Evaluation of performance of different machine learning techniques for mapping landslide susceptibility associated with extreme rainfall events in southern Brazil

Evaluation of performance of different machine learning techniques for mapping landslide susceptibility associated with extreme rainfall events in southern Brazil | 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 Evaluation of performance of different machine learning techniques for mapping landslide susceptibility associated with extreme rainfall events in southern Brazil Lucas Krein Rademann, Lucas da Silva Ribeiro, Romario Trentin, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6457135/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 13 Oct, 2025 Read the published version in Natural Hazards → Version 1 posted 5 You are reading this latest preprint version Abstract Landslides represent a primary geological process that triggers hazards in steep slope areas, affecting infrastructure and sometimes causing loss of life. Susceptibility mapping is a critical component in the mitigation of landslide-induced disasters, providing technical expertise to support public policy decisions. In May 2024, a significant rainfall event occurred in Southern Brazil, leading to multiple landslides and the transgression of previously established limits of slope stability. Hence, it became necessary to study the landslide susceptibility of this region. Given the complex nature of the landslide process, machine learning tools were used to map the landslide susceptibility using Random Forest (RF), Artificial Neural Network (ANN) and Scoring Sheet (SC) models to compare the performance of these models. The geo-environmental parameters of slope, elevation, slope orientation, catchment area, and curvature were used to train the models. All three models were effective in mapping susceptibility, but the ANN model exhibited the most consistent results, demonstrating a higher frequency of true positives and enhanced accuracy in its classification. The analysis revealed that slope gradient was a key factor in determining susceptibility, with high slope areas being more susceptible, particularly on northeast and east-facing slopes. The data analyzed in this study refers to an extreme rainfall event where the geomorphic thresholds are different from the standards expected for landslide occurrence, making it difficult to determine susceptibility using traditional methods. However, the ML models demonstrated high accuracy in determining the spatial distribution of susceptibility, providing a faster and more accurate analysis. Landslides Geomorphology Machine Learning Brazil Extreme Rainfall Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. INTRODUCTION Landslides are the most common risk-triggering geological processes in steep terrain, often causing damage to property and human life (Petley, 2012 ; Du et al., 2017 ; Palmisano; Vitone; Cotecchia; 2018 ). Landslides are gravitational movements characterized by a high average velocity of short duration, with a well-defined volume and rupture surface, where the detached material is thrown off the slope, typically resulting in long scars (Guidicini and Nieble, 1984 ; Fernandes and Amaral, 1996; Maciel Filho and Nummer, 2011 ). The process is triggered by several natural factors such as climate, relief, soil and lithology and anthropogenic factors that act to trigger landslides (Sisay et al., 2024 ), making the study of this phenomenon complex, with a large number of parameters and interrelationships between them (Li & Chen, 2019 ; Mao et al., 2022 ; Pourghasemi & Rahmati, 2018 ). Mitigation of landslides requires knowledge of the conditions in which these processes are triggered (Du et al., 2017 ), and susceptibility mapping is a critical step in assessing the risk of landslides and has been widely carried out in the scientific community (Quevedo et al., 2020 ; Schirmer & Robaina, 2023 ; Sisay et al., 2024 ; Zêzere et al., 2017 ). Landslide susceptibility mapping identifies areas that are more prone to landslides based on their geomorphological characteristics. This process estimates the spatial probability of landslide occurrence without considering the intensity of the event or its recurrence (Mao et al., 2022 ; Zêzere et al., 2017 ). As highlighted by Mendonça et al. ( 2024 ), the study of susceptibility to landslide events is an effective way to mitigate disasters caused by these phenomena and to provide technical support for public decision-making on sustainable municipal land use and occupation. Most susceptibility studies are carried out quantitatively using statistical techniques (Guzzetti et al., 1999 ; Reichenbach et al., 2018 ; Sisay et al., 2024 ; Zêzere et al., 2017 ), but in recent years there has been a rise in the use of machine learning (ML) techniques for mapping susceptibility to landslides (Li & Chen, 2019 ; Mao et al., 2022 ; Quevedo et al., 2020 ). Multivariate analysis using ML techniques allows an integrated analysis of geo-environmental parameters in a GIS (Geographic Information System) environment with great precision (Liu et al., 2023 ; Wang et al., 2023 ; Were et al., 2023 ). This method has proven superior to traditional statistical-based methods, as it can model high-dimensional and non-linear data sets, allowing for the assessment of complex environmental interactions (Bouramtane et al., 2022 ). The central region of the state of Rio Grande do Sul, in southern Brazil, has a recurrence of landslides and has been the subject of several studies on susceptibility, using different methodologies (Cardozo et al., 2021 ; Schirmer & Robaina, 2023 ). At the end of April and beginning of May 2024, the region was affected by a series of landslides, in which new thresholds of slope stability were exceeded due to an extreme rainfall event, the largest rainfall event ever recorded for the region. During this period, the accumulated volume of rain reached 408.3 mm in four days (INMET, 2024), representing a value around 60% higher than the historical monthly average for the period (257.2 mm). This event reinforces the need for further studies on susceptibility to landslides in the region, considering the increasing frequency of these extreme events in this part of the globe (Sanches et al., 2019 ). ML models have the ability to learn the relationship between the occurrence of landslides and the preceding physical conditions, which avoids the use of ready-made models that are already assumed in a structured way and performs better than bivariate and multivariate statistical methods (Dickson & Perry, 2016 ; Pourghasemi & Rahmati, 2018 ). However, each ML algorithm has a different way of learning, produces different results, and demands comparative analysis between models. The objective of this study is to use ML techniques to map landslide susceptibility considering the extreme rainfall event of 2024 in the Cerro Comprido watershed, southern Brazil (Fig. 1 ). The Random Forest, Neural Network and Scoring Sheet algorithms were used, and the accuracy of the models used was evaluated through validation and comparison of the data. 2. MATERIAL AND METHODS The survey of the landslides in the study area was based on the Soobitsky ( 2024 ) database, carried out using RapidEye images with a spatial resolution of 5m and 3m, generated by the Super Dove sensor (PSB.SD), available in the PlanetExplorer repository (Planet Team, 2025 ). A visual inspection was performed on the images from May 5 to 15, 2024, about two weeks after the extreme rainfall event, to identify the landslides by changes in texture and color. Once the landslides were identified, the rupture locations were vectorized, refining and enriching the existing database. Watersheds were delineated and morphometric parameters were measured using the Digital Terrain Model (DTM) resulting from the Japanese Aerospace Exploration Agency's Advanced Land Observing Satellite (ALOS) mission (JAXA, 2006) with the Phased Array L-band Synthetic Aperture Radar (PALSAR) provided by the Alaska Satellite Facility (ASF) at 12.5 m spatial resolution. The Watershed tool in ArcPro 3.3 software was used to automatically delineate the watersheds in the study area and extract their drainage networks. Information on the geoenvironmental characteristics of the study area was obtained from the MDT. Elevation was obtained directly from the MDT, while slope in percent was obtained using the Slope tool. The aspect tool was used to calculate the aspect of the slopes, resulting in 8 different directions. To understand the dynamics of water on the slope, information on the plan and curvature profile of the slopes was obtained using the Curvature tool. To analyze the catchment area, the flow accumulation was calculated, which resulted in the number of pixels draining into each pixel in the study area. Considering the area of each pixel as 156.5 m², the drained pixels were multiplied by the area of a pixel to obtain the catchment area of each pixel. Once the physical variables were defined, the methodology for applying the ML model was applied, as shown in Fig. 2 . The database for training and testing the machine learning algorithms was created using the points where landslides occurred in the study area (59 points) and the creation of random sampling points in the study area in a 1:1 ratio. Information was extracted from the points in the training file in the form of continuous data on elevation, slope, catchment area, and curvature, and in the form of categorical data on aspect and presence or absence of landslides. Three supervised machine learning algorithms were used after defining the training database: Artificial Neural Network (ANN), Random Forest (RF) and Scoring Sheet (SC). Neural network-based algorithms simulate the neural networks of the human brain. In the ANN method, "neurons" act to solve some complex problems by extracting trends and detecting patterns (Abdolrasol et al., 2021 ; Binetti et al., 2024 ). These neurons have the ability to learn complex relationships between input and output variables through nonlinear analysis (Pourghasemi & Rahmati, 2018 ). The Random Forest method uses recursive binary splitting to develop multiple uncorrelated decision trees, using two-thirds of the sample to develop a tree and the remaining sample to assess its predictive accuracy and the importance of environmental covariates (Bouramtane et al., 2022 ; Were et al., 2023 ). This model is widely used in environmental analysis due to its adaptability to regression and classification tasks (Bouramtane et al., 2022 ; Hasanuzzaman & Shit, 2024 ; Pourghasemi & Rahmati, 2018 ). The scoring sheet method begins by determining risk scores, which are sparse linear models with integer coefficients that predict risk. In this way, integer values are assigned to the individuals or parameters analyzed, which determine their probability (Liu et al., 2022 ). By defining weights for the different parameters and defining a risk score for intervals between parameter classes, this method is most similar to traditional methods of defining weights and matrices. The models were trained and validated using Orange 3.38 software, with the artificial neural network (ANN) model being trained with 100 neurons and a maximum of 500 iterations. The literature suggests that increasing the number of neurons can enhance the performance of the model, particularly in terms of processing time. It is recommended that the minimum number of neurons be equal to or greater than the number of variables (Mendonça et al., 2024 ; Yotov et al., 2020 ). The choice of 100 neurons as the optimal number was determined through empirical testing, which revealed that higher values did not necessarily lead to improved accuracy. The Random Forest model was trained using a maximum of 200 trees and a minimum data subset size of 5, where the greater the number of decision trees, the more robust the result. The information gain limit was linked to the number of parameters