditional ones such as coastal blue, green 1, yellow, and red-edge, ranked among the lowest positions in both algorithms.\u003c/p\u003e \u003cp\u003eThe importance of DTM in discriminating forest classes may be related to anthropogenic pressures and topographic conditions that influence the degree of conservation and the structural development of forest remnants [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In densely populated urban areas of the Atlantic Forest biome, which have historically undergone Brazilian economic cycles, such as in Niter\u0026oacute;i, remnants tend to concentrate in areas of difficult access, generally in rugged terrain and at higher altitudes, where human pressure is less intense. From a topographic perspective, steep upper slopes near the crystalline massif summits are characterized by shallow soils, which limit the development of large trees but favor vegetation conservation due to reduced anthropogenic impact. Conversely, lower slopes accumulate colluvial material and present deeper, less inclined soils, conditions that allow vegetation development. However, their proximity to human activities subjects them to greater pressure and risk of degradation. Thus, a contrast emerges: at higher altitudes, conservation is favored by distance from human occupation, although vegetation size is limited; at lower altitudes, forest development potential is greater, but proximity to anthropogenic activities compromises conservation [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHurskainen et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] and Pittman and Hu [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e] highlighted the importance of auxiliary variables in land cover classification, particularly attributes such as elevation, NDVI, and slope. In the present study, the inclusion of these variables contributed significantly to the differentiation among high, medium, and low forest classes, especially in areas where spectral responses were similar across classes.\u003c/p\u003e \u003cp\u003eThe variable importance analysis indicated that topographic attributes played a key role in discriminating structural classes, suggesting that vegetation organization is strongly associated with environmental gradients controlled by relief. This behavior is consistent with the findings of Pittman and Hu [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e], who demonstrated that geomorphometric variables can exhibit high predictive power by capturing variations in moisture, drainage, and solar exposure.\u003c/p\u003e \u003cp\u003eIn this context, areas located in different positions within the landscape tend to present variations in vegetation height and density, which explains the improvement in classification accuracy when these variables are incorporated. Thus, even without species-level distinction, it was possible to separate structural forest classes, evidencing that geomorphometry acts as an important conditioning factor of vegetation cover heterogeneity.\u003c/p\u003e \u003cp\u003eThe RF classification is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The forest classes totaled 6,131 ha, corresponding to 45.8% of the municipality of Niter\u0026oacute;i. The largest area was classified as high-canopy forest, representing 44%, followed by medium-canopy forest with 37%, and low-canopy forest with 19%. Thus, nearly 80% of the remnants exhibit a medium to advanced successional stage, which is a remarkably high index for urban areas within the Atlantic Forest biome. The grassland class, in turn, comprises only 517 ha, the smallest area among the four vegetation classes present on the slopes. On the other hand, the edges of the forest remnants are almost entirely surrounded by buildings, highlighting both their high vulnerability and the critical importance of forest protection policies.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study clearly demonstrated that geomorphometric variables, particularly elevation (DTM) and vegetation height (DHM), play a central role in the classification of forest physiognomy in urban environments with rugged terrain, surpassing the contribution of traditional spectral attributes. DTM obtained the highest permutation importance value in RF and ranked among the top three variables in SVM. DHM also stood out in both classifications with similar importance values. Conversely, although NDVI was prominent in SVM, isolated spectral bands\u0026mdash;especially those in the visible range\u0026mdash;consistently ranked lowest in importance.\u003c/p\u003e \u003cp\u003eFurthermore, it was observed that the classification produced by the RF algorithm yielded the highest accuracy indicators, with overall accuracy above 90% and, most notably, excellent results for producer\u0026rsquo;s accuracy (precision), user\u0026rsquo;s accuracy (recall), and F-score. This indicates superior performance compared to the SVM algorithm, in addition to presenting fewer salt-and-pepper type noise artifacts. The RF classification also achieved higher producer\u0026rsquo;s accuracy, user\u0026rsquo;s accuracy, and F-score values than SVM for forest classes, highlighting its improved ability to separate classes with similar characteristics.