Classification of agroforestry systems by photo-interpretation of satellite imagery

Classification of agroforestry systems by photo-interpretation of satellite imagery | 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 Classification of agroforestry systems by photo-interpretation of satellite imagery Ouadya Tahiri, Damien Beillouin, Patrice Dumas, Rémi Prudhomme, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6000362/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Effective and large-scale monitoring of agroforestry (AF) systems is essential to assess the environmental benefits of agroforestry and support sustainable land management strategies. However, a standardized method for classifying these systems using satellite imagery is still missing. Here, we present a novel operational framework to classify agroforestry systems into three categories—Alley cropping, Scattered agroforestry, and Hedgerows—and to distinguish these systems from Cropland without trees, Natural trees, and Orchards. The proposed procedure relies on a classification tree based on photo-interpretation of satellite imagery. The accuracy and robustness of this classification tree were evaluated by five interpreters across 300 agroforestry and non-agroforestry plots spanning all continents. Results show that the classification tree accurately distinguishes agroforestry categories from one another and from non-agroforestry systems, with an overall accuracy ranging from 0.75 to 0.81 depending on the interpreter. After eliminating the interpreters’ errors, the potential classification accuracy increases to 0.86. While hedgerows were accurately classified in most cases with an omission error of 2% and no commission error (0%), the study revealed challenges in differentiating between Alley cropping and Orchards which were frequently confounded. Similarly, plots with Scattered agroforestry were also misclassified as Natural trees leading to a commission error of 19% for this class. Despite these limitations, the proposed classification tree represents a valuable tool for large-scale monitoring of agroforestry systems. Future adaptations of this framework could address regional specificities, further improving its applicability and accuracy. agroforestry classification tree classification errors photo-interpretation Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Agroforestry systems, which integrate trees within agricultural landscapes (Nair 1993 ) are increasingly recognized for their multifaceted contributions to biodiversity conservation (Torralba et al. 2016 ), soil health enhancement (Zhu et al. 2020 ), and climate change mitigation (Chapman et al. 2020 ). However, these systems are highly heterogeneous (Nair 1993 ), with diverse spatial arrangements and components (e.g., crops, pasture animals, trees), leading to considerable variability in their ecological and economic performance. In particular, studies have demonstrated that carbon sequestration and yield values can strongly differ across different agroforestry system types (Beillouin et al. 2020 ; Cardinael et al. 2018 ; Ivezić et al. 2021 ). The impacts of agroforestry systems are also highly context-dependent, varying by factors such as soil types, climatic conditions, and local management practices (Kuyah et al. 2019 ). The global distribution of these systems remains uneven, with significant regional disparities in their prevalence and adoption. This highlights the need for accurate, high-resolution spatial maps that capture the diversity of agroforestry systems and account for the contextual variability in their environmental impacts. In the absence of such maps, and clear differentiation between agroforestry practices, current global land use models fail to adequately represent these systems, resulting in substantial uncertainties regarding their environmental benefits (Terasaki Hart et al. 2023 ). The spatial heterogeneity of agroforestry systems complicates their mapping from satellite imagery (Escobar-López et al. 2024 ; Ndao et al. 2021 ), as some agroforestry systems may be confounded with other land use types, such as orchards or natural vegetation (Bolívar-Santamaría & Reu 2021 ; Lesiv et al. 2022 ). Given the growing global emphasis on the ambition of agroforestry practices (Rosenstock et al. 2019 ) and its environmental significance, developing reliable and scalable classification systems based on remote sensing techniques is critical. Such classification systems are essential for accurately mapping agroforestry systems, tracking their expansion, monitoring their ecological and economic outcomes, and assessing their role in climate change mitigation across diverse regions, in particular, when coupled with robust measurement, reporting, and verification (MRV) frameworks (Batjes et al. 2024 ). Photo interpretation refers to the process of identifying, analyzing, and extracting information from aerial or satellite imagery based on visible features (Lillesand et al. 2015 ). This technique has been extensively applied across various fields, such as forestry (Brovelli et al. 2020 ; Lesiv et al. 2022 ), agriculture (Belgiu & Csillik 2018 ; Hussain et al. 2023 ), and archeology (Doneus & Doneus 2024 ). A key advantage of this approach is its ability to enable remote monitoring without the need for on-site presence. With the rise of machine learning and the growing accessibility of satellite data, photo interpretation is increasingly used to produce extensive databases, often including large volumes of labeled points across vast geographical areas (Stanimirova et al. 2023 ). Such databases can subsequently be used to train and test machine-learning models to generate land-use maps (Brandt & Stolle 2021 ; Lesiv et al. 2022 ). Photo-interpretation has been widely used to detect trees outside forested areas (Brandt & Stolle, 2021 ; Sarti et al. 2021 ). However, these approaches are often insufficient to distinguish different types of agroforestry systems, and agroforestry from other tree-dominated land use, such as savannahs. Lesiv et al. ( 2022 ) used visual interpretation of satellite imagery to distinguish agroforestry from other tree-based ecosystems as part of their global forest management database initiative. While their approach improves georeferenced data availability on agroforestry, the absence of clear criteria defining the agroforestry class reduces the reproducibility of the approach. In addition, in their final classification, the different agroforestry systems were not distinguished. Consequently, identifying agroforestry systems from satellite imagery remains a significant challenge, necessitating the development of an operational and standardized classification framework. Such a framework must address the inherent ambiguity between agroforestry systems with scattered trees (e.g. parklands) and non-agroforestry natural landscapes (e.g., savannahs). Additionally, the delineation of agroforestry systems remains poorly defined, particularly when these systems co-exist, or are adjacent to other land use types. Agroforestry units have been delineated in some studies using segmentation techniques, such as object-based image analysis (OBIA) combined with other geospatial data (Ndao et al. 2021 ). However, these techniques can become computationally exhaustive when applied to large-scale studies, as they require significantly longer processing times (Guirado et al. 2017 ). This study aims to develop a robust and systematic framework for differentiating and classifying agroforestry systems using photo-interpretation by analyzing the structural characteristics of these systems. The approach is based on the premise that distinct agroforestry systems can be characterized by unique spatial configurations of trees and associated crops or pastures. Our specific objectives are threefold: i) identify key land cover characteristics that can reliably differentiate agroforestry from non-agroforestry systems using photo-interpretation, ii) develop an operational classification tree based on these characteristics, enabling the discrimination of various agroforestry systems from other land use types, iii) evaluate this classification tree using a set of satellite images covering multiple land use types and geographical regions. By achieving these goals, we aim to provide a globally applicable and effective tool for large-scale mapping of agroforestry systems. 2. Materials and methods We followed a three-step approach to classify agroforestry systems based on photo-interpretation. In step 1, we conducted a literature review in order to identify key land use characteristics that could serve as reliable indicators to distinguish different agroforestry systems. In step 2, we developed a streamlined and generalizable classification tool by translating these indicators into simple decision rules. In step 3, we validated the resulting classification tool by testing it with different interpreters using 300 land use reference plots. 2.1. Step 1. Identification of land use characteristics to classify agroforestry and non-agroforestry systems Our objective is to categorize the agroforestry systems, as defined by Nair et al. (2009) and refined by Cardinael et al. ( 2018 ), based on the spatial arrangement of trees both within-plot and in the surrounding area. Some agroforestry systems, however, were excluded from the scope of this study, particularly those characterized by closed canopy systems (e.g., shaded perennial crops, and multistrata systems), which are difficult to distinguish from forests using image-based techniques. Additionally, sequential agroforestry systems, such as improved fallows, were not considered due to their temporal dynamics (Friday et al. 1999 ), which makes them unsuitable for image-based classification. Based on these considerations, we focused on three categories of agroforestry systems : (1) Alley cropping, where trees are arranged in parallel rows; (2) Hedgerows, where trees surround agricultural fields; and (3) Scattered agroforestry systems, in which trees are dispersed within pasturelands or croplands (e.g., Parklands , Dehesa , and Montados ). Tree presence and their spatial arrangement alone may be insufficient to reliably differentiate between agroforestry and some other land-use types. This is especially challenging in distinguishing alley cropping from orchards as they often have similar tree arrangements (i.e., parallel rows). To address this challenge, we assumed that tree row spacing is narrower in orchards compared to alley cropping systems, driven by the need for higher fruit yield (Gomez-del-Campo et al. 2017 ; Rallo et al. 2013 ). Another potential source of confusion arises when distinguishing between scattered agroforestry systems and savannah or some other forms of dispersed natural tree cover as these systems share the same spatial configuration and other ecological and dynamic characteristics (FAO 1999 ). To tackle this issue, we assumed that natural systems are located further from human structures compared to scattered agroforestry systems, which tend to be situated closer to human settlements (FAO 1999 ). Grazing is one of the most mobile and irregular activities that characterize these agroforestry systems. However, it is difficult to define a spatial threshold as it varies according to several factors (Liao 2018 ). In this case, we start from the assumption that pastoralists would first utilize patches of rangelands with better forage quality and lower travel costs near the settlement (Liao et al. 2017 ). The observation scale is critical for classification, and landscape heterogeneity may lead to misclassification. For instance, hedgerows may be confused with large cropland fields surrounded by patches of trees. Beyond a given distance from the hedgerow, a plot should be classified as cropland without trees rather than agroforestry. To avoid misclassification, we cut out the landscape in observation units assumed to be representative for the purpose of agroforestry classification. To handle the issues presented in this section, the classification integrates information from three spatial scales: 30m x 30m Observation Unit : This unit serves as the primary level for classifying individual sites. We selected this size because of the radius of influence of agroforestry trees on variables such as soil parameters (Cardinael et al. 2015 ; Pardon et al. 2017 ), crop yield (Roupsard et al. 2020 ), and soil organism abundance (Vaupel et al. 2023 ) typically ranges between 10m and 30m. Additionally, this scale aligns with widely used remote sensing imagery data, such as Landsat 8 & 9 with 30m-spatial resolution ( www.usgs.gov/landsat-missions ). At this scale, tree presence, shadow patterns, crown shapes, and surrounding land use characteristics are visually assessed to initiate the classification process. 10,000m² Land Unit (1 ha) : This scale is used to calculate tree density, serving as a threshold for distinguishing agroforestry systems with scattered trees from denser forest systems. A threshold of 50 trees per hectare was applied, in accordance with the recommendations of the French Agroforestry Association (AFAF) (2014) to establish an agroforestry system. The chosen 1 ha area scale allows the interpreter to estimate tree density relatively quickly, by visual inspection and is large enough to avoid measurement errors at low density. 