analyzed. The SC model, on the other hand, was trained to use a maximum of 6 decision parameters and to assign a weight of up to 5, which was considered sufficient after systematically testing the data. Following the training of the models, they were applied to the data from the study area, with one point for every 12.5m contained within the watersheds studied. The database contains information on the same geo-environmental parameters as the training file. An index of 0 to 1 probability of landslide occurrence is generated for each pixel and each of the ML models. Values close to 1 indicate higher susceptibility. The data was then spatially discretized for the specified study area, resulting in the definition of three distinct classes: less than 0.5, 0.5 to 0.75, and greater than 0.75. The models were validated using cross-validation in the Orange 3.38 software. The following efficiency parameters were extracted: The Area Under the ROC Curve (AUC), which is a graph of the true positive rate against the false negative rate, with values close to 1 indicating good quality; classification accuracy, which is the proportion of correctly classified data; precision, which is the proportion of true positives among the instances classified as positive; and recall, which is the proportion of true positives among all the positive instances in the data. 3. RESULTS 3.1 GEOENVIRONMENTAL PARAMETERS The Cerro Comprido watershed comprises the geomorphological units defined as Ramps of Jacuí colluvium-alluvium deposits, hills in sedimentary rocks, and association of hills and ridges at the edge of the plateau (Schirmer; Robaina; Trentin, 2013 ). The main aspects of the Ramps of Jacuí colluvium-alluvium deposits unit are represented by the low slope inclination, less than 5%, and the altimetric variation between 40 and 90 meters, associated with recent deposits of the main channel of the Jacuí River. The hills in the sedimentary rocks are defined by a gentle and undulating relief, with a slope of up to 15% and altitudes ranging from 90 to 200 meters, with a predominant bedrock of sandstone of fluvial origin, often friable under the action of weathering and erosion. The association of hills and ridges on the edge of the plateau is made up of volcanic rocks at the top and sandstone with layers of mudstone at the base. There are sections of interbedded sandstone with high-angle crossbeds that mark the overburden contacts where outcrops appear, forming small plateaus between escarpments. This unit is characterized by slopes greater than 15%, with altitudes ranging from 120 to 480 meters (Schirmer; Robaina; Trentin, 2013 ). The physical characteristics of the study area (Fig. 3 ) were analyzed based on elevation (A), slope (B), catchment area (C), aspect (D) and slope curvature (E), variables used for ML processing. The data from the study area indicates a range of elevations from 59 to 545 meters, resulting in an altimetric variation of 486 meters. The most significant class range in terms of area is 200 to 250 meters, where predominate hills in sedimentary rocks. In terms of slope, the predominant slopes are less than 10%, making them flat to gently undulating. However, there are also slopes in areas of strongly undulating relief, ranging from 20–45%. In certain points, the presence of slopes with gradients exceeding 75% has been observed, associated with the hillsides close to the top of the watershed. The catchment area is a variable representing the area draining into the central pixel, which is an indication of the accumulation of water flows. In this variable, higher values indicate greater concentrations of runoff, which means more water available and more susceptibility to landslide. In the Cerro Comprido basin, catchment areas classified as greater than 10,000 m² have been identified as exhibiting high flow concentrations, thereby delineating the drainage network. The aspect of the slope exerts a significant influence on various environmental factors, including wind direction, precipitation, and solar exposure. In the study area, the primary slope orientations are aligned north and northeast. The west-facing slope, however, is the least significant in terms of the area it covers. Still on the subject of slope characteristics, the analysis of slope curvature for the study area reveals that the major slope plane and profile configuration is classified as a Divergent-Concave type, indicated by positive values, these are areas of water distribution on the hillside. Conversely, the negative values indicate a Convergent-Convex type configuration, what indicates flow concentration and more prone to landslides areas. 3.2 SUSCEPTIBILITY ASSESSMENT The susceptibility map, generated using the ANN model, shows that 7.7% of the study area is defined as medium susceptibility and 7.9% as high susceptibility to landslides (Fig. 4 ). The areas exhibiting medium and high susceptibility are predominantly concentrated on the slopes that face east, southeast, and northeast. Areas of low susceptibility (< 0.5) cover about 85% of the study area, especially in the lower areas with less steep slopes near the main channel. The analysis of the susceptibility resulting from the model generated with the RF algorithm shows that 18.4% of the areas have medium and high susceptibility (0.5–0.75 and > 0.75), and the remaining 81.5% have low susceptibility. In this model, the areas of high susceptibility occur in similar regions of the map resulting from the application of the ANN, but with less occurrence in the top of the hills. The model indicates susceptibility for slopes facing north and southwest. The RF model attributed the greatest significance to the slope parameter, where the average of the landslide points is 56%, with a score of 0.413, which is the main parameter for defining susceptibility to landslides in this model, followed by altitude (0.186) and catchment area (0.120). The map resulting from the SC model shows a greater occurrence of areas with medium and high susceptibility than the previous models, with 24.5% of the study area. When considering the distribution of susceptible areas, this model demonstrated the greatest susceptible area on the slopes north and northwest of the study area, which is the area with the greatest divergence between models and with less occurrence of landslides (Fig. 5). In this model, the algorithm identified only the northeast aspect as a distinctive indicator of high susceptibility, assigning more data intervals related to low susceptibility, such as slopes of less than 25% and south-facing slopes. In general, the models considered similar limits for determining areas of medium and high susceptibility (Table 1 ), especially in the variables of curvature and slope orientation, where only the SC model showed a lower occurrence on the south face, in disagreement with the other methods which indicated the southwest slopes as less susceptible. Table 1 Summary of the geoenvironmental variables in the pixels classified as medium and high susceptibility in the three models used. Variable ANN RF SC Slope (%) Max. 213.34 213.35 213.35 Min. 1.41 11.31 13.03 Mean 52.72 56.05 59.42 Elevation (m) Max. 542 537 537 Min. 71 82 68 Mean 341.37 345.57 360.28 Curvature Max. 16.64 16.63 16.64 Min. -11.52 -11.52 -11.52 Mean 0.18 0.09 0.42 Catchment Area (m²) Max. 728437.50 883593.80 883593.80 Min. 0 0 0 Mean 3003.13 2745.31 2485.94 Aspect Higher occurrence Northeast Northeast Northeast Lower occurrence Southwest Southwest South The slope variable demonstrated variation between the models in terms of the minimum and average limits for determining susceptible areas. The ANN model identified the lowest slopes as susceptible areas, with a minimum and average of 1.41% and 52.72%, respectively. The RF model considered slopes with gradients above 11% as susceptible areas, with an average of 56%. This value is similar to that defined by the SC model, which indicated even higher values, referring to slopes with gradients above 13% and an average of almost 60%. The findings of the last two models are close to the values indicated in the study conducted by Ribeiro et al. ( 2025 ), in the same study area, where they highlighted the occurrence of landslides with a higher frequency ratio at gradients from 35 to 40% and reaching the critical point at gradients of 70 to 75%. In terms of elevation, the maximum values are similar between the three models, while the minimum values of ANN and SC are closer, and the values presented by the RF model indicate the highest minimum elevations (82m). The average altitude of the susceptibility points was found to be similar between the ANN and RF models (~ 345m) and around 15 meters higher for the SC model. Despite these differences, the findings from all three models align with the results from previous studies, which have identified the 290 to 440m range as the most susceptible (Ribeiro et al., 2025 ). An analysis of the catchment area indicates that the RF and SC models exhibit equivalent maximum catchment areas, exceeding 883,000 square meters. In contrast, the ANN model exhibits a comparatively diminished maximum catchment area, both in terms of maximum and average values. 3.3 PERFORMANCE OF MACHINE LEARNING MODELS The classification quality of the models used in the study was assessed by cross-validation (Table 2 ). The AUC parameter indicates the ratio of the true positive to the false positive, with a value of 1 indicating perfect classification. All models showed values above 0.8 for this parameter, indicating good accuracy in classifying susceptible areas, with a slightly higher value for the SC model. Table 2 Quality parameters of the analyzed models. Model AUC Classification Accuracy Precision Recall Neural Network 0.851 0.814 0.815 0.814 Random Forest 0.860 0.797 0.797 0.797 Scoring Sheet 0.868 0.788 0.799 0.788 When further quality parameters of the models were evaluated, the ANN exhibited the highest level of accuracy, precision, and recall, with the RF and SC models ranking second and third, respectively. This finding suggests that the ANN model exhibits superior consistency in its results, manifesting as a greater prevalence of accurate classifications and true positives. In contrast, the SC model, despite having a good consistency in the AUC value, performs poorly in the other quality parameters, with greater confusion in the model classification, which has already been observed by (Aguirre-Gutiérrez et al., 2013 ) in some models with spatial distribution of data. These discrepancies can be attributed to the distinct methodologies employed by machine learning (ML) models in assigning importance to geo-environmental variables. The landslide susceptibility value assigned to the rupture points exhibited variation according to the model employed. Higher values were observed in the ANN model (Fig. 6 A), with an average value close to 0.8 and low occurrence of landslides in areas with values below 0.5, indicating good model consistency. When the RF model was evaluated, the results were satisfactory, but with a lower average susceptibility in areas where landslides occurred (0.7), and a concentration of landslides between 0.7 and 0.85. Notably, the SC model exhibits a greater prevalence of sparse susceptibility values, while values greater than 0.7 are indicative of a satisfactory model. The model assigned an average susceptibility below 0.7 (0.67) to the points where landslides occurred, with a higher incidence of landslides in areas with a susceptibility below 0.5. The frequency density of the susceptibility value for landslides in the study area (Fig. 6 B) was analysed, revealing that the Neural Network model exhibited the greatest number of landslides at high susceptibility, with the peak of the curve very close to 1, indicating good performance in predicting the true values. The Random Forest model exhibited a curve peak at slightly lower values and at a lower density. The Scoring Sheet model, on the other hand, showed a