\u003c/p\u003e \u003cp\u003eTaken together, these results emphasize the relevance of geomorphometric variables in explaining the structure and distribution of forest remnants in urban areas, influencing both their development and conservation status. The proposed methodology contributes to advancing vegetation mapping in complex environments and underscores the importance of integrating geomorphometric variables into remote sensing studies. This approach can be applied across different contexts to enhance understanding of spatial patterns and landscape transformation processes, providing valuable support for conservation strategies and sustainable management.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCRediT authorship contribution statement\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eM. M. Cotrim\u003c/b\u003e: Writing \u0026ndash; review \u0026amp; editing, Writing \u0026ndash; original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Conceptualization, Investigation. \u003cb\u003eC. N. Francisco\u003c/b\u003e: Writing \u0026ndash; review \u0026amp; editing, Writing \u0026ndash; original draft, Formal analysis, Data curation, Supervision, Project administration. \u003cb\u003eP. J. F. Fernandes\u003c/b\u003e: Writing \u0026ndash; review \u0026amp; editing, Supervision, Formal analysis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eCRediT authorship contribution statement\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eM. M. Cotrim:\u0026nbsp;\u003c/strong\u003eWriting \u0026ndash; review \u0026amp; editing, Writing \u0026ndash; original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Conceptualization, Investigation. \u003cstrong\u003eC. N. Francisco:\u0026nbsp;\u003c/strong\u003eWriting \u0026ndash; review \u0026amp; editing, Writing \u0026ndash; original draft, Formal analysis, Data curation, Supervision, Project administration. \u003cstrong\u003eP. J. F. Fernandes:\u0026nbsp;\u003c/strong\u003eWriting \u0026ndash; review \u0026amp; editing, Supervision, Formal analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDeclaration of competing interest\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eThis study was financed in part by the Coordena\u0026ccedil;\u0026atilde;o de Aperfei\u0026ccedil;oamento de Pessoal de N\u0026iacute;vel Superior \u0026ndash; Brasil (CAPES) \u0026ndash; Finance Code 001. The Article Processing Charge (APC) for the publication of this research was funded by the \u003cstrong\u003eCoordination for the Improvement of Higher Education Personnel \u0026ndash; CAPES\u003c/strong\u003e (Research Organization Registry identifier ROR: 00x0ma614). For the purposes of open access, the authors have assigned the \u003cstrong\u003eCreative Commons CC BY license\u003c/strong\u003e to any accepted version of the article.\u003c/p\u003e\n\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eData and code necessary to replicate this research can be found in:https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9. \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAbdalla Ldos, Oliveira S, de Volot\u0026atilde;o LC S (2015) Modelagem din\u0026acirc;mica da vegeta\u0026ccedil;\u0026atilde;o baseada em aut\u0026ocirc;matos celulares: um estudo de caso da regenera\u0026ccedil;\u0026atilde;o da Mata Atl\u0026acirc;ntica no distrito de Aldeia Velha - RJ. Rev Bras Cartogr 67(3)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlba E et al (2022) Compara\u0026ccedil;\u0026atilde;o entre algoritmos de aprendizado de m\u0026aacute;quina para a identifica\u0026ccedil;\u0026atilde;o de floresta tropical sazonalmente seca. Anu\u0026aacute;rio Inst Geoci\u0026ecirc;ncias 45\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlimjan G, Sun T, Liang Y, Jumahun H, Guan Y (2018) A new technique for remote sensing image classification based on combinatorial algorithm of SVM and KNN. Int J Pattern Recognit Artif Intell 32(07):1859012. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1142/S0218001418590127\u003c/span\u003e\u003cspan address=\"10.1142/S0218001418590127\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlshammari AH, Bencsik G, Ali AH (2026) A survey of six classical classifiers, including algorithms, methodological characteristics, foundational variants, and recent advances. Algorithms 19(1):37\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnwer HA, Mohamed T, Hassan A (2025) Assessing vegetation dynamics in Al Jazirah, Sudan using NDVI-based remote sensing techniques. J Saudi Soc Agric Sci 24(18). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s44447-025-00011-0\u003c/span\u003e\u003cspan address=\"10.1007/s44447-025-00011-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAziz G, Minallah N, Saeed A, Khan MA, Tariq U, Khan MA (2024) Remote sensing based forest cover classification using machine learning. Sci Rep 14:69. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-023-50863-1\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-50863-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaines O, Wilkes P, Disney M (2020) Quantifying urban forest structure with open-access remote sensing data sets. Urban Urban Green 50:126653. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ufug.2020.126653\u003c/span\u003e\u003cspan address=\"10.1016/j.ufug.2020.126653\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelgiu M, Drăguţ L (2016) Random forest in remote sensing: a review of applications and future directions. ISPRS J Photogramm Remote Sens 114:24\u0026ndash;31. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.isprsjprs.2016.01.011\u003c/span\u003e\u003cspan address=\"10.1016/j.isprsjprs.2016.01.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBispo PC et al (2012) Effects of the geomorphometric characteristics of the local terrain on floristic composition in the central Brazilian Amazon. Austral Ecol 37:491\u0026ndash;499\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreiman L (2001) Random forests. Mach Learn 45(1):5\u0026ndash;32\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBurak DL, Silveira TCD, Machado LV, Mendon\u0026ccedil;a EDS (2023) Soil biological properties along a topographic gradient in Brazil\u0026rsquo;s Atlantic Forest biome. Diversity 15(10):1097. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/d15101097\u003c/span\u003e\u003cspan address=\"10.3390/d15101097\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCastro FC, Centeno TM (2005) Segmenta\u0026ccedil;\u0026atilde;o de imagens geradas por perfilamento a laser para delimita\u0026ccedil;\u0026atilde;o de \u0026aacute;rvores individuais em uma \u0026aacute;rea de reflorestamento de eucaliptos. XII Simp\u0026oacute;sio Brasileiro de Sensoriamento Remoto, Goi\u0026acirc;nia, Brasil, pp 16\u0026ndash;21. abril 2005\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: synthetic minority over-sampling technique. J Artif Intell Res 16:321\u0026ndash;357\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen S, Lei F, Zang Z, Zhang M (2023) Forest mapping using a VGG16-UNet++ \u0026amp; stacking model based on Google Earth Engine in the urban area. IEEE Geosci Remote Sens Lett 20:1\u0026ndash;5 Art. 2502205. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/LGRS.2023.3274128\u003c/span\u003e\u003cspan address=\"10.1109/LGRS.2023.3274128\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen J (1960) A coefficient of agreement for nominal scales. Educ Psychol Meas 20(1):37\u0026ndash;46\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eConselho Nacional do Meio Ambiente (CONAMA) (1994) Resolu\u0026ccedil;\u0026atilde;o n\u0026ordm; 006, de 4 de maio de 1994. Disposi\u0026ccedil;\u0026otilde;es gerais para a classifica\u0026ccedil;\u0026atilde;o de \u0026aacute;reas de preserva\u0026ccedil;\u0026atilde;o ambiental e unidades de conserva\u0026ccedil;\u0026atilde;o, e diretrizes para a elabora\u0026ccedil;\u0026atilde;o de programas de conserva\u0026ccedil;\u0026atilde;o e manejo dessas \u0026aacute;reas. Bras\u0026iacute;lia: CONAMA. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ibama.gov.br/sophia/cnia/legislacao/MMA/RE0006-040594.PDF\u003c/span\u003e\u003cspan address=\"https://www.ibama.gov.br/sophia/cnia/legislacao/MMA/RE0006-040594.PDF\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (ibama.gov.br in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCongalton R, Green K (2009) Assessing the accuracy of remotely sensed data: principles and practices. CRC/Taylor \u0026amp; Francis, Boca Raton, p 183\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273\u0026ndash;297\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCorte APD, Broadbent EN, Santos LHO (2022) Aplica\u0026ccedil;\u0026otilde;es do LiDAR para o invent\u0026aacute;rio florestal \u0026ndash; enfoque unidade de \u0026aacute;rea. S\u0026atilde;o Paulo: AT. ISBN 978-65-258-0122-3. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.22533/at.ed.223222904\u003c/span\u003e\u003cspan address=\"10.22533/at.ed.223222904\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCusworth S, Gkoutos GV, Acharjee A (2024) A novel generative adversarial networks modelling for the class imbalance problem in high dimensional omics data. BMC Med Inf Decis Mak 24\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCotrim MM, Francisco CN, Fernandes PJF (no prelo) Avalia\u0026ccedil;\u0026atilde;o do potencial de algoritmos n\u0026atilde;o supervisionados para discrimina\u0026ccedil;\u0026atilde;o de fisionomias florestais\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeering DW, Rouse JW, Haas RH, Schell JA (1975) Measuring forage production of grazing units from Landsat MSS data. Int Symp Remote Sens Environ 10 Ann Arbor Proc Vol 2:1169\u0026ndash;1178\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDouzas G et al (2019) Imbalanced learning in land cover classification: improving minority classes\u0026rsquo; prediction accuracy using the geometric SMOTE algorithm. Remote Sens 11(24):3040. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs11243040\u003c/span\u003e\u003cspan address=\"10.3390/rs11243040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDubayah R, Knox R, Hofton M, Blair JB, Drake J (2000) Land surface characterization using LiDAR remote sensing. In: Hill M, Aspinall R (eds) Spatial Information for Land Use Management. International Publishers Direct, Singapore\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuda RO, Hart PE, Stork DG (2000) Pattern classification, 2nd edn. Wiley-Interscience, New York\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eErlhal GJ et al (1991) Um sistema de segmenta\u0026ccedil;\u0026atilde;o e classifica\u0026ccedil;\u0026atilde;o de imagem de sat\u0026eacute;lite: aplica\u0026ccedil;\u0026otilde;es em meteorologia e geoprocessamento. Anais, S\u0026atilde;o Jos\u0026eacute; dos Campos, SP\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFaceli K et al (2021) Intelig\u0026ecirc;ncia artificial: uma abordagem de aprendizado de m\u0026aacute;quina. Blucher, S\u0026atilde;o Paulo\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernandes PJF, Sami JA (2023) Script para classifica\u0026ccedil;\u0026atilde;o de remanescentes florestais. Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9\u003c/span\u003e\u003cspan address=\"https://acrobat.adobe.com/id/urn:aaid:sc:VA6C2:370a75b8-de9b-439c-a7e6-af7c173663e9\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (acrobat.adobe.com in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFisher A, Rudin C, Dominici F (2019) All models are wrong, but many are useful: learning a variable\u0026rsquo;s importance by studying an entire class of prediction models simultaneously. arXiv. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/1801.01489\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/1801.01489\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFix E, Hodges JL Jr (1951) Discriminatory analysis, nonparametric discrimination. USAF School of Aviation Medicine, Randolph Field, Tex., Project 21-49-004, Rept 4, AF41(128)-31\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGezahagn R, Mengistu F (2025) Morphometric and land use land cover analysis for the management of water resources in Guder sub-basin. Ethiopia Appl Water Sci 15:18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s13201-024-02325-w\u003c/span\u003e\u003cspan address=\"10.1007/s13201-024-02325-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHao X et al (2023) Urban land cover classification using high-resolution remote sensing imagery: challenges and methods. Remote Sens 15(6):1252\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHartling S, Sagan V, Sidike P, Maimaitijiang M, Carron J (2019) Urban tree species classification using WorldView-2/3 and LiDAR data fusion approach and deep learning. Sensors 19(6):1284. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/s19061284\u003c/span\u003e\u003cspan address=\"10.3390/s19061284\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang R, Yao W, Xu Z, Cao L, Shen X (2022) Information fusion approach for biomass estimation in a plateau mountainous forest using a synergistic system comprising UAS-based digital camera and LiDAR. Comput Electron Agric 198\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHUB SIGeo (2023) Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.sigeo.niteroi.rj.gov.br/\u003c/span\u003e\u003cspan address=\"https://www.sigeo.niteroi.rj.gov.br/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (sigeo.niteroi.rj.gov.br in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHurskainen P, Adhikari H, Siljander M, Pellinen J, Heiskanen J (2019) Auxiliary datasets improve accuracy of object-based land use/land cover classification in heterogeneous savanna landscapes. Remote Sens Environ 233:111354. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rse.2019.111354\u003c/span\u003e\u003cspan address=\"10.1016/j.rse.2019.111354\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHyypp\u0026auml; J, Kelle O, Lehikoinen M, Inkinen M (2001) A segmentation-based method to retrieve stem volume estimates from 3-D tree height models produced by laser scanners. IEEE Trans Geosci Remote Sens 39(5):969\u0026ndash;975\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIBGE (2009) Manual t\u0026eacute;cnico de geomorfologia. 2\u0026ordf; ed. IBGE, Rio de Janeiro\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJennings MD et al (2006) Description, documentation, and evaluation of associations and alliances within the U.S. national vegetation classification, version 4.5. Ecological Society of America, Vegetation Classification Panel, Washington, DC, 119 p\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang Y, Li X, Peng L, Li C, Song T (2025) Assessing the potential of multi-seasonal Sentinel-2 satellite imagery combined with airborne LiDAR for urban tree species identification. Sci Rep 15:25107. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-025-10971-6\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-10971-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJunk WJ, Bayley PB, Sparks RE (1989) The flood pulse concept in river-floodplain. Oil Fat Ind 7(7):249\u0026ndash;250\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi C, Jie W, Lei W, Luanyun H, Peng G (2014) Comparison of classification algorithms and training sample sizes in urban land classification with Landsat Thematic Mapper imagery. Remote Sens 6(2):964\u0026ndash;983. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs6020964\u003c/span\u003e\u003cspan address=\"10.3390/rs6020964\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi S, Dragicevic S, Castro FA, Sester M, Winter S, Coltekin A, Pettit C, Jiang B, Haworth J, Stein A, Cheng T (2016) Geospatial big data handling theory and methods: a review and research challenges. ISPRS J Photogramm Remote Sens 115:119\u0026ndash;133\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Y, Li Z, Wang Z, Wang K, Yu Q (2023) Semi-supervised remote sensing image scene classification with prototype-based consistency. J Remote Sens 16:1023\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi M, Cao S, Zhu Z, Wang Z, Myneni RB, Piao S (2023b) Earth Syst Sci Data 15:4181\u0026ndash;4203. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5194/essd-15-4181-2023\u003c/span\u003e\u003cspan address=\"10.5194/essd-15-4181-2023\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing) Spatiotemporally consistent global dataset of the GIMMS Normalized Difference Vegetation Index (PKU GIMMS NDVI) from 1982 to 2022\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacarringue T et al (2022) Developments in land use and land cover classification techniques in remote sensing: a review. J Geogr Inf Syst 14(3):123\u0026ndash;145\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMapBiomas B (2024) Legenda da Cole\u0026ccedil;\u0026atilde;o 9: descri\u0026ccedil;\u0026atilde;o das classes. Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://brasil.mapbiomas.org/wp-content/uploads/sites/4/2024/08/Legenda-Colecao-9-Descricao-Detalhada-PDF-PT.pdf\u003c/span\u003e\u003cspan address=\"https://brasil.mapbiomas.org/wp-content/uploads/sites/4/2024/08/Legenda-Colecao-9-Descricao-Detalhada-PDF-PT.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (brasil.mapbiomas.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaulani Y, Surendro K (2025) Detailed land use classification model based on vegetation indices and texture features. Remote Sens Appl Soc Environ 40:101786. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rsase.2025.101786\u003c/span\u003e\u003cspan address=\"10.1016/j.rsase.2025.101786\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaulik U, Chakraborty D (2017) Remote sensing image classification: a survey of support-vector-machine-based advanced techniques. IEEE Geosci Remote Sens Mag 5(1):33\u0026ndash;52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/MGRS.2016.2641240\u003c/span\u003e\u003cspan address=\"10.1109/MGRS.2016.2641240\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNascimento RFF, Alc\u0026acirc;ntara EH, Kampel M, Stech JL, Novo EMLM, Fonseca LMG (2009) O algoritmo support vector machines (SVM): avalia\u0026ccedil;\u0026atilde;o da separa\u0026ccedil;\u0026atilde;o \u0026oacute;tima de classes em imagens CCD-CBERS-2. Anais, pp 2079\u0026ndash;2086\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeto D, Seabra V, Correia M (2015) An\u0026aacute;lise da distribui\u0026ccedil;\u0026atilde;o da ocupa\u0026ccedil;\u0026atilde;o urbana no munic\u0026iacute;pio de Niter\u0026oacute;i a partir do mapeamento de uso e cobertura da terra. UERJ/FFP, S\u0026atilde;o Gon\u0026ccedil;alo, RJ\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeyns R, Canters F (2022) Mapping of urban vegetation with high-resolution remote sensing: a review. Remote Sens 14(4):1031\u0026ndash;1059. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs14041031\u003c/span\u003e\u003cspan address=\"10.3390/rs14041031\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOc\u0026oacute;n JP, Stavros EN, Steinberg SJ, Robertson J, Gillespie TW (2024) Remote sensing approaches to identify trees to species-level in the urban forest: a review. Prog Phys Geogr Earth Environ 48(3):438\u0026ndash;453. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/03091333241252520\u003c/span\u003e\u003cspan address=\"10.1177/03091333241252520\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOliveira-Filho AT, Fontes MAL (2000) Patterns of floristic differentiation among Atlantic forests in southeastern Brazil and the influence of climate. Biotropica 32(4):793\u0026ndash;810. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1744-7429.2000.tb00619.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1744-7429.2000.tb00619.