2 km Radius : This broader scale evaluates the proximity of human-made structures to distinguish natural scattered tree systems (e.g., savannah) from agroforestry systems with scattered trees (e.g., parklands). The 2 km distance threshold reflects the average daily walking distance of pastoralists from a “base camp” to a grazing site (Liao et al. 2017 ). In this context, the "base camp" is typically any human-made structure, such as a house, agricultural plot, or road. Roads, in particular, are key indicators, as they enhance accessibility and enable the development of agroforestry systems (Bennett 2017 ). This 2 km distance threshold between the “base camp” and the observation unit is used to distinguish scattered agroforestry systems and natural ecosystems. Based on this analysis, we identify and select the following land use characteristics for classification: (i) the presence or absence of trees; (ii) the presence of geometric patterns indicative of human-managed landscapes; (iii) the spatial arrangement of trees (trees in parallel rows, tree edges, or dispersed trees); (iv) the proximity to of human structures (e.g., houses, roads); and (v) tree density, which helps differentiate between various tree-dominated systems. These characteristics collectively provide a robust foundation for distinguishing agroforestry systems from other land-use types while ensuring generalizability and practical applicability. 2.2. Step 2. Definition of the classification tree To categorise land use systems into distinct classes, we followed the method proposed by ​Caumont and Ivanaj ( 2017 ) based on similarity profiles using previously defined variables. This approach focuses on identifying the most discriminative variables and grouping systems according to three principles: i) Exclusivity —each system is assigned to a single group; ii) Homogeneity — systems within each group share similar characteristics, minimizing intra-group variability while maximizing inter-group differences; and iii) Parsimony —a minimal number of groups is used to summarize the diversity of systems. The resulting classification tree (Fig. 1 ) uses a hierarchical structure with seven decision nodes, allowing for systematic discrimination between land use types. It classifies 30x30m land units into seven distinct land use types: Alley cropping, Hedgerows, Scattered agroforestry, Natural trees, Cropland without trees, Orchards, and others. The first classification criterion in the classification process is the presence of trees in the 30x30m observation unit. Trees are identified visually based on their shadow, crown shape, and surrounding land use characteristics (Brandt & Stolle 2021 ). The next classification steps are determined by six other closed questions (Yes/No) formulated based on the variables identified in the previous step, such as the trees' spatial arrangement, tree density, proximity to human activities, and other visual variables. Ultimately, we get both a user-friendly and globally applicable classification tree. Its simplicity, rooted in a limited set of decision rules, allows non-specialists to systematically and independently classify seven categories of land use types. 2.3. Step3. Assessing the classification tree 2.3.1. Georeferenced sites To assess the accuracy of the decision tree, we compiled a database of georeferenced sites, each labeled and corresponding to a specific land use system. To collect these labeled sites, a search was performed on three databases, Scopus, Web of Science, and Google Scholar, for studies reporting georeferenced data of the selected systems between 2016 and 2024. Additional sources such as reports from funded research projects (e.g., AGFORWARD) were consulted. These searches were conducted using keywords such as alley cropping, tree-based intercropping, hedgerows, and scattered agroforestry. Additionally, we collected data from three pre-existing databases for croplands without trees (Laso Bayas et al. 2017 ; Lesiv et al. 2019 ) and natural trees (Lesiv et al. 2022 ). The complete list of data sources is available in Annex A. These labeled sites serve as ground-truth data, having been validated by experts. Each study was reviewed to ensure compliance with the following selection criteria: i) Temporal alignment - only points collected between 2016 and 2024 were retained to match the aerial imagery used for photo interpretation. ii) System specificity : Studies with vague or ambiguous descriptions of agroforestry systems or those potentially conflating different system types were excluded. iii) Georeferencing precision : Sites with imprecise coordinates were discarded unless this information could be corrected based on the annexes of the papers or any other indicator that could help to geolocate the point. Georeferenced points were cross-verified using Google Earth ( https://earth.google.com/ ) and Bing Maps ( https://www.bing.com/maps ) to ensure that each site met the basic characteristic of the specified system, for instance, we did not include a point as agroforestry or natural trees in the validation dataset when there is no visible tree on the satellite imagery, as this would distort the interpretation and accuracy of our framework. As a result, 134 studies were retained and distributed as follows: 45 focused on Alley cropping, 21 on hedgerows, 35 on Scattered AF, and 35 on orchards. From the selected studies, we were able to identify 150 georeferenced agroforestry sites (50 sites per agroforestry system) and 150 georeferenced non-agroforestry sites to evaluate the accuracy of the classification tree. We tried to set a threshold of 50 points per system to guarantee a balanced database and ensure that rare classes are well represented. We supplemented the classes originally with less than 50 sites with additional georeferenced images obtained from AGFORWARD reports ( www.agforward.eu ) to reach the threshold. 2.3.2. Implementation of the classification tree The 300 georeferenced sites were uploaded to Collect Earth Online ( www.collect.earth ) , a user-friendly platform designed for satellite imagery interpretation (Saah et al. 2019 )​. For each site, a 30x30m observation unit was defined at the center of the georeferenced coordinates. To facilitate the implementation of the classification tree (Fig. 2 ), we delineated the three observation units presented in step 1: i) 30x30m Square : The primary unit of classification. ii) 100x100m Square : Used to calculate tree density. iii) 2km Radius : Used to account for proximity to human-made structures. The sequential classification rules derived from the classification tree were programmed into the platform. Five interpreters, including one administrator, who are also the authors of this paper, applied these rules to classify the satellite imagery. A comprehensive training guide was provided to introduce the interpreters to the platform (see Annex B) and the classification protocol. To ensure unbiased assessments, the interpreters (other than the administrator) did not have access to the true categories of the systems represented in the images and worked independently without communication. Subsequently, the administrator compiled and compared their classifications (Annex C) against reference data for each site. 2.3.3. Metrics