peak density at values close to RF, but with a higher occurrence of values close to 0.5, indicating greater inconsistency in the results. A comparative analysis of the results obtained from the ANN model, which demonstrated the optimal performance in terms of landslide susceptibility classification within the designated study area when contrasted with alternative models, reveals a notable degree of congruence between the outcomes of the ANN and RF models, as evidenced by Fig. 7 A. This observation is consistent with the equivalence between the two models in some scenarios, as observed by Pacheco Quevedo et al. ( 2019 ). When represented in a scatter plot, the classification points of these two models show a positive correlation of 0.862, with greater similarity at the extremes of high and low susceptibility. Furthermore, an analysis of the determination of susceptibility in locations where landslides have occurred reveals that the landslide points classified with low susceptibility by the ANN model also have low susceptibility in the RF model. However, this is not the case for the points classified with low susceptibility in the RF model, which receive higher values in the ANN model, demonstrating the higher accuracy of the ANN model. The SC model demonstrated a reduced level of correlation with the ANN model (0.743), with notable disparities between landslide points identified as low-susceptibility by the SC model and those identified as high-susceptibility by the ANN model. These disparities are particularly evident in the bottom right corner of the graph, highlighting the distinction in their predictive capacities. Additionally, there are no landslides classified as low susceptibility in the ANN and high in the SC, which is evidence of the greater assertiveness of the neural network model. In general, the three models demonstrated satisfactory performance in the assessment of landslide susceptibility in the study area. However, the ANN model demonstrated a better performance considering that it has better precision, accuracy, and consistency of true positives, which has already been observed by (Aditian et al., 2018 ) in the study of landslides in Indonesia. Furthermore, the presence of uncertain points in the ANN model was also observed in the other models, suggesting a necessity for additional variables during the model training process. 4. CONCLUSION The efficacy of the machine learning techniques in mapping landslide susceptibility was demonstrated by the three models, which were able to identify the areas most susceptible to landslides with a high degree of precision and accuracy, as indicated by an AUC above 0.800. The analysis revealed that slope gradient was a key factor in determining susceptibility, with high slope areas being more susceptible, particularly on northeast and east-facing slopes. The ANN model demonstrated the greatest consistency in its results, showing a higher incidence of true positives and greater accuracy in its classification. The model exhibited a reduced occurrence of landslides in areas of low susceptibility and a high concentration of landslides in areas of high and medium susceptibility, demonstrating its superior performance and lack of overfitting and demonstrating its potential to be applied to the database with great accuracy. Despite the extensive utilization of the random forest model in the literature for analyzing complex environmental phenomena, such as landslides, this model demonstrated slightly lower precision and accuracy than the artificial neural network. The spatial distribution exhibited by the Random Forest model was comparable to that of the ANN model, however, a greater number of landslides occurred in areas exhibiting low susceptibility. This finding suggests that the Random Forest model performs less effectively than the ANN model. The SC model, which is rarely employed in this type of study, performed poorly in determining susceptibility despite having an operating logic similar to traditional susceptibility mapping methods. The data analyzed in this study refers to an extreme rainfall event where the geomorphic thresholds are different from the standards expected for landslide occurrence, making it difficult to determine susceptibility using traditional methods. However, the ML models demonstrated high accuracy in determining the spatial distribution of susceptibility, providing a faster and more accurate analysis. In addition, it is essential to evaluate the performance of the models taking into account a larger number of geo-environmental parameters, such as lithology and soil types, at appropriate scales, which could increase the accuracy of the predictions and amplify the differences between the models. Further studies are required for the study area, taking into account normal precipitation conditions and conducting inventories over a time series. This will enable the assessment of the recurrence and return time of events. This approach will facilitate the refinement of susceptibility mapping, thereby enabling the identification of regions where risk may be overestimated in analyses based exclusively on extreme events. The incorporation of long-term data is also crucial, as it serves to reduce uncertainties, thereby ensuring more robust predictions for applications in territorial planning, disaster prevention and risk management. Declarations FUNDING DECLARATION This work was supported by the Brazilian National Council for Scientific and Technological Development (CNPq), grant number 302449/2022-1 and the Foundation for Research Support of the State of Rio Grande do Sul (FAPERGS), grant number 24/2551- 0002134-5. COMPETING INTEREST The authors have no relevant financial or non-financial interests to disclose. AUTHOR CONTRINUTIONS All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Lucas Krein Rademann and Lucas da Silva Ribeiro. The first draft of the manuscript was written by Lucas Krein Rademann and Romario Trentin and Luis Eduardo de Souza Robaina commented on previous versions of the manuscript. All authors read and approved the final manuscript. References Abdolrasol, M. G. M., Hussain, S. M. S., Ustun, T. S., Sarker, M. R., Hannan, M. A., Mohamed, R., Ali, J. A., Mekhilef, S., & Milad, A. (2021). Artificial Neural Networks Based Optimization Techniques: A Review. Electronics, 10(21), 2689. https://doi.org/10.3390/electronics10212689 Aditian, A., Kubota, T., & Shinohara, Y. (2018). Comparison of GIS-based landslide susceptibility models using frequency ratio, logistic regression, and artificial neural network in a tertiary region of Ambon, Indonesia. Geomorphology, 318, 101–111. https://doi.org/10.1016/j.geomorph.2018.06.006 Aguirre-Gutiérrez, J., Carvalheiro, L. G., Polce, C., Van Loon, E. E., Raes, N., Reemer, M., & Biesmeijer, J. C. (2013). Fit-for-Purpose: Species Distribution Model Performance Depends on Evaluation Criteria – Dutch Hoverflies as a Case Study. PLoS ONE, 8(5), e63708. https://doi.org/10.1371/journal.pone.0063708 Binetti, M. S., Massarelli, C., & Uricchio, V. F. (2024). Machine Learning in Geosciences: A Review of Complex Environmental Monitoring Applications. Machine Learning and Knowledge Extraction, 6(2), 1263–1280. https://doi.org/10.3390/make6020059 Bouramtane, T., Hilal, H., Rezende-Filho, A. T., Bouramtane, K., Barbiero, L., Abraham, S., Valles, V., Kacimi, I., Sanhaji, H., Torres-Rondon, L., De Castro, D. D., Vieira Santos, J. D. C., Ouardi, J., Beqqali, O. E., Kassou, N., & Morarech, M. (2022). Mapping Gully Erosion Variability and Susceptibility Using Remote Sensing, Multivariate Statistical Analysis, and Machine Learning in South Mato Grosso, Brazil. Geosciences, 12(6), 235. https://doi.org/10.3390/geosciences12060235 Cardozo, G. L., Zanandrea, F., Michel, G. P., & Kobiyama, M. (2021). Inventário de movimentos de massa na bacia hidrográfica do rio Mascarada/RS. Ciência e Natura, 43, e31. https://doi.org/10.5902/2179460X43594 Dickson, M. E., & Perry, G. L. W. (2016). Identifying the controls on coastal cliff landslides using machine-learning approaches. Environmental Modelling & Software, 76, 117–127. https://doi.org/10.1016/j.envsoft.2015.10.029 Du, G., Zhang, Y., Iqbal, J., Yang, Z., & Yao, X. (2017). Landslide susceptibility mapping using an integrated model of information value method and logistic regression in the Bailongjiang watershed, Gansu Province, China. Journal of Mountain Science , 14(2), 249–268. https://doi.org/10.1007/s11629-016-4126-9 Fernandes, N. F.; Amaral, C. P. Movimentos de massa: uma abordagem geológico-geomorfológica. (2003). In: Guerra, A. J. T. and Cunha, S. B. (org.) Geomorfologia e Meio Ambiente. 4ª. ed. Rio de Janeiro: Bertrand. cap. 3, p. 123-194. Guidicini, G.; Nieble, C. M. (1984). Estabilidade de Taludes Naturais e de Escavação. São Paulo: Edgard Blücher. p.196. Guzzetti, F., Carrara, A., Cardinali, M., & Reichenbach, P. (1999). Landslide hazard evaluation: A review of current techniques and their application in a multi-scale study, Central Italy. Geomorphology , 31 (1–4), 181–216. https://doi.org/10.1016/S0169-555X(99)00078-1 Hasanuzzaman, M., & Shit, P. (2024). Assessment of gully erosion susceptibility using four data-driven models AHP, FR, RF and XGBoosting machine learning algorithms. Natural Hazards Research, S2666592124000362. https://doi.org/10.1016/j.nhres.2024.05.001 INMET - Instituto Nacional de Meteorologia. Eventos Extremos de Abril de 2024 no Brasil, Brasília-DF, (2024). Disponível em: https://portal.inmet.gov.br/uploads/notastecnicas/Nota-Extremos-revisado.pdf. Acesso em: 29 jul. 2024. JAXA - Japan Aerospace Exploration Agency ALOS PALSAR L1.0, (2006). Disponível: https://asf.alaska.edu/datasets/daac/alos-palsar/. Acesso em: 01 jun. 2024. Li, Y., & Chen, W. (2019). Landslide Susceptibility Evaluation Using Hybrid Integration of Evidential Belief Function and Machine Learning Techniques. Water , 12 (1), 113. https://doi.org/10.3390/w12010113 Liu, G., Arabameri, A., Santosh, M., & Nalivan, O. A. (2023). Optimizing machine learning algorithms for spatial prediction of gully erosion susceptibility with four training scenarios. Environmental Science and Pollution Research, 30(16), 46979–46996. https://doi.org/10.1007/s11356-022-25090-2 Liu, J., Zhong, C., Li, B., Seltzer, M., & Rudin, C. (2022). FasterRisk: Fast and Accurate Interpretable Risk Scores (No. arXiv:2210.05846). arXiv. https://doi.org/10.48550/arXiv.2210.05846 Maciel Filho, C. L.; Nummer, A. V. (2011). Introdução à geologia de engenharia. Santa Maria: Editora da UFSM. Mao, Y., Mwakapesa, D. S., Li, Y., Xu, K., Nanehkaran, Y. A., & Zhang, M. (2022). Assessment of landslide susceptibility using DBSCAN-AHD and LD-EV methods. Journal of Mountain Science , 19 (1), 184–197. https://doi.org/10.1007/s11629-020-6491-7 Mendonça, R. R., Oliveira, G. G. D., & Tornquist, C. G. (2024). Landslide Susceptibility Modeling Using Artificial Neural Networks in the Municipality of Joinville, southern Brazil. Revista Brasileira de Geomorfologia , 25 (4). https://doi.org/10.20502/rbg.v25i4.2513 Pacheco Quevedo, R., Guasselli, L. A., Garcia De Oliveira, G., & Chimelo Ruiz, L. F. (2019). Modelagem de áreas suscetíveis a movimentos de massa: Avaliação comparativa de técnicas de amostragem, aprendizagem de máquina e modelos digitais de elevação. Geosciences = Geociências, 38(3), 781–795. https://doi.org/10.5016/geociencias.v38i3.14019 Palmisano, F., Vitone, C., & Cotecchia, F. (2018). Assessment of Landslide Damage to Buildings at the Urban Scale. J ournal of Performance of Constructed Facilities , 32(4), 04018055. https://doi.org/10.1061/(ASCE)CF.1943-5509.0001201 Petley, D.; (2012). Global patterns of loss of life from landslides. Geology ; 40 (10): 927–930. doi: https://doi.org/10.1130/G33217.1 Planet Team. Planet Application Program Interface: In Space for Life on Earth. San Francisco, CA: Planet Labs, 2025, p. n/a. Disponível em: https://api.planet.com. Acesso em: feveiro de 2025. Pourghasemi, H. R., & Rahmati, O. (2018). Prediction of the landslide