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOliveira-Filho AT, Tameir\u0026atilde;o Neto E, Carvalho WAC, Werneck M, Brina AE, Vidal CV, Rezende SC, Pereira JAA (2005) An\u0026aacute;lise flor\u0026iacute;stica do compartimento arb\u0026oacute;reo de \u0026aacute;reas de Floresta Atl\u0026acirc;ntica sensu lato na regi\u0026atilde;o das Bacias do Leste (Bahia, Minas Gerais, Esp\u0026iacute;rito Santo e Rio de Janeiro). Rodrigu\u0026eacute;sia 56(87):185\u0026ndash;235. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/2175-78602005568715\u003c/span\u003e\u003cspan address=\"10.1590/2175-78602005568715\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOppong J, Namwamba JB, Twumasi YA, Ning ZH, Asare-Ansah AB, Akinrinwoye C, Annan J (2023) Urbanization and urban forest loss: a spatial analysis of five metropolitan districts in Ghana. Geol Ecol Landsc 1\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/24749508.2023.2202439\u003c/span\u003e\u003cspan address=\"10.1080/24749508.2023.2202439\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhan TN, Kuch V, Lehnert LW (2020) Land cover classification using Google Earth Engine and random forest classifier\u0026mdash;the role of image composition. Remote Sens 12(15):2411\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhiri D, Morgenroth J (2017) Developments in Landsat land cover classification methods: a review. Remote Sens 12(20):3340\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePittman RC, Hu B (2023) Contribution of topographic features and categorization uncertainty for a tree species classification in the boreal biome of Northern Ontario. GIScience Remote Sens 60(1):2214994. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/15481603.2023.2214994\u003c/span\u003e\u003cspan address=\"10.1080/15481603.2023.2214994\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlanet (2026) Understanding PlanetScope instruments. Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://developers.planet.com/docs/apis/data/sensors/ (developers.planet.com\u003c/span\u003e\u003cspan address=\"https://developers.planet.com/docs/apis/data/sensors/ (developers.planet.com\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePMN (2018) Atlas das unidades de conserva\u0026ccedil;\u0026atilde;o do munic\u0026iacute;pio de Niter\u0026oacute;i. Coordena\u0026ccedil;\u0026atilde;o t\u0026eacute;cnica: Amanda Jevaux da Silva Sousa. Edi\u0026ccedil;\u0026atilde;o Pedro Bittencourt, Niter\u0026oacute;i, RJ\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRossiter DG (2004) Technical note: statistical methods for accuracy assessment of classified thematic maps. Int Inst Geo-inf Sci Earth Obs (ITC), Enschede\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSanlang S, Cao S, Du M, Mo Y, Chen Q, He W (2021) Integrating aerial LiDAR and very-high-resolution images for urban functional zone mapping. Remote Sens 13(13):2573. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs13132573\u003c/span\u003e\u003cspan address=\"10.3390/rs13132573\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSobral BS, Oliveira-J\u0026uacute;nior JF, Gois G, Terassi PMB, Muniz-J\u0026uacute;nior JGR (2018) Variabilidade espa\u0026ccedil;o-temporal e interanual da chuva no estado do Rio de Janeiro. Rev Bras Climatol 22:281\u0026ndash;308\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidt K, Dikau R (1999) Landform classification of the eastern Andes of Ecuador based on a digital elevation model. Geomorphology 28(1\u0026ndash;2):1\u0026ndash;15\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScikit-learn (2024a) Ensemble methods. Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit-learn.org/0.16/modules/ensemble.html\u003c/span\u003e\u003cspan address=\"https://scikit-learn.org/0.16/modules/ensemble.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (scikit-learn.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScikit-learn (2024b) KNeighborsClassifier. Dispon\u0026iacute;vel em: https://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003escikit-learn.org/dev/modules/generated/sklearn.neighbors.KNeighborsClassifier.html#sklearn.neighbors.KNeighborsClassifier\u003c/span\u003e\u003cspan address=\"http://scikit-learn.org/dev/modules/generated/sklearn.neighbors.KNeighborsClassifier.html#sklearn.neighbors.KNeighborsClassifier\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (scikit-learn.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScikit-learn (2024c) RandomForestClassifier. Dispon\u0026iacute;vel em: https://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003escikit-learn.org/1.5/modules/generated/sklearn.ensemble.RandomForestClassifier.html\u003c/span\u003e\u003cspan address=\"http://scikit-learn.org/1.5/modules/generated/sklearn.ensemble.RandomForestClassifier.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (scikit-learn.