used for evaluation The performance of the interpreters’ classifications was evaluated according to four metrics: Overall accuracy, Commission error, Omission error, and Cohen’s Kappa (McHugh 2012 ). Additionally, the consistency among the five interpreters was assessed using the Fleiss’ Kappa metric (Fleiss 1971 ). These metrics were used to quantify both individual (i.e., per interpreter) and collective (aggregated) classification performances. The accuracy is defined as the proportion of labeled sites correctly classified. From this, the error rate (%) was derived as: (1-Accuracy)*100. Cohen’s Kappa (k) was used to measure the agreement between each interpreter’s classification and the true labeled classes, accounting for chance agreement. It is calculated as: $$\:k\:=\:\frac{{p}_{o\:}-\:{p}_{e}}{1\:-\:{p}_{e}}$$ where \(\:{p}_{o\:}\) is the observed agreement, representing the proportion of points where the interpreter's classification matched the true label $$\:{p}_{o}\:=\:\frac{{\sum\:}_{c=1}^{C}T{P}_{c}}{N}$$ with \(\:T{P}_{c}\) the number of true positives for class c , \(\:C\:\) is the total number of classes (here, C = 6), and \(\:N\:\) the total number of data points. \(\:{p}_{e}\) is the expected agreement by chance, calculated as: $$\:{p}_{e}\:=\:{\sum\:}_{c=1}^{C}\left(\frac{Number\:of\:points\:classified\:in\:class\:c\:*\:True\:number\:of\:point\:in\:class\:c}{{N}^{2}}\right)$$ Fleiss’s kappa, used for assessing consistency among the five interpreters, modifies the calculations of \(\:{p}_{o}\:\) and \(\:{p}_{e}\:\) accounts for multiple raters: $$\:{p}_{o}\:=\frac{1}{N\:\times\:R\:\left(R\:-1\right)}{\sum\:}_{i\:=\:1}^{N}{\sum\:}_{c=1}^{C}{n}_{ic}\left({n}_{ic}-1\right)$$ $$\:{p}_{e}\:=\:{\sum\:}_{c={1}_{\:}^{\:}}^{C}{p}_{c}^{2}$$ where \(\:\:{p}_{c}\:=\:\frac{1}{N\times\:R}{\sum\:}_{i=1}^{N}{n}_{ic}\) is the proportion of all assignments to class c. \(\:R\:\) : is the number of interpreters (here, R = 5). \(\:{n}_{ic}\) : is the number of interpreters who assigned point i to class c. While Kappa values provide an overall performance measure, they do not describe the error levels associated with the different classes. To identify strengths and weaknesses in classification for individual classes, we calculated the following metrics: Omission Error ( \(\:Eo\) ) represents the proportion of sites belonging to a given class c incorrectly classified in another class: \(\:\:Eo\:=1-(\:F{N}_{c}\:/(T{P}_{c}\:+F{N}_{c})\) ) Commission Error ( \(\:Ec\) ) represents the proportion of sites incorrectly allocated to a class c : $$\:Ec\:=1-(T{P}_{c}\:/\:(\:T{P}_{c}\:+F{P}_{c}\left)\right)$$ For instance, misclassifying an alley cropping point as an orchard results in an omission error for the "Alley cropping" class and a commission error for the "Orchard" class. The metrics and their 95% confidence intervals were computed based on bootstrap techniques to account for uncertainty and variability in performance measures. 2.3.4. Interpreters’ errors vs. Classification tree errors To disentangle sources of misclassification, two error types were distinguished: Interpreter’s Error : This type of error occurs when a site that was misclassified by at least one interpreter (excluding the administrator) is correctly classified by the administrator (interpreter 1) after a careful implementation of the classification tree. To be as certain as possible that the application of the tree by the interpreter was incorrect, the administrator goes through the tree and checks each decision node to determine the precise step at which the interpreter made a mistake, in order to objectivize the error. The administrator can also use the reported site type to understand why the interpreter made a mistake. Classification Tree Error : This type of error occurs when all interpreters (including the administrator) misclassify a site even though they implement the classification tree correctly. This differentiation allowed us to isolate errors attributable to the classification tree itself from those resulting from interpreters’ skill variability and thus calculate the potential accuracy of the decision tree. The metrics described above were recalculated for each error type, for the interpreter errors and the classification tree errors, separately. 3. Results 3.1. Geographical distribution of the labeled sites The 300 labeled sites used to evaluate our classification method provide comprehensive coverage across all continents, ensuring a broad and balanced geographical distribution (Fig. 3 a). Although Europe accounts for the highest proportion (35%) of sites, all agroforestry systems — alley cropping, hedgerows, and scattered agroforestry— and tree non-agroforestry systems —orchards, croplands without trees, and natural trees — are represented in all continents. Some land use classes are more frequently represented in some regions than others. For instance, in Russia, the “Natural trees” class is predominant as forests cover a large share of this country (FAO, 2020 ). 3.2. Accuracy of the interpreters’ classifications Misclassifications are geographically dispersed, with no clear regional pattern, as shown in the map presented in Fig. 3 b. This map depicts the distribution of well-classified and misclassified sites by Interpreter 1, based on photo-interpretation using the Collect Earth Online platform. Overall, the classification accuracy ranges from 0.75 to 0.81, depending on the interpreters, with an average of 0.79 (Table.1). These results indicate that the least accurate interpreter was able to classify 75% (95%CI=[70%, 80%]) of sites correctly, while the most accurate achieved 81% (95%CI=[76%, 85%]) accuracy. The coefficient of variation (CV) across the five interpreters is 2.89%, reflecting low dispersion around the mean accuracy and low variability among interpreters. Part of the between interpreters variation can be attributed to certain interpreters not assessing all labeled sites due to difficulties in accessing or visualizing the images. Consequently, the samples of sites evaluated by the five interpreters are slightly different. While accuracy requirements depend on the context (Story & Congalton 1986 ), an accuracy of 79% is considered sufficient for large-scale analyses focusing on broad patterns, as suggested by Charles and Olson ( 2008 ). Additionally, all interpreters’ accuracy values are at least 4.5 times higher than the 17% accuracy expected from random classification with six classes. This high classification performance is further confirmed by the values of Cohen’s kappa (Table 1 ) ranging from 0.70 (95%CI= [0.64, 0.76]) to 0.77 (95%CI= [0.71, 0.81]). According to Landis and Koch ( 1977b ), these values reflect a “ substantial” agreement between the interpreters and the reference data. Consistent with the CV of accuracy values, the coefficient of variation of the kappa values among the five interpreters is