susceptibility: Which algorithm, which precision? CATENA, 162, 177–192. https://doi.org/10.1016/j.catena.2017.11.022 Quevedo, R. P., Oliveira, G. G. D., & Guasselli, L. A. (2020). Mapeamento de Suscetibilidade a Movimentos de Massa a partir de Redes Neurais Artificiais. Anuário do Instituto de Geociências - UFRJ , 43 (2). https://doi.org/10.11137/2020_2_128_138 Reichenbach, P., Rossi, M., Malamud, B. D., Mihir, M., & Guzzetti, F. (2018). A review of statistically-based landslide susceptibility models. Earth-Science Reviews, 180, 60–91. https://doi.org/10.1016/j.earscirev.2018.03.001 Ribeiro, L. da S., Rademann, L. K., Robaina, L. E. de S., Schnorr, G. G., & Trentin, R. (2025). Análise da relação entre a ocorrência de deslizamentos e os atributos do relevo no evento extremo de precipitação no sul do Brasil, maio de 2024. Sociedade & Natureza, 37(1). https://doi.org/10.14393/SN-v37-2025-74694 Sanches, F., Verdum, R., Fisch, G., Gass, S. L. B., & Rocha, V. M. (2019). Extreme Rainfall Events in the Southwest of Rio Grande do Sul (Brazil) and Its Association with the Sandization Process. American Journal of Climate Change, 08(04), 441–453. https://doi.org/10.4236/ajcc.2019.84024 Schirmer, G. J.; Robaina, L. E. S. Trentin, R. (2013). Unidades geomorfológicas em municípios da Quarta Colônia do Rio Grande do Sul. Geografa Ensino & Pesquisa, v. 17, n.2 p. 199-212, maio/ago. https://doi.org/10.5902/223649949244 Schirmer, G. J., & Robaina, L. E. de S. (2023). Mapeamento de áreas susceptíveis a desastres naturais da Quarta Colônia-RS com o base no zoneamento geoambiental. Geografia Ensino & Pesquisa , 27 , e67900–e67900. https://doi.org/10.5902/2236499467900 Sisay, T., Tesfaye, G., Jothimani, M., Reda, T. M., & Tadese, A. (2024). Landslide susceptibility mapping using combined geospatial, FR and AHP models: A case study from Ethiopia’s highlands. Discover Sustainability , 5 (1), 474. https://doi.org/10.1007/s43621-024-00730-4 SOOBITSKY, R. Manual inventory of landslides in Brazil, 2024-04-29. (2024). NASA GSFC. Juang CS, Stanley TA, and Kirschbaum DB. (2019). Using citizen science to expand the global map of landslides: Introducing the Cooperative Open Online Landslide Rep. Wang, Y., Zhang, Y., & Chen, H. (2023). Gully erosion susceptibility prediction in Mollisols using machine learning models. Journal of Soil and Water Conservation, 78(5), 385–396. https://doi.org/10.2489/jswc.2023.00019 Were, K., Kebeney, S., Churu, H., Mutio, J. M., Njoroge, R., Mugaa, D., Alkamoi, B., Ng’etich, W., & Singh, B. R. (2023). Spatial Prediction and Mapping of Gully Erosion Susceptibility Using Machine Learning Techniques in a Degraded Semi-Arid Region of Kenya. Land, 12(4), 890. https://doi.org/10.3390/land12040890. Yotov, K., Hadzhikolev, E., & Hadzhikoleva, S. (2020). Determining the Number of Neurons in Artificial Neural Networks for Approximation, Trained with Algorithms Using the Jacobi Matrix. TEM Journal , 1320–1329. https://doi.org/10.18421/TEM94-02 Zêzere, J. L., Pereira, S., Melo, R., Oliveira, S. C., & Garcia, R. A. C. (2017). Mapping landslide susceptibility using data-driven methods. Science of The Total Environment , 589 , 250–267. https://doi.org/10.1016/j.scitotenv.2017.02.188 Cite Share Download PDF Status: Published Journal Publication published 13 Oct, 2025 Read the published version in Natural Hazards → Version 1 posted Reviewers agreed at journal 20 May, 2025 Reviewers invited by journal 07 May, 2025 Editor invited by journal 03 May, 2025 Editor assigned by journal 16 Apr, 2025 First submitted to journal 15 Apr, 2025 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-6457135","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":453335789,"identity":"8f82d6a3-cb2a-45bd-b591-93e084e9a93b","order_by":0,"name":"Lucas Krein Rademann","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYBACPijN2ACmKhgYDAhpYUPVcoZkLYxtxGiRSH748eceO9l+scPPHhfOOyxvzt58gOFHxTY8WtKMpXmeJRvPnJ1mbjxz22HDnT3HEhh7ztzGoyWHQZrhAHPihtsJZtK82w4zbriRY8DM2IZXC/PPHwfqE/ffTv8mzTvnsD0xWtgkeA4cTtwgnQO0pQHIIKiF55mZNc+B48YzbueUSfMcS0/ecOZYwkF8fuFnT35888eBatn+2enbpHlqrG03HG8++OBHBW4t6KAZTB4gWj0Q1JGieBSMglEwCkYIAAAPl1eXtOAmxwAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-3341-3357","institution":"Federal University of Santa Maria: Universidade Federal de Santa Maria","correspondingAuthor":true,"prefix":"","firstName":"Lucas","middleName":"Krein","lastName":"Rademann","suffix":""},{"id":453335790,"identity":"54449109-dcac-44ba-80f3-06abe4010ae7","order_by":1,"name":"Lucas da Silva Ribeiro","email":"","orcid":"","institution":"Federal University of Santa Maria: Universidade Federal de Santa Maria","correspondingAuthor":false,"prefix":"","firstName":"Lucas","middleName":"da Silva","lastName":"Ribeiro","suffix":""},{"id":453335791,"identity":"c839b35a-a886-45b6-a7f4-65725e4652c0","order_by":2,"name":"Romario Trentin","email":"","orcid":"","institution":"Federal University of Santa Maria: Universidade Federal de Santa Maria","correspondingAuthor":false,"prefix":"","firstName":"Romario","middleName":"","lastName":"Trentin","suffix":""},{"id":453335792,"identity":"ce925a12-d755-4a66-b502-c252def41cc6","order_by":3,"name":"Luis Eduardo de Souza Robaina","email":"","orcid":"","institution":"Federal University of Santa Maria: Universidade Federal de Santa Maria","correspondingAuthor":false,"prefix":"","firstName":"Luis","middleName":"Eduardo de Souza","lastName":"Robaina","suffix":""}],"badges":[],"createdAt":"2025-04-15 17:41:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6457135/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6457135/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s11069-025-07694-2","type":"published","date":"2025-10-13T15:58:36+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82504111,"identity":"fc82adcc-3047-484b-8df9-008ab826a579","added_by":"auto","created_at":"2025-05-12 09:21:09","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":203489,"visible":true,"origin":"","legend":"\u003cp\u003eLocalization of study area.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/b2448a19a8aaf067600017e2.jpg"},{"id":82505492,"identity":"b44f857c-6f59-46de-ac4c-e0c3a8b73176","added_by":"auto","created_at":"2025-05-12 09:29:09","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":47306,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the methodology used to create the machine learning models.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/07f0fc5a6f9f535075fd6086.jpg"},{"id":82504114,"identity":"2251efd4-eeb1-44e1-925e-aa53685ed80f","added_by":"auto","created_at":"2025-05-12 09:21:09","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":115252,"visible":true,"origin":"","legend":"\u003cp\u003eGeoenvironmental parameters analyzed in the study. A – Elevation; B – Slope; C – Catchment Area; D – Aspect; E – Curvature.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/f8fd77b7a27fbb3721c36d04.jpg"},{"id":82505495,"identity":"8c0a188e-fe44-465f-98bd-5b3e70369c03","added_by":"auto","created_at":"2025-05-12 09:29:09","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":230239,"visible":true,"origin":"","legend":"\u003cp\u003eMap of landslide susceptibility in the three machine learning models studied.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/98521482c229476623e2ad06.jpg"},{"id":82504112,"identity":"b622f1cc-7e58-400e-b92d-bce28ab76b45","added_by":"auto","created_at":"2025-05-12 09:21:09","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":127879,"visible":true,"origin":"","legend":"\u003cp\u003e3D perspective view of the map of landslide susceptibility in the Cerro Comprido watershed in the three machine learning models.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/cbed1caf7491d9704a4ba0f1.jpg"},{"id":82505851,"identity":"f9b5863c-5a7c-4c3d-8b59-8eac9ef9769c","added_by":"auto","created_at":"2025-05-12 09:37:09","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":66925,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of the predicted susceptibility to landslides in the three models evaluated (A) and the density of the susceptibility value predicted by each model (B).\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/51796b5ef195f45913e7d0ab.jpg"},{"id":82504120,"identity":"0e739a8b-ae0d-402f-8f4c-dfb142351367","added_by":"auto","created_at":"2025-05-12 09:21:09","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":134505,"visible":true,"origin":"","legend":"\u003cp\u003ePredicted susceptibility scatter for the Cerro Comprido catchment in the ANN and RF model (A) and in the ANN and SC model (B).\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/bc3f7fc0576f28a2af60ecbb.jpg"},{"id":93956366,"identity":"a382b85e-28f3-4728-8455-30a9494da24f","added_by":"auto","created_at":"2025-10-20 16:11:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1489272,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6457135/v1/d74f8ceb-27ff-4e83-aae9-7d5146a29dc6.pdf"}],"financialInterests":"","formattedTitle":"Evaluation of performance of different machine learning techniques for mapping landslide susceptibility associated with extreme rainfall events in southern Brazil","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eLandslides are the most common risk-triggering geological processes in steep terrain, often causing damage to property and human life (Petley, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Du et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Palmisano; Vitone; Cotecchia; \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Landslides are gravitational movements characterized by a high average velocity of short duration, with a well-defined volume and rupture surface, where the detached material is thrown off the slope, typically resulting in long scars (Guidicini and Nieble, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e1984\u003c/span\u003e; Fernandes and Amaral, 1996; Maciel Filho and Nummer, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe process is triggered by several natural factors such as climate, relief, soil and lithology and anthropogenic factors that act to trigger landslides (Sisay et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), making the study of this phenomenon complex, with a large number of parameters and interrelationships between them (Li \u0026amp; Chen, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Mao et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Pourghasemi \u0026amp; Rahmati, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Mitigation of landslides requires knowledge of the conditions in which these processes are triggered (Du et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), and susceptibility mapping is a critical step in assessing the risk of landslides and has been widely carried out in the scientific community (Quevedo et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Schirmer \u0026amp; Robaina, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sisay et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Z\u0026ecirc;zere et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eLandslide susceptibility mapping identifies areas that are more prone to landslides based on their geomorphological characteristics. This process estimates the spatial probability of landslide occurrence without considering the intensity of the event or its recurrence (Mao et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Z\u0026ecirc;zere et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). As highlighted by Mendon\u0026ccedil;a et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), the study of susceptibility to landslide events is an effective way to mitigate disasters caused by these phenomena and to provide technical support for public decision-making on sustainable municipal land use and occupation.