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScikit-learn (2024d) SVC. Dispon\u0026iacute;vel em: https://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003escikit-learn.org/1.5/modules/generated/sklearn.svm.SVC.html#sklearn.svm.SVC\u003c/span\u003e\u003cspan address=\"http://scikit-learn.org/1.5/modules/generated/sklearn.svm.SVC.html#sklearn.svm.SVC\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (scikit-learn.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScikit-learn (2024e) Cross-validation: evaluating estimator performance. Dispon\u0026iacute;vel em: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit-learn.org/stable/modules/cross_validation.html\u003c/span\u003e\u003cspan address=\"https://scikit-learn.org/stable/modules/cross_validation.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (scikit-learn.org in Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSega M, Xiao Y (2011) Multivariate random forests. Wiley Interdiscip Rev Data Min Knowl Discov 1(1):80\u0026ndash;87\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeabra VS (2012) An\u0026aacute;lise da paisagem em apoio aos estudos de favorabilidade \u0026agrave; recupera\u0026ccedil;\u0026atilde;o florestal na bacia hidrogr\u0026aacute;fica do rio S\u0026atilde;o Jo\u0026atilde;o. Tese de Doutorado, PPGG-UFRJ, Rio de Janeiro, p 189\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSovann C et al (2025) Land cover classification of tropical land cover mapping. Remote Sens 17(9):1551. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs17091551\u003c/span\u003e\u003cspan address=\"10.3390/rs17091551\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValeriano MM (2008) TOPODATA: guia para utiliza\u0026ccedil;\u0026atilde;o de dados geomorfom\u0026eacute;tricos locais. Relat\u0026oacute;rio T\u0026eacute;cnico INPE-15318-RPQ/818, INPE, S\u0026atilde;o Jos\u0026eacute; dos Campos\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVeloso HP et al (1991) Classifica\u0026ccedil;\u0026atilde;o da vegeta\u0026ccedil;\u0026atilde;o brasileira, adaptada a um sistema universal. Minist\u0026eacute;rio da Economia, Fazenda e Planejamento; IBGE, Rio de Janeiro\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValkenborg D et al (2023) Supervised learning. Am J Orthod Dentofac Orthop 164(1):146\u0026ndash;149\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Atlantic Forest, Forest remnants, Supervised classification, Algorithms, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-9593796/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9593796/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study aims to evaluate the relative contribution of spectral and geomorphometric attributes in the supervised classification of forest physiognomies in urban remnants of the Atlantic Forest, using machine learning algorithms. The performance of three algorithms—Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM)—was compared. The dataset consisted of twelve attributes, including spectral variables (bands and vegetation index) and non-spectral variables (slope and vegetation height), derived from very high spatial resolution data of forests located in the municipality of Niterói, part of the Rio de Janeiro Metropolitan Region, which encompasses approximately 50% of its territory. Three classification tests were conducted, and the Random Forest (RF) algorithm achieved the best performance, with an overall accuracy of 96% and F-score values above 90%. In comparison, the Support Vector Machine (SVM) obtained an overall accuracy of 88% and F-score values exceeding 74%. Non-spectral attributes, particularly elevation and vegetation height, showed the greatest importance in classification, with permutation indices of 0.48 and 0.12, respectively, while the remaining attributes scored below 0.04. The supervised classification distinguished three physiognomic classes: tall forest, medium forest, and low forest. Beyond demonstrating the superior performance of RF in classifying forest remnants, this study highlights the relevance of geomorphometric variables—especially elevation—in characterizing vegetation physiognomy, whether due to anthropogenic influence or the role of topography in plant physiology.\u003c/p\u003e","manuscriptTitle":"Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-06 04:23:57","doi":"10.21203/rs.3.rs-9593796/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"42a52e3b-f510-4984-b883-7cdb0780c38f","owner":[],"postedDate":"May 6th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-05T21:46:39+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T10:26:01+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T10:25:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"GeoInformatica","date":"2026-05-02T12:34:53+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-08T15:19:24+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-06 04:23:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9593796","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9593796","identity":"rs-9593796","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}