lower than 5% (3.53%). Additionally, Fleiss' kappa of 0.86 reveals a strong overall agreement among all interpreters (Landis & Koch 1977a ). Table 1 Overall accuracy and Cohen’s kappa with their 95% confidence intervals for each interpreter, along with their average (Avg) and coefficient of variation (CV). Values of n indicate the number of sites assessed by each interpreter Interpreter Overall accuracy 95% CI interval Cohen's kappa 95% interval 1 (n = 283) 0.81 0.76–0.85 0.77 0.71–0.82 2 (n = 269) 0.81 0.76–0.85 0.77 0.71–0.82 3 (n = 263) 0.75 0.70–0.80 0.70 0.64–0.76 4 (n = 264) 0.8 0.75–0.84 0.76 0.7–0.81 5 (n = 263) 0.78 0.73–0.83 0.74 0.68–0.79 Avg 0.79 - 0.75 - CV (%) 2.89 - 3.53 - Table 2 presents the classification errors for each land-use class. Orchards exhibit the highest error rate, averaging 53.3% across interpreters, with individual rates ranging from 48–66%. Scattered agroforestry has the second highest error rate, averaging 38%, followed by Alley cropping with an average error rate of 25.5%. In contrast, Cropland without trees, Hedgerows, and Natural trees show much lower average error rates of 10.1%, 2.8%, and 2.0% respectively. While error rates vary among interpreters, no single interpreter achieves the lowest error across all classes. For example, interpreter 1 minimizes error rates for Alley cropping, Scattered agroforestry, Croplands without trees, and Natural trees, whereas interpreter 5 performs best for Hedgerows and Orchards. Table 2 Error rate ((1-Accuracy) *100) per land-use class and interpreter, along with their average (Avg) and coefficient of variation (CV). Values of n indicate the number of sites assessed by each interpreter or for each land use class Interpreter Alley cropping Hedgerows Scattered AF Orchards Croplands WT Natural trees 1 (n = 283) 17.39 2.17 28.26 65.96 4.17 0 2 (n = 269) 22.22 4.88 43.18 48.89 4.17 2.00 3 (n = 263) 42.50 2.27 46.34 48.84 8.70 6.12 4 (n = 264) 19.51 4.55 37.50 54.76 10.00 0 5 (n = 263) 25.64 0 34.15 47.83 23.26 2.00 Avg 25.5 2.8 37.9 53.3 10.1 2.0 CV (%) 35.2 64.2 17 12.8 70 110.4 3.3. Potential accuracy of the classification tree The accuracy metrics were calculated a second time on all labeled sites after removing interpretation errors. This second run isolates errors attributable solely to the tree's limitations in distinguishing between classes, removing the influence of interpreters' misinterpretations. Consequently, the resulting error rates represent the theoretical lower bounds of classification errors achievable with perfect implementation free from interpreter-related inaccuracies. After the removal of interpreters’ errors, the classification tree demonstrates improved performance, achieving an average accuracy of 0.86 and a Cohen’s kappa of 0.83 (Table 3 ). Nevertheless, approximately 15% of classification errors persist, reflecting inherent limitations of the decision tree rather than errors stemming from interpretation. The small fluctuations of the accuracy and Cohen’s kappa between the five groups of sites (coefficient of variation < 1.6%) are not due to differences between experts but from variations in the sets of images assessed. As mentioned above, some interpreters did not evaluate all 300 labeled sites, leading to slight fluctuations in the sample sizes and associated metrics. Table 3 Accuracy and Cohen’s kappa with 95% confidence intervals, along with their average (Avg) and coefficient of variation (CV) after excluding the interpreter-related error. The samples of sites considered in Tables 1 and 2 were carefully checked by Interpreter 1 to assess the residual error Sample of sites Overall accuracy 95% interval Cohen's kappa 95% interval 1 (n = 267) 0.85 0.81–0.90 0.82 0.77–0.87 2 (n = 250) 0.87 0.83–0.91 0.84 0.79–0.89 3 (n = 235) 0.84 0.79–0.89 0.81 0.75–0.86 4 (n = 244) 0.86 0.82–0.91 0.84 0.79–0.89 5 (n = 237) 0.87 0.82–0.91 0.84 0.79–0.89 Avg 0.86 - 0.83 - CV (%) 1.36 - 1.52 - Figure 4 illustrates the performance of the classification tree for each land use class after removing the interpreters-related errors. The off-diagonal cells of the confusion matrix (Fig. 4 a) reveal two primary sources of misclassification, i.e. confusion between Orchards and Alley Cropping, and between Scattered AF and Natural trees. The commission and omission errors (Fig. 4 b) provide further insight into these misclassifications. Specifically, 36% of the sites are misclassified as Alley cropping, predominantly from Orchards class (Fig. 4 a), while 58% of the Orchard sites are misclassified into other classes, primarily Alley cropping. In contrast, Alley Cropping shows lower omissions errors (14%), and the confusion matrix indicates that the misclassification of Alley Cropping as Orchards occurs less frequently than the reverse. This pattern indicates a higher propensity for Orchards to be misclassified compared to Alley Cropping, even when interpreter-related errors are eliminated. Concerning the confusion between Scattered AF and Natural Trees, Scattered AF has an omission error of 23% and a commission error of 10%, whereas Natural Trees has no omission error and a commission error of 18%. These results indicate that nearly one out of four sites of Scattered AF is misclassified, while all Natural trees sites are correctly identified. This pattern suggests that the confusion mainly arises from misclassifying Scattered trees as Natural trees, rather than the opposite. Merging the Alley Cropping and Orchards classes results in a notable improvement in the classification tree's performance. The overall accuracy increased from 0.86 to 0.94 (reducing the error by 6%), while Cohen’s kappa rises from 0.83 to 0.92, whichLandis and Koch ( 1977b ) described as an “almost perfect” agreement. The omission error of the new combined class Alley Cropping/Orchard drops to 4%, and the commission error is reduced to 1.7%. 4. Discussion We developed a classification tree to identify three agroforestry systems - Hedgerows , Scattered agroforestry, and Alley cropping - and three non-agroforestry systems — Cropland without trees , Natural trees , and Orchards — using a photo-interpretation approach. This novel method relies on straightforward classification rules, making the proposed approach accessible to interpreters with basic knowledge of agroforestry and agriculture. Moreover, our classification tree can be easily applied using the free and open-source Collect Earth Online platform (Saah et al., 2019 ). Despite its simplicity, the classification tree demonstrates high performance, achieving robust results in distinguishing between agroforestry and non-agroforestry systems. Indeed, based on a dataset of 300 labeled sites representing agroforestry and non-agroforestry systems across all continents, we demonstrated that the lower limit of classification error rate achievable with perfectly trained users would be 14%. By merging the classes “Alley Cropping” and “Orchards,” we further reduced the error rate to approximately 10%. This balance between ease of use and reliability makes it an innovative tool for large-scale agroforestry mapping and monitoring. 