\u003c/p\u003e \u003cp\u003eMost susceptibility studies are carried out quantitatively using statistical techniques (Guzzetti et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Reichenbach et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Sisay et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Z\u0026ecirc;zere et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), but in recent years there has been a rise in the use of machine learning (ML) techniques for mapping susceptibility to landslides (Li \u0026amp; Chen, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Mao et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Quevedo et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Multivariate analysis using ML techniques allows an integrated analysis of geo-environmental parameters in a GIS (Geographic Information System) environment with great precision (Liu et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wang et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Were et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This method has proven superior to traditional statistical-based methods, as it can model high-dimensional and non-linear data sets, allowing for the assessment of complex environmental interactions (Bouramtane et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe central region of the state of Rio Grande do Sul, in southern Brazil, has a recurrence of landslides and has been the subject of several studies on susceptibility, using different methodologies (Cardozo et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Schirmer \u0026amp; Robaina, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). At the end of April and beginning of May 2024, the region was affected by a series of landslides, in which new thresholds of slope stability were exceeded due to an extreme rainfall event, the largest rainfall event ever recorded for the region. During this period, the accumulated volume of rain reached 408.3 mm in four days (INMET, 2024), representing a value around 60% higher than the historical monthly average for the period (257.2 mm). This event reinforces the need for further studies on susceptibility to landslides in the region, considering the increasing frequency of these extreme events in this part of the globe (Sanches et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eML models have the ability to learn the relationship between the occurrence of landslides and the preceding physical conditions, which avoids the use of ready-made models that are already assumed in a structured way and performs better than bivariate and multivariate statistical methods (Dickson \u0026amp; Perry, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Pourghasemi \u0026amp; Rahmati, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). However, each ML algorithm has a different way of learning, produces different results, and demands comparative analysis between models. The objective of this study is to use ML techniques to map landslide susceptibility considering the extreme rainfall event of 2024 in the Cerro Comprido watershed, southern Brazil (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The Random Forest, Neural Network and Scoring Sheet algorithms were used, and the accuracy of the models used was evaluated through validation and comparison of the data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"2. MATERIAL AND METHODS","content":"\u003cp\u003eThe survey of the landslides in the study area was based on the Soobitsky (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) database, carried out using RapidEye images with a spatial resolution of 5m and 3m, generated by the Super Dove sensor (PSB.SD), available in the PlanetExplorer repository (Planet Team, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). A visual inspection was performed on the images from May 5 to 15, 2024, about two weeks after the extreme rainfall event, to identify the landslides by changes in texture and color. Once the landslides were identified, the rupture locations were vectorized, refining and enriching the existing database.\u003c/p\u003e \u003cp\u003eWatersheds were delineated and morphometric parameters were measured using the Digital Terrain Model (DTM) resulting from the Japanese Aerospace Exploration Agency's Advanced Land Observing Satellite (ALOS) mission (JAXA, 2006) with the Phased Array L-band Synthetic Aperture Radar (PALSAR) provided by the Alaska Satellite Facility (ASF) at 12.5 m spatial resolution. The Watershed tool in ArcPro 3.3 software was used to automatically delineate the watersheds in the study area and extract their drainage networks.\u003c/p\u003e \u003cp\u003eInformation on the geoenvironmental characteristics of the study area was obtained from the MDT. Elevation was obtained directly from the MDT, while slope in percent was obtained using the Slope tool. The aspect tool was used to calculate the aspect of the slopes, resulting in 8 different directions. To understand the dynamics of water on the slope, information on the plan and curvature profile of the slopes was obtained using the Curvature tool.\u003c/p\u003e \u003cp\u003eTo analyze the catchment area, the flow accumulation was calculated, which resulted in the number of pixels draining into each pixel in the study area. Considering the area of each pixel as 156.5 m\u0026sup2;, the drained pixels were multiplied by the area of a pixel to obtain the catchment area of each pixel.\u003c/p\u003e \u003cp\u003eOnce the physical variables were defined, the methodology for applying the ML model was applied, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The database for training and testing the machine learning algorithms was created using the points where landslides occurred in the study area (59 points) and the creation of random sampling points in the study area in a 1:1 ratio. Information was extracted from the points in the training file in the form of continuous data on elevation, slope, catchment area, and curvature, and in the form of categorical data on aspect and presence or absence of landslides.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThree supervised machine learning algorithms were used after defining the training database: Artificial Neural Network (ANN), Random Forest (RF) and Scoring Sheet (SC). Neural network-based algorithms simulate the neural networks of the human brain. In the ANN method, \"neurons\" act to solve some complex problems by extracting trends and detecting patterns (Abdolrasol et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Binetti et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These neurons have the ability to learn complex relationships between input and output variables through nonlinear analysis (Pourghasemi \u0026amp; Rahmati, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe Random Forest method uses recursive binary splitting to develop multiple uncorrelated decision trees, using two-thirds of the sample to develop a tree and the remaining sample to assess its predictive accuracy and the importance of environmental covariates (Bouramtane et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Were et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This model is widely used in environmental analysis due to its adaptability to regression and classification tasks (Bouramtane et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Hasanuzzaman \u0026amp; Shit, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Pourghasemi \u0026amp; Rahmati, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The scoring sheet method begins by determining risk scores, which are sparse linear models with integer coefficients that predict risk. In this way, integer values are assigned to the individuals or parameters analyzed, which determine their probability (Liu et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). By defining weights for the different parameters and defining a risk score for intervals between parameter classes, this method is most similar to traditional methods of defining weights and matrices.\u003c/p\u003e \u003cp\u003eThe models were trained and validated using Orange 3.38 software, with the artificial neural network (ANN) model being trained with 100 neurons and a maximum of 500 iterations. The literature suggests that increasing the number of neurons can enhance the performance of the model, particularly in terms of processing time. It is recommended that the minimum number of neurons be equal to or greater than the number of variables (Mendon\u0026ccedil;a et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Yotov et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The choice of 100 neurons as the optimal number was determined through empirical testing, which revealed that higher values did not necessarily lead to improved accuracy. The Random Forest model was trained using a maximum of 200 trees and a minimum data subset size of 5, where the greater the number of decision trees, the more robust the result. The information gain limit was linked to the number of parameters analyzed. The SC model, on the other hand, was trained to use a maximum of 6 decision parameters and to assign a weight of up to 5, which was considered sufficient after systematically testing the data.\u003c/p\u003e \u003cp\u003eFollowing the training of the models, they were applied to the data from the study area, with one point for every 12.5m contained within the watersheds studied. The database contains information on the same geo-environmental parameters as the training file. An index of 0 to 1 probability of landslide occurrence is generated for each pixel and each of the ML models. Values close to 1 indicate higher susceptibility. The data was then spatially discretized for the specified study area, resulting in the definition of three distinct classes: less than 0.5, 0.5 to 0.75, and greater than 0.75.\u003c/p\u003e \u003cp\u003eThe models were validated using cross-validation in the Orange 3.38 software. The following efficiency parameters were extracted: The Area Under the ROC Curve (AUC), which is a graph of the true positive rate against the false negative rate, with values close to 1 indicating good quality; classification accuracy, which is the proportion of correctly classified data; precision, which is the proportion of true positives among the instances classified as positive; and recall, which is the proportion of true positives among all the positive instances in the data.\u003c/p\u003e"},{"header":"3. RESULTS","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 GEOENVIRONMENTAL PARAMETERS\u003c/h2\u003e \u003cp\u003eThe Cerro Comprido watershed comprises the geomorphological units defined as Ramps of Jacu\u0026iacute; colluvium-alluvium deposits, hills in sedimentary rocks, and association of hills and ridges at the edge of the plateau (Schirmer; Robaina; Trentin, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). The main aspects of the Ramps of Jacu\u0026iacute; colluvium-alluvium deposits unit are represented by the low slope inclination, less than 5%, and the altimetric variation between 40 and 90 meters, associated with recent deposits of the main channel of the Jacu\u0026iacute; River. The hills in the sedimentary rocks are defined by a gentle and undulating relief, with a slope of up to 15% and altitudes ranging from 90 to 200 meters, with a predominant bedrock of sandstone of fluvial origin, often friable under the action of weathering and erosion. The association of hills and ridges on the edge of the plateau is made up of volcanic rocks at the top and sandstone with layers of mudstone at the base. There are sections of interbedded sandstone with high-angle crossbeds that mark the overburden contacts where outcrops appear, forming small plateaus between escarpments. This unit is characterized by slopes greater than 15%, with altitudes ranging from 120 to 480 meters (Schirmer; Robaina; Trentin, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe physical characteristics of the study area (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) were analyzed based on elevation (A), slope (B), catchment area (C), aspect (D) and slope curvature (E), variables used for ML processing. The data from the study area indicates a range of elevations from 59 to 545 meters, resulting in an altimetric variation of 486 meters. The most significant class range in terms of area is 200 to 250 meters, where predominate hills in sedimentary rocks.