4.1. Sources of error and perspectives for improvement The classification error rate with non-expert ranged from 19–25%, depending on the interpreter. Errors associated with interpreters are manageable and could be minimized through targeted training. Previous studies on photo interpretation have shown intensive training sessions can significantly reduce photo interpretation errors (Lesiv et al., 2022 ). Importantly, interpreters' errors often vary across individuals and images, suggesting that a collaborative, consensus-based approach could further enhance classification consistency. In contrast, classification errors inherent to the decision tree itself are systemic, arising from its structural limitations. A low rate of residual errors persists despite eliminating interpretation errors and reflects the intrinsic difficulty of distinguishing between certain land-use classes. A major source of error involved the frequent misclassification of orchards as alley cropping (more frequently than the reverse), likely due to the similarity between their spatial arrangement as both feature trees planted in parallel rows. This similarity stems from shared agronomic principles, such as optimal tree spacing for maximizing productivity and facilitating mechanization (Gomez-del-Campo et al. 2017 ; Menzel & Le Lagadec 2017 ). Furthermore, orchards and alley cropping can correspond to different stages of land use change. For instance, adding crops to orchards can create alley cropping systems, while removing crops can revert alley cropping to orchards (Sánchez-Navarro et al. 2023 ). This dynamic nature underlines the necessity for high temporal and spatial resolution imagery to accurately monitor land-use transitions. However, widely used platforms like Google Earth or Microsoft Bing are limited by the infrequency of image updates. It reduces their utility for tracking rapid land-use change, as they capture images over extended intervals, limiting their ability to monitor fine temporal variations. In addition, there might be some issues in the labeling of alley cropping and orchard because some studies may refer to alley cropping as orchard agroforestry (Zayani et al. 2023 ) making it unclear whether to attribute a given plot to orchard or alley cropping. Another challenge emerged in distinguishing scattered agroforestry from natural forests, particularly in regions where dense tree cover coincides with human activity, such as silvopastoral systems beneath tree canopies (Coble et al. 2020 ; Öllerer et al. 2019 ), This is confirmed by the commission error in the 'Natural Trees' class, where 19% of the points (Fig. 4 b) classified as natural trees were actually scattered agroforestry. Even though the tree density criterion prevents a correct classification in those cases, this criterion remains important, as scattered agroforestry systems are often found in transformed forest landscapes influenced by human activity, with low tree density, such as the Parklands in Africa (Taïbi et al. 2019 ), the Spanish Dehesas (Gaspar et al. 2009 ), and the Montados in Portugal (Pinto-Correia et al. 2011 ). In addition to the tree’s density, our classification tree relies on proximity to human settlements to differentiate agroforestry from natural forests by defining the 2km distance radius. This approach aligns with the method of Lesiv et al. ( 2022 ), who define natural forests as areas without detectable disturbances within a specific spatial distance (with a radius of approximately 0.5km). 4.2. Implications for future applications Collecting validation data remains a challenging step, largely due to a shortage of labeled and georeferenced data. Detailed information on agroforestry systems is often lacking, even in scientific publications. Moreover, precise coordinates are not always provided, reducing the number of usable validation plots. Our classification tree offers a new operational tool for future applications, such as collecting georeferenced and labeled data in areas including one or several types of agroforestry systems. Such data are essential for assessing the impact of agroforestry on biomass, crop production, biodiversity, and carbon stocks. It constitutes an interesting and less costly alternative to on-site data observation. While remote sensing and classification trees are increasingly used for large-scale agroforestry mapping and monitoring, they cannot fully replace the need for detailed field-based observations, especially in complex agroforestry systems. Closed-canopy systems, such as shaded perennial crop systems (e.g., cacao, coffee, and tropical tree crop agroforestry), present particular challenges for remote sensing due to their complex vertical structures and the dynamic interactions between trees and crops (Bolívar-Santamaría & Reu 2024 ). These systems, with their dense canopies, often elude accurate classification through traditional photo-interpretation techniques, particularly when distinguishing between different crops or when the canopy structure conceals underlying biodiversity. To address these limitations, future applications would benefit from an integrated approach that combines remote sensing with on-the-ground validation efforts. Advanced technologies such as LiDAR (Light Detection and Ranging) can significantly improve the characterization of complex canopy structures by providing detailed three-dimensional data on vegetation height, density, and stratification (Thapa et al. 2023 ). When combined with high-resolution satellite imagery, LiDAR enables more accurate differentiation between agroforestry systems and other land uses, especially in landscapes with dense and complex vegetation. This hybrid approach is particularly advantageous for systems like shaded agroforestry, where subtle distinctions between trees and crops can be difficult to capture with optical imagery alone. LiDAR’s ability to detect vertical structure and fine-scale vegetation characteristics offers a more comprehensive view of these systems, which are often inadequately represented through traditional remote sensing methods (Dagar et al. 2024 ). Another promising direction for enhancing agroforestry monitoring is the use of crowdsourced data. Citizen science initiatives, where local communities contribute ground-based observations on land-use practices and environmental conditions, can play a pivotal role in validating satellite-based classifications and expanding the geographic coverage of agroforestry data collection (L.C. & L. 2023). By incorporating local knowledge into global datasets, these initiatives offer the potential for more granular insights into agroforestry systems, particularly in remote or understudied areas. Additionally, crowdsourcing is a scalable, cost-effective strategy for data collection, particularly in regions with limited access to high-end remote sensing technologies. It allows for broader participation and facilitates the rapid accumulation of valuable data, enhancing the accuracy of agroforestry mapping. 