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn terms of slope, the predominant slopes are less than 10%, making them flat to gently undulating. However, there are also slopes in areas of strongly undulating relief, ranging from 20\u0026ndash;45%. In certain points, the presence of slopes with gradients exceeding 75% has been observed, associated with the hillsides close to the top of the watershed.\u003c/p\u003e \u003cp\u003eThe catchment area is a variable representing the area draining into the central pixel, which is an indication of the accumulation of water flows. In this variable, higher values indicate greater concentrations of runoff, which means more water available and more susceptibility to landslide. In the Cerro Comprido basin, catchment areas classified as greater than 10,000 m\u0026sup2; have been identified as exhibiting high flow concentrations, thereby delineating the drainage network.\u003c/p\u003e \u003cp\u003eThe aspect of the slope exerts a significant influence on various environmental factors, including wind direction, precipitation, and solar exposure. In the study area, the primary slope orientations are aligned north and northeast. The west-facing slope, however, is the least significant in terms of the area it covers. Still on the subject of slope characteristics, the analysis of slope curvature for the study area reveals that the major slope plane and profile configuration is classified as a Divergent-Concave type, indicated by positive values, these are areas of water distribution on the hillside. Conversely, the negative values indicate a Convergent-Convex type configuration, what indicates flow concentration and more prone to landslides areas.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 SUSCEPTIBILITY ASSESSMENT\u003c/h2\u003e \u003cp\u003eThe susceptibility map, generated using the ANN model, shows that 7.7% of the study area is defined as medium susceptibility and 7.9% as high susceptibility to landslides (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The areas exhibiting medium and high susceptibility are predominantly concentrated on the slopes that face east, southeast, and northeast. Areas of low susceptibility (\u0026lt;\u0026thinsp;0.5) cover about 85% of the study area, especially in the lower areas with less steep slopes near the main channel.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe analysis of the susceptibility resulting from the model generated with the RF algorithm shows that 18.4% of the areas have medium and high susceptibility (0.5\u0026ndash;0.75 and \u0026gt;\u0026thinsp;0.75), and the remaining 81.5% have low susceptibility. In this model, the areas of high susceptibility occur in similar regions of the map resulting from the application of the ANN, but with less occurrence in the top of the hills. The model indicates susceptibility for slopes facing north and southwest. The RF model attributed the greatest significance to the slope parameter, where the average of the landslide points is 56%, with a score of 0.413, which is the main parameter for defining susceptibility to landslides in this model, followed by altitude (0.186) and catchment area (0.120).\u003c/p\u003e \u003cp\u003eThe map resulting from the SC model shows a greater occurrence of areas with medium and high susceptibility than the previous models, with 24.5% of the study area. When considering the distribution of susceptible areas, this model demonstrated the greatest susceptible area on the slopes north and northwest of the study area, which is the area with the greatest divergence between models and with less occurrence of landslides (Fig.\u0026nbsp;5). In this model, the algorithm identified only the northeast aspect as a distinctive indicator of high susceptibility, assigning more data intervals related to low susceptibility, such as slopes of less than 25% and south-facing slopes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn general, the models considered similar limits for determining areas of medium and high susceptibility (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), especially in the variables of curvature and slope orientation, where only the SC model showed a lower occurrence on the south face, in disagreement with the other methods which indicated the southwest slopes as less susceptible.\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\u003eSummary of the geoenvironmental variables in the pixels classified as medium and high susceptibility in the three models used.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eSlope\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMax.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e213.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e213.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e213.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMin.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e13.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e52.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e56.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e59.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eElevation\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(m)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMax.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e542\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMin.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e341.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e345.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e360.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eCurvature\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMax.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e16.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMin.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-11.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-11.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-11.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eCatchment Area\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(m\u0026sup2;)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMax.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e728437.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e883593.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e883593.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMin.\u003c/em\u003e\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\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eMean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3003.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2745.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2485.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"1\" nameend=\"c6\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eAspect\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eHigher occurrence\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNortheast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNortheast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eNortheast\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eLower occurrence\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSouthwest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSouthwest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eSouth\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 slope variable demonstrated variation between the models in terms of the minimum and average limits for determining susceptible areas. The ANN model identified the lowest slopes as susceptible areas, with a minimum and average of 1.41% and 52.72%, respectively. The RF model considered slopes with gradients above 11% as susceptible areas, with an average of 56%. This value is similar to that defined by the SC model, which indicated even higher values, referring to slopes with gradients above 13% and an average of almost 60%. The findings of the last two models are close to the values indicated in the study conducted by Ribeiro et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), in the same study area, where they highlighted the occurrence of landslides with a higher frequency ratio at gradients from 35 to 40% and reaching the critical point at gradients of 70 to 75%.\u003c/p\u003e \u003cp\u003eIn terms of elevation, the maximum values are similar between the three models, while the minimum values of ANN and SC are closer, and the values presented by the RF model indicate the highest minimum elevations (82m). The average altitude of the susceptibility points was found to be similar between the ANN and RF models (~\u0026thinsp;345m) and around 15 meters higher for the SC model. Despite these differences, the findings from all three models align with the results from previous studies, which have identified the 290 to 440m range as the most susceptible (Ribeiro et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAn analysis of the catchment area indicates that the RF and SC models exhibit equivalent maximum catchment areas, exceeding 883,000 square meters. In contrast, the ANN model exhibits a comparatively diminished maximum catchment area, both in terms of maximum and average values.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 PERFORMANCE OF MACHINE LEARNING MODELS\u003c/h2\u003e \u003cp\u003eThe classification quality of the models used in the study was assessed by cross-validation (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The AUC parameter indicates the ratio of the true positive to the false positive, with a value of 1 indicating perfect classification. All models showed values above 0.8 for this parameter, indicating good accuracy in classifying susceptible areas, with a slightly higher value for the SC model.\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\u003eQuality parameters of the analyzed models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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 \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\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eClassification Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNeural Network\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.851\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.815\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRandom Forest\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.860\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.797\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.797\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.797\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eScoring Sheet\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.788\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.788\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\u003eWhen further quality parameters of the models were evaluated, the ANN exhibited the highest level of accuracy, precision, and recall, with the RF and SC models ranking second and third, respectively. This finding suggests that the ANN model exhibits superior consistency in its results, manifesting as a greater prevalence of accurate classifications and true positives. In contrast, the SC model, despite having a good consistency in the AUC value, performs poorly in the other quality parameters, with greater confusion in the model classification, which has already been observed by (Aguirre-Guti\u0026eacute;rrez et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) in some models with spatial distribution of data. These discrepancies can be attributed to the distinct methodologies employed by machine learning (ML) models in assigning importance to geo-environmental variables.\u003c/p\u003e \u003cp\u003eThe landslide susceptibility value assigned to the rupture points exhibited variation according to the model employed. Higher values were observed in the ANN model (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA), with an average value close to 0.8 and low occurrence of landslides in areas with values below 0.5, indicating good model consistency. When the RF model was evaluated, the results were satisfactory, but with a lower average susceptibility in areas where landslides occurred (0.7), and a concentration of landslides between 0.7 and 0.85. Notably, the SC model exhibits a greater prevalence of sparse susceptibility values, while values greater than 0.7 are indicative of a satisfactory model. The model assigned an average susceptibility below 0.7 (0.67) to the points where landslides occurred, with a higher incidence of landslides in areas with a susceptibility below 0.5.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe frequency density of the susceptibility value for landslides in the study area (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB) was analysed, revealing that the Neural Network model exhibited the greatest number of landslides at high susceptibility, with the peak of the curve very close to 1, indicating good performance in predicting the true values. The Random Forest model exhibited a curve peak at slightly lower values and at a lower density. The Scoring Sheet model, on the other hand, showed a peak density at values close to RF, but with a higher occurrence of values close to 0.5, indicating greater inconsistency in the results.\u003c/p\u003e \u003cp\u003eA comparative analysis of the results obtained from the ANN model, which demonstrated the optimal performance in terms of landslide susceptibility classification within the designated study area when contrasted with alternative models, reveals a notable degree of congruence between the outcomes of the ANN and RF models, as evidenced by Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eA. This observation is consistent with the equivalence between the two models in some scenarios, as observed by Pacheco Quevedo et al. (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). When represented in a scatter plot, the classification points of these two models show a positive correlation of 0.862, with greater similarity at the extremes of high and low susceptibility. Furthermore, an analysis of the determination of susceptibility in locations where landslides have occurred reveals that the landslide points classified with low susceptibility by the ANN model also have low susceptibility in the RF model. However, this is not the case for the points classified with low susceptibility in the RF model, which receive higher values in the ANN model, demonstrating the higher accuracy of the ANN model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe SC model demonstrated a reduced level of correlation with the ANN model (0.743), with notable disparities between landslide points identified as low-susceptibility by the SC model and those identified as high-susceptibility by the ANN model. These disparities are particularly evident in the bottom right corner of the graph, highlighting the distinction in their predictive capacities. Additionally, there are no landslides classified as low susceptibility in the ANN and high in the SC, which is evidence of the greater assertiveness of the neural network model.\u003c/p\u003e \u003cp\u003eIn general, the three models demonstrated satisfactory performance in the assessment of landslide susceptibility in the study area. However, the ANN model demonstrated a better performance considering that it has better precision, accuracy, and consistency of true positives, which has already been observed by (Aditian et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) in the study of landslides in Indonesia. Furthermore, the presence of uncertain points in the ANN model was also observed in the other models, suggesting a necessity for additional variables during the model training process.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. CONCLUSION","content":"\u003cp\u003eThe efficacy of the machine learning techniques in mapping landslide susceptibility was demonstrated by the three models, which were able to identify the areas most susceptible to landslides with a high degree of precision and accuracy, as indicated by an AUC above 0.800. The analysis revealed that slope gradient was a key factor in determining susceptibility, with high slope areas being more susceptible, particularly on northeast and east-facing slopes.\u003c/p\u003e \u003cp\u003eThe ANN model demonstrated the greatest consistency in its results, showing a higher incidence of true positives and greater accuracy in its classification. The model exhibited a reduced occurrence of landslides in areas of low susceptibility and a high concentration of landslides in areas of high and medium susceptibility, demonstrating its superior performance and lack of overfitting and demonstrating its potential to be applied to the database with great accuracy.\u003c/p\u003e \u003cp\u003eDespite the extensive utilization of the random forest model in the literature for analyzing complex environmental phenomena, such as landslides, this model demonstrated slightly lower precision and accuracy than the artificial neural network. The spatial distribution exhibited by the Random Forest model was comparable to that of the ANN model, however, a greater number of landslides occurred in areas exhibiting low susceptibility. This finding suggests that the Random Forest model performs less effectively than the ANN model. The SC model, which is rarely employed in this type of study, performed poorly in determining susceptibility despite having an operating logic similar to traditional susceptibility mapping methods.\u003c/p\u003e \u003cp\u003eThe data analyzed in this study refers to an extreme rainfall event where the geomorphic thresholds are different from the standards expected for landslide occurrence, making it difficult to determine susceptibility using traditional methods. However, the ML models demonstrated high accuracy in determining the spatial distribution of susceptibility, providing a faster and more accurate analysis.\u003c/p\u003e \u003cp\u003eIn addition, it is essential to evaluate the performance of the models taking into account a larger number of geo-environmental parameters, such as lithology and soil types, at appropriate scales, which could increase the accuracy of the predictions and amplify the differences between the models. Further studies are required for the study area, taking into account normal precipitation conditions and conducting inventories over a time series. This will enable the assessment of the recurrence and return time of events. This approach will facilitate the refinement of susceptibility mapping, thereby enabling the identification of regions where risk may be overestimated in analyses based exclusively on extreme events. The incorporation of long-term data is also crucial, as it serves to reduce uncertainties, thereby ensuring more robust predictions for applications in territorial planning, disaster prevention and risk management.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFUNDING DECLARATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Brazilian National Council for Scientific and Technological Development (CNPq), grant number 302449/2022-1 and the Foundation for Research Support of the State of Rio Grande do Sul (FAPERGS), grant number 24/2551- 0002134-5.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOMPETING INTEREST\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRINUTIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Lucas Krein Rademann and Lucas da Silva Ribeiro. The first draft of the manuscript was written by Lucas Krein Rademann and Romario Trentin and Luis Eduardo de Souza Robaina commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbdolrasol, M. G. M., Hussain, S. M. S., Ustun, T. S., Sarker, M. R., Hannan, M. A., Mohamed, R., Ali, J. A., Mekhilef, S., \u0026amp; Milad, A. (2021). Artificial Neural Networks Based Optimization Techniques: A Review. Electronics, 10(21), 2689. https://doi.org/10.3390/electronics10212689\u003c/li\u003e\n\u003cli\u003eAditian, A., Kubota, T., \u0026amp; Shinohara, Y. (2018). Comparison of GIS-based landslide susceptibility models using frequency ratio, logistic regression, and artificial neural network in a tertiary region of Ambon, Indonesia. Geomorphology, 318, 101\u0026ndash;111. https://doi.org/10.1016/j.geomorph.2018.06.006\u003c/li\u003e\n\u003cli\u003eAguirre-Guti\u0026eacute;rrez, J., Carvalheiro, L. G., Polce, C., Van Loon, E. E., Raes, N., Reemer, M., \u0026amp; Biesmeijer, J. C. (2013). Fit-for-Purpose: Species Distribution Model Performance Depends on Evaluation Criteria \u0026ndash; Dutch Hoverflies as a Case Study. PLoS ONE, 8(5), e63708. https://doi.org/10.1371/journal.pone.0063708\u003c/li\u003e\n\u003cli\u003eBinetti, M. S., Massarelli, C., \u0026amp; Uricchio, V. F. (2024). Machine Learning in Geosciences: A Review of Complex Environmental Monitoring Applications. Machine Learning and Knowledge Extraction, 6(2), 1263\u0026ndash;1280. https://doi.org/10.3390/make6020059\u003c/li\u003e\n\u003cli\u003eBouramtane, T., Hilal, H., Rezende-Filho, A. T., Bouramtane, K., Barbiero, L., Abraham, S., Valles, V., Kacimi, I., Sanhaji, H., Torres-Rondon, L., De Castro, D. D., Vieira Santos, J. D. C., Ouardi, J., Beqqali, O. E., Kassou, N., \u0026amp; Morarech, M. (2022). Mapping Gully Erosion Variability and Susceptibility Using Remote Sensing, Multivariate Statistical Analysis, and Machine Learning in South Mato Grosso, Brazil. Geosciences, 12(6), 235. https://doi.org/10.3390/geosciences12060235 \u003c/li\u003e\n\u003cli\u003eCardozo, G. L., Zanandrea, F., Michel, G. P., \u0026amp; Kobiyama, M. (2021). Invent\u0026aacute;rio de movimentos de massa na bacia hidrogr\u0026aacute;fica do rio Mascarada/RS. Ci\u0026ecirc;ncia e Natura, 43, e31. https://doi.org/10.5902/2179460X43594\u003c/li\u003e\n\u003cli\u003eDickson, M. E., \u0026amp; Perry, G. L. W. (2016). Identifying the controls on coastal cliff landslides using machine-learning approaches. Environmental Modelling \u0026amp; Software, 76, 117\u0026ndash;127. https://doi.org/10.1016/j.envsoft.2015.10.029\u003c/li\u003e\n\u003cli\u003eDu, G., Zhang, Y., Iqbal, J., Yang, Z., \u0026amp; Yao, X. (2017). Landslide susceptibility mapping using an integrated model of information value method and logistic regression in the Bailongjiang watershed, Gansu Province, China. \u003cem\u003eJournal of Mountain Science\u003c/em\u003e, 14(2), 249\u0026ndash;268. https://doi.org/10.1007/s11629-016-4126-9\u003c/li\u003e\n\u003cli\u003eFernandes, N. F.; Amaral, C. P. Movimentos de massa: uma abordagem geol\u0026oacute;gico-geomorfol\u0026oacute;gica. (2003). In: Guerra, A. J. T. and Cunha, S. B. (org.) Geomorfologia e Meio Ambiente. 4\u0026ordf;. ed. Rio de Janeiro: Bertrand. cap. 3, p. 123-194.\u003c/li\u003e\n\u003cli\u003eGuidicini, G.; Nieble, C. M. (1984). Estabilidade de Taludes Naturais e de Escava\u0026ccedil;\u0026atilde;o. S\u0026atilde;o Paulo: Edgard Bl\u0026uuml;cher. p.196. \u003c/li\u003e\n\u003cli\u003eGuzzetti, F., Carrara, A., Cardinali, M., \u0026amp; Reichenbach, P. (1999). Landslide hazard evaluation: A review of current techniques and their application in a multi-scale study, Central Italy. \u003cem\u003eGeomorphology\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(1\u0026ndash;4), 181\u0026ndash;216. https://doi.org/10.1016/S0169-555X(99)00078-1\u003c/li\u003e\n\u003cli\u003eHasanuzzaman, M., \u0026amp; Shit, P. (2024). Assessment of gully erosion susceptibility using four data-driven models AHP, FR, RF and XGBoosting machine learning algorithms. Natural Hazards Research, S2666592124000362. https://doi.org/10.1016/j.nhres.2024.05.001\u003c/li\u003e\n\u003cli\u003eINMET - Instituto Nacional de Meteorologia. Eventos Extremos de Abril de 2024 no Brasil, Bras\u0026iacute;lia-DF, (2024). Dispon\u0026iacute;vel em: https://portal.inmet.gov.br/uploads/notastecnicas/Nota-Extremos-revisado.pdf. Acesso em: 29 jul. 2024.