5. Conclusion Despite advancements in photo-interpretation platforms and the availability of high-resolution satellite imagery, operational systems are still lacking for accurately classifying agroforestry systems. Here, we developed a simple method to classify agroforestry systems into three categories—alley cropping, scattered agroforestry, and hedgerows—and to distinguish these systems from non-agroforestry land uses, i.e. cropland without trees, natural trees, and orchards. The proposed method relies on a classification tree that can be easily implemented by photo-interpretation of satellite imagery. Based on 300 observations of agroforestry and non-agroforestry plots covering all continents, we showed that its classification error rate ranged from 19 to 25%, depending on the interpreter. After eliminating the interpreter’s errors, the classification error rate was reduced to 14%. The main source of error stemmed from confusion between alley cropping and orchards and, to a lower extent, between scattered trees and natural trees. The fusion of these pairs of categories decreases the classification error to 6%. Looking ahead, the proposed classification tree offers significant potential for large-scale mapping of agroforestry systems. By improving the accuracy and scalability of agroforestry monitoring, this tool could contribute to more effective assessments of the ecosystem services and climate adaptation benefits provided by agroforestry systems, especially in the context of global climate change. Declarations Competing Interests The authors have no conflicts of interest to declare that are relevant to the content of this article. Funding This work benefited from the French state aid managed by the ANR under the "Investissements d'avenir" programme with the reference ANR-16-CONV-0003 as a partial stipend alongside with a partial stipend from CIRAD. The funding sources had no role in the study design, data collection, analysis, or decision to publish. Author Contribution All authors wrote and reviewed the main manuscript textO.T. analyzed data and prepared figures Data Availability Data is provided within the manuscript or supplementary information files References AFAF (2014) L’AGROFORESTERIE EN 10 QUESTIONS. https://www.agroforesterie.fr/wp-content/uploads/2022/07/afaf-agroforesterie-en-10-questions.pdf .Accessed June 2024 Batjes, N. H., Ceschia, E., Heuvelink, G. B. M., Demenois, J., le Maire, G., Cardinael, R., Arias-Navarro, C., & van Egmond, F. (2024). Towards a modular, multi-ecosystem monitoring, reporting and verification (MRV) framework for soil organic carbon stock change assessment. Carbon Management , 15 (1). https://doi.org/10.1080/17583004.2024.2410812 Beillouin, D., Ben-Ari, T., & Makowski, D. (2020). Erratum: Evidence map of crop diversification strategies at the global scale (Environmental Research Letters (2019) (14) (123001) DOI: 10.1088/1748-9326/ab4449). In Environmental Research Letters (Vol. 15, Issue 1). Institute of Physics Publishing. https://doi.org/10.1088/1748-9326/ab5ffb Belgiu, M., & Csillik, O. (2018). Sentinel-2 cropland mapping using pixel-based and object-based time-weighted dynamic time warping analysis. Remote Sensing of Environment , 204 , 509–523. https://doi.org/10.1016/j.rse.2017.10.005 Bennett, V. J. (2017). Effects of Road Density and Pattern on the Conservation of Species and Biodiversity. Current Landscape Ecology Reports , 2 (1), 1–11. https://doi.org/10.1007/s40823-017-0020-6 Bolívar-Santamaría, S., & Reu, B. (2021). Detection and characterization of agroforestry systems in the Colombian Andes using sentinel-2 imagery. Agroforestry Systems , 95 (3), 499–514. https://doi.org/10.1007/s10457-021-00597-8 Bolívar-Santamaría, S., & Reu, B. (2024). Assessing canopy structure in Andean (agro)forests using 3D UAV remote sensing. Agroforestry Systems , 98 (5), 1225–1241. https://doi.org/10.1007/s10457-023-00865-9 Brandt, J., & Stolle, F. (2021). A global method to identify trees outside of closed-canopy forests with medium-resolution satellite imagery. International Journal of Remote Sensing , 42 (5), 1713–1737. https://doi.org/10.1080/01431161.2020.1841324 Brovelli, M. A., Sun, Y., & Yordanov, V. (2020). Monitoring Forest Change in the Amazon Using Multi-Temporal Remote Sensing Data and Machine Learning Classification on Google Earth Engine. ISPRS International Journal of Geo-Information , 9 (10), 580. https://doi.org/10.3390/ijgi9100580 Cardinael, R., Chevallier, T., Barthès, B. G., Saby, N. P. A., Parent, T., Dupraz, C., Bernoux, M., & Chenu, C. (2015). Impact of alley cropping agroforestry on stocks, forms and spatial distribution of soil organic carbon — A case study in a Mediterranean context. Geoderma , 259–260 , 288–299. https://doi.org/10.1016/j.geoderma.2015.06.015 Cardinael, R., Umulisa, V., Toudert, A., Olivier, A., Bockel, L., & Bernoux, M. (2018). Revisiting IPCC Tier 1 coefficients for soil organic and biomass carbon storage in agroforestry systems. Environmental Research Letters , 13 (12). https://doi.org/10.1088/1748-9326/aaeb5f Caumont, D., & Ivanaj, S. (2017). Chapitre 5. Comment construire et valider une typologie. In Analyse des données (pp. 161–198). Dunod. https://shs.cairn.info/analyse-des-donnees--9782100758517-page-161?lang=fr Chapman, M., Walker, W. S., Cook-Patton, S. C., Ellis, P. W., Farina, M., Griscom, B. W., & Baccini, A. (2020). Large climate mitigation potential from adding trees to agricultural lands. Global Change Biology , 26 (8), 4357–4365. https://doi.org/10.1111/gcb.15121 Charles, E., & Olson, J. (2008). IS 80% ACCURACY GOOD ENOUGH? Pecora 17 – The Future of Land Imaging…Going Operational . https://www.asprs.org/a/publications/proceedings/pecora17/0026.pdf Accessed November 2024 Coble, A. P., Contosta, A. R., Smith, R. G., Siegert, N. W., Vadeboncoeur, M., Jennings, K. A., Stewart, A. J., & Asbjornsen, H. (2020). Influence of forest-to-silvopasture conversion and drought on components of evapotranspiration. Agriculture, Ecosystems & Environment , 295 , 106916. https://doi.org/10.1016/j.agee.2020.106916 Dagar, J. C., Gupta, S. R., & Dimobe, K. (2024). Agroforestry from a Global Perspective: Recent Developments, Technological Advancements, and Emerging Research Trends. In Agroforestry Solutions for Climate Change and Environmental Restoration (pp. 3–32). Springer Nature Singapore. https://doi.org/10.1007/978-981-97-5004-7_1 Doneus, N., & Doneus, M. (2024). Intensively Cultivated Roman Villae Estates: Case Study of Medulin Bay (Istria, Croatia). Archaeological Prospection , 31 (2), 75–97. https://doi.org/10.1002/arp.1933 Escobar-López, A., Castillo-Santiago, M. Á., Mas, J. F., Hernández-Stefanoni, J. L., & López-Martínez, J. O. (2024). Identification of coffee agroforestry systems using remote sensing data: a review of methods and sensor data. Geocarto International , 39 (1). https://doi.org/10.1080/10106049.2023.2297555 FAO. (1999). CHAPTER I: AGROFORESTRY PARKLAND SYSTEMS. In Agroforestry parklands in sub-Saharan Africa (FAO conservation guide) . Food & Agriculture Organization of the United Nations (FAO). https://www.fao.org/4/x3940e/X3940E11.htm Accessed 2024 FAO. (2020). Global Forest Resources Assessment 2020 . https://www.fao.org/interactive/forest-resources-assessment/2020/en/. Accessed November 2024 Fleiss, J. L. (1971). Measuring nominal scale agreement among many raters. Psychological Bulletin , 76 (5), 378–382. https://doi.org/10.1037/h0031619 Friday, K. S. ., Drilling, M. Elmo., & Garrity, D. P. . (1999). Imperata grassland rehabilitation using agroforestry and assisted natural regeneration . International Centre for Research in Agroforestry, Southeast Asian Regional Research Programme. Gaspar, P., Mesías, F. J., Escribano, M., & Pulido, F. (2009). Sustainability in Spanish Extensive Farms (Dehesas): An Economic and Management Indicator-Based Evaluation. Rangeland Ecology & Management , 62 (2), 153–162. https://doi.org/10.2111/07-135.1 Gomez-del-Campo, M., Connor, D. J., & Trentacoste, E. R. (2017). Long-term Effect of Intra-Row Spacing on Growth and Productivity of Super-High Density Hedgerow Olive Orchards (cv. Arbequina). Frontiers in Plant Science , 8 . https://doi.org/10.3389/fpls.2017.01790 Guirado, E., Tabik, S., Alcaraz-Segura, D., Cabello, J., & Herrera, F. (2017). Deep-learning Versus OBIA for Scattered Shrub Detection with Google Earth Imagery: Ziziphus lotus as Case Study. Remote Sensing , 9 (12), 1220. https://doi.org/10.3390/rs9121220 Hussain, M. H., Abuhani, D. A., Khan, J., ElMohandes, M., Zualkernan, I., & Ali, T. (2023). A Light-Weight Cropland Mapping Model Using Satellite Imagery. Sensors , 23 (15), 6729. https://doi.org/10.3390/s23156729 Ivezić, V., Yu, Y., & Werf, W. van der. (2021). Crop Yields in European Agroforestry Systems: A Meta-Analysis. In Frontiers in Sustainable Food Systems (Vol. 5). Frontiers Media S.A. https://doi.org/10.3389/fsufs.2021.606631 Kuyah, S., Whitney, C. W., Jonsson, M., Sileshi, G. W., Öborn, I., Muthuri, C. W., & Luedeling, E. (2019). Agroforestry delivers a win-win solution for ecosystem services in sub-Saharan Africa. A meta-analysis. Agronomy for Sustainable Development , 39 (5). https://doi.org/10.1007/s13593-019-0589-8 Landis, J. R., & Koch, G. G. (1977a). An Application of Hierarchical Kappa-type Statistics in the Assessment of Majority Agreement among Multiple Observers. Biometrics , 33 (2), 363. https://doi.org/10.2307/2529786 Landis, J. R., & Koch, G. G. (1977b). The Measurement of Observer Agreement for Categorical Data. Biometrics , 33 (1), 159–174. https://doi.org/10.2307/2529310 Laso Bayas, J. C., Lesiv, M., Waldner, F., Schucknecht, A., Duerauer, M., See, L., Fritz, S., Fraisl, D., Moorthy, I., McCallum, I., Perger, C., Danylo, O., Defourny, P., Gallego, J., Gilliams, S., Akhtar, I. ul H., Baishya, S. J., Baruah, M., Bungnamei, K., … Wilson, J. W. (2017). A global reference database of crowdsourced cropland data collected using the Geo-Wiki platform. Scientific Data , 4 (1), 170136. https://doi.org/10.1038/sdata.2017.136 L.C., V. D., & L., F. (2023). One Million Voices Global Review: A review and analysis of existing citizen science initiatives and projects supporting agroecology and agroecological transitions . Center for International Forestry Research (CIFOR). https://doi.org/10.17528/cifor-icraf/008921 Lesiv, M., Laso Bayas, J. C., See, L., Duerauer, M., Dahlia, D., Durando, N., Hazarika, R., Kumar Sahariah, P., Vakolyuk, M., Blyshchyk, V., Bilous, A., Perez‐Hoyos, A., Gengler, S., Prestele, R., Bilous, S., Akhtar, I. ul H., Singha, K., Choudhury, S. B., Chetri, T., … Fritz, S. (2019). Estimating the global distribution of field size using crowdsourcing. Global Change Biology , 25 (1), 174–186. https://doi.org/10.1111/gcb.14492 Lesiv, M., Schepaschenko, D., Buchhorn, M., See, L., Dürauer, M., Georgieva, I., Jung, M., Hofhansl, F., Schulze, K., Bilous, A., Blyshchyk, V., Mukhortova, L., Brenes, C. L. M., Krivobokov, L., Ntie, S., Tsogt, K., Pietsch, S. A., Tikhonova, E., Kim, M., … Fritz, S. (2022). Global forest management data for 2015 at a 100 m resolution. Scientific Data , 9 (1). https://doi.org/10.1038/s41597-022-01332-3 Liao, C. (2018). Quantifying multi-scale pastoral mobility: Developing a metrics system and using GPS-Tracking data for evaluation. Journal of Arid Environments , 153 , 88–97. https://doi.org/10.1016/j.jaridenv.2017.12.012 Liao, C., Clark, P. E., DeGloria, S. D., & Barrett, C. B. (2017). Complexity in the spatial utilization of rangelands: Pastoral mobility in the Horn of Africa. Applied Geography , 86 , 208–219. https://doi.org/10.1016/j.apgeog.2017.07.003 Lillesand, T., W. Kiefer, R., & Chipman, J. (2015). Concepts and Foundations of Remote Sensing 1. In WILEY (Ed.), Remote Sensing and Image Interpretation (7th Edition, pp. 59–84). McHugh, M. L. (2012). Interrater reliability: the kappa statistic. Biochemia Medica , 276–282. https://doi.org/10.11613/BM.2012.031 Menzel, C. M., & Le Lagadec, M. D. (2017). Can the productivity of mango orchards be increased by using high-density plantings? Scientia Horticulturae , 219 , 222–263. https://doi.org/10.1016/j.scienta.2016.11.041 Nair, P. K. R. . (1993). An introduction to agroforestry. Kluwer Academic Publishers in cooperation with International Centre for Research in Agroforestry. https://apps.worldagroforestry.org/Units/Library/Books/PDFs/32_An_introduction_to_agroforestry.pdf?n Accessed 2024 Ndao, B., Leroux, L., Gaetano, R., Diouf, A. A., Soti, V., Bégué, A., Mbow, C., & Sambou, B. (2021). Landscape heterogeneity analysis using geospatial techniques and a priori knowledge in Sahelian agroforestry systems of Senegal. Ecological Indicators , 125 . https://doi.org/10.1016/j.ecolind.2021.107481 Öllerer, K., Varga, A., Kirby, K., Demeter, L., Biró, M., Bölöni, J., & Molnár, Z. (2019). Beyond the obvious impact of domestic livestock grazing on temperate forest vegetation – A global review. Biological Conservation , 237 , 209–219. https://doi.org/10.1016/j.biocon.2019.07.007 Pardon, P., Reubens, B., Reheul, D., Mertens, J., De Frenne, P., Coussement, T., Janssens, P., & Verheyen, K. (2017). Trees increase soil organic carbon and nutrient availability in temperate agroforestry systems. Agriculture, Ecosystems & Environment , 247 , 98–111. https://doi.org/10.1016/j.agee.2017.06.018 Pinto-Correia, T., Ribeiro, N., & Sá-Sousa, P. (2011). Introducing the montado, the cork and holm oak agroforestry system of Southern Portugal. Agroforestry Systems , 82 (2