\u003c/li\u003e\n\u003cli\u003eJAXA - Japan Aerospace Exploration Agency ALOS PALSAR L1.0, (2006). Dispon\u0026iacute;vel: https://asf.alaska.edu/datasets/daac/alos-palsar/. Acesso em: 01 jun. 2024.\u003c/li\u003e\n\u003cli\u003eLi, Y., \u0026amp; Chen, W. (2019). Landslide Susceptibility Evaluation Using Hybrid Integration of Evidential Belief Function and Machine Learning Techniques. \u003cem\u003eWater\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(1), 113. https://doi.org/10.3390/w12010113\u003c/li\u003e\n\u003cli\u003eLiu, G., Arabameri, A., Santosh, M., \u0026amp; Nalivan, O. A. (2023). Optimizing machine learning algorithms for spatial prediction of gully erosion susceptibility with four training scenarios. Environmental Science and Pollution Research, 30(16), 46979\u0026ndash;46996. https://doi.org/10.1007/s11356-022-25090-2\u003c/li\u003e\n\u003cli\u003eLiu, J., Zhong, C., Li, B., Seltzer, M., \u0026amp; Rudin, C. (2022). FasterRisk: Fast and Accurate Interpretable Risk Scores (No. arXiv:2210.05846). arXiv. https://doi.org/10.48550/arXiv.2210.05846\u003c/li\u003e\n\u003cli\u003eMaciel Filho, C. L.; Nummer, A. V. (2011). Introdu\u0026ccedil;\u0026atilde;o \u0026agrave; geologia de engenharia. Santa Maria: Editora da UFSM. \u003c/li\u003e\n\u003cli\u003eMao, Y., Mwakapesa, D. S., Li, Y., Xu, K., Nanehkaran, Y. A., \u0026amp; Zhang, M. (2022). Assessment of landslide susceptibility using DBSCAN-AHD and LD-EV methods. \u003cem\u003eJournal of Mountain Science\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(1), 184\u0026ndash;197. https://doi.org/10.1007/s11629-020-6491-7\u003c/li\u003e\n\u003cli\u003eMendon\u0026ccedil;a, R. R., Oliveira, G. G. D., \u0026amp; Tornquist, C. G. (2024). Landslide Susceptibility Modeling Using Artificial Neural Networks in the Municipality of Joinville, southern Brazil. \u003cem\u003eRevista Brasileira de Geomorfologia\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e(4). https://doi.org/10.20502/rbg.v25i4.2513\u003c/li\u003e\n\u003cli\u003ePacheco Quevedo, R., Guasselli, L. A., Garcia De Oliveira, G., \u0026amp; Chimelo Ruiz, L. F. (2019). Modelagem de \u0026aacute;reas suscet\u0026iacute;veis a movimentos de massa: Avalia\u0026ccedil;\u0026atilde;o comparativa de t\u0026eacute;cnicas de amostragem, aprendizagem de m\u0026aacute;quina e modelos digitais de eleva\u0026ccedil;\u0026atilde;o. Geosciences = Geoci\u0026ecirc;ncias, 38(3), 781\u0026ndash;795. https://doi.org/10.5016/geociencias.v38i3.14019\u003c/li\u003e\n\u003cli\u003ePalmisano, F., Vitone, C., \u0026amp; Cotecchia, F. (2018). Assessment of Landslide Damage to Buildings at the Urban Scale. J\u003cem\u003eournal of Performance of Constructed Facilities\u003c/em\u003e, 32(4), 04018055. https://doi.org/10.1061/(ASCE)CF.1943-5509.0001201\u003c/li\u003e\n\u003cli\u003ePetley, D.; (2012). Global patterns of loss of life from landslides. \u003cem\u003eGeology\u003c/em\u003e; 40 (10): 927\u0026ndash;930. doi: https://doi.org/10.1130/G33217.1\u003c/li\u003e\n\u003cli\u003ePlanet Team. Planet Application Program Interface: In Space for Life on Earth. San Francisco, CA: Planet Labs, 2025, p. n/a. Dispon\u0026iacute;vel em: https://api.planet.com. Acesso em: feveiro de 2025.\u003c/li\u003e\n\u003cli\u003ePourghasemi, H. R., \u0026amp; Rahmati, O. (2018). Prediction of the landslide susceptibility: Which algorithm, which precision? CATENA, 162, 177\u0026ndash;192. https://doi.org/10.1016/j.catena.2017.11.022\u003c/li\u003e\n\u003cli\u003eQuevedo, R. P., Oliveira, G. G. D., \u0026amp; Guasselli, L. A. (2020). Mapeamento de Suscetibilidade a Movimentos de Massa a partir de Redes Neurais Artificiais. \u003cem\u003eAnu\u0026aacute;rio do Instituto de Geoci\u0026ecirc;ncias - UFRJ\u003c/em\u003e, \u003cem\u003e43\u003c/em\u003e(2). https://doi.org/10.11137/2020_2_128_138\u003c/li\u003e\n\u003cli\u003eReichenbach, P., Rossi, M., Malamud, B. D., Mihir, M., \u0026amp; Guzzetti, F. (2018). A review of statistically-based landslide susceptibility models. Earth-Science Reviews, 180, 60\u0026ndash;91. https://doi.org/10.1016/j.earscirev.2018.03.001\u003c/li\u003e\n\u003cli\u003eRibeiro, L. da S., Rademann, L. K., Robaina, L. E. de S., Schnorr, G. G., \u0026amp; Trentin, R. (2025). An\u0026aacute;lise da rela\u0026ccedil;\u0026atilde;o entre a ocorr\u0026ecirc;ncia de deslizamentos e os atributos do relevo no evento extremo de precipita\u0026ccedil;\u0026atilde;o no sul do Brasil, maio de 2024. Sociedade \u0026amp; Natureza, 37(1). https://doi.org/10.14393/SN-v37-2025-74694\u003c/li\u003e\n\u003cli\u003eSanches, F., Verdum, R., Fisch, G., Gass, S. L. B., \u0026amp; Rocha, V. M. (2019). Extreme Rainfall Events in the Southwest of Rio Grande do Sul (Brazil) and Its Association with the Sandization Process. American Journal of Climate Change, 08(04), 441\u0026ndash;453. https://doi.org/10.4236/ajcc.2019.84024\u003c/li\u003e\n\u003cli\u003eSchirmer, G. J.; Robaina, L. E. S. Trentin, R. (2013). Unidades geomorfol\u0026oacute;gicas em munic\u0026iacute;pios da Quarta Col\u0026ocirc;nia do Rio Grande do Sul. Geografa Ensino \u0026amp; Pesquisa, v. 17, n.2 p. 199-212, maio/ago. https://doi.org/10.5902/223649949244\u003c/li\u003e\n\u003cli\u003eSchirmer, G. J., \u0026amp; Robaina, L. E. de S. (2023). Mapeamento de \u0026aacute;reas suscept\u0026iacute;veis a desastres naturais da Quarta Col\u0026ocirc;nia-RS com o base no zoneamento geoambiental. \u003cem\u003eGeografia Ensino \u0026amp; Pesquisa\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e, e67900\u0026ndash;e67900. https://doi.org/10.5902/2236499467900\u003c/li\u003e\n\u003cli\u003eSisay, T., Tesfaye, G., Jothimani, M., Reda, T. M., \u0026amp; Tadese, A. (2024). Landslide susceptibility mapping using combined geospatial, FR and AHP models: A case study from Ethiopia\u0026rsquo;s highlands. \u003cem\u003eDiscover Sustainability\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(1), 474. https://doi.org/10.1007/s43621-024-00730-4\u003c/li\u003e\n\u003cli\u003eSOOBITSKY, R. Manual inventory of landslides in Brazil, 2024-04-29. (2024). NASA GSFC. Juang CS, Stanley TA, and Kirschbaum DB. (2019). Using citizen science to expand the global map of landslides: Introducing the Cooperative Open Online Landslide Rep.\u003c/li\u003e\n\u003cli\u003eWang, Y., Zhang, Y., \u0026amp; Chen, H. (2023). Gully erosion susceptibility prediction in Mollisols using machine learning models. Journal of Soil and Water Conservation, 78(5), 385\u0026ndash;396. https://doi.org/10.2489/jswc.2023.00019\u003c/li\u003e\n\u003cli\u003eWere, K., Kebeney, S., Churu, H., Mutio, J. M., Njoroge, R., Mugaa, D., Alkamoi, B., Ng\u0026rsquo;etich, W., \u0026amp; Singh, B. R. (2023). Spatial Prediction and Mapping of Gully Erosion Susceptibility Using Machine Learning Techniques in a Degraded Semi-Arid Region of Kenya. Land, 12(4), 890. https://doi.org/10.3390/land12040890.\u003c/li\u003e\n\u003cli\u003eYotov, K., Hadzhikolev, E., \u0026amp; Hadzhikoleva, S. (2020). Determining the Number of Neurons in Artificial Neural Networks for Approximation, Trained with Algorithms Using the Jacobi Matrix. \u003cem\u003eTEM Journal\u003c/em\u003e, 1320\u0026ndash;1329. https://doi.org/10.18421/TEM94-02\u003c/li\u003e\n\u003cli\u003eZ\u0026ecirc;zere, J. L., Pereira, S., Melo, R., Oliveira, S. C., \u0026amp; Garcia, R. A. C. (2017). Mapping landslide susceptibility using data-driven methods. \u003cem\u003eScience of The Total Environment\u003c/em\u003e, \u003cem\u003e589\u003c/em\u003e, 250\u0026ndash;267. https://doi.org/10.1016/j.scitotenv.2017.02.188\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"natural-hazards","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nhaz","sideBox":"Learn more about [Natural Hazards](https://www.springer.com/journal/11069)","snPcode":"11069","submissionUrl":"https://submission.nature.com/new-submission/11069/3","title":"Natural Hazards","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Landslides, Geomorphology, Machine Learning, Brazil, Extreme Rainfall","lastPublishedDoi":"10.21203/rs.3.rs-6457135/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6457135/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLandslides represent a primary geological process that triggers hazards in steep slope areas, affecting infrastructure and sometimes causing loss of life. Susceptibility mapping is a critical component in the mitigation of landslide-induced disasters, providing technical expertise to support public policy decisions. In May 2024, a significant rainfall event occurred in Southern Brazil, leading to multiple landslides and the transgression of previously established limits of slope stability. Hence, it became necessary to study the landslide susceptibility of this region. Given the complex nature of the landslide process, machine learning tools were used to map the landslide susceptibility using Random Forest (RF), Artificial Neural Network (ANN) and Scoring Sheet (SC) models to compare the performance of these models. The geo-environmental parameters of slope, elevation, slope orientation, catchment area, and curvature were used to train the models. All three models were effective in mapping susceptibility, but the ANN model exhibited the most consistent results, demonstrating a higher frequency of true positives and enhanced accuracy in its classification. The analysis revealed that slope gradient was a key factor in determining susceptibility, with high slope areas being more susceptible, particularly on northeast and east-facing slopes. The data analyzed in this study refers to an extreme rainfall event where the geomorphic thresholds are different from the standards expected for landslide occurrence, making it difficult to determine susceptibility using traditional methods. However, the ML models demonstrated high accuracy in determining the spatial distribution of susceptibility, providing a faster and more accurate analysis.\u003c/p\u003e","manuscriptTitle":"Evaluation of performance of different machine learning techniques for mapping landslide susceptibility associated with extreme rainfall events in southern Brazil","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-12 09:21:05","doi":"10.21203/rs.3.rs-6457135/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2025-05-20T10:05:09+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-07T14:51:27+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"Natural Hazards","date":"2025-05-03T15:15:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-16T11:53:03+00:00","index":"","fulltext":""},{"type":"submitted","content":"Natural Hazards","date":"2025-04-15T13:39:51+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"natural-hazards","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nhaz","sideBox":"Learn more about [Natural Hazards](https://www.springer.com/journal/11069)","snPcode":"11069","submissionUrl":"https://submission.nature.com/new-submission/11069/3","title":"Natural Hazards","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"e0ccb6d2-40c0-455f-a3b9-ee9407918a7c","owner":[],"postedDate":"May 12th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-10-20T16:08:08+00:00","versionOfRecord":{"articleIdentity":"rs-6457135","link":"https://doi.org/10.1007/s11069-025-07694-2","journal":{"identity":"natural-hazards","isVorOnly":false,"title":"Natural Hazards"},"publishedOn":"2025-10-13 15:58:36","publishedOnDateReadable":"October 13th, 2025"},"versionCreatedAt":"2025-05-12 09:21:05","video":"","vorDoi":"10.1007/s11069-025-07694-2","vorDoiUrl":"https://doi.org/10.1007/s11069-025-07694-2","workflowStages":[]},"version":"v1","identity":"rs-6457135","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6457135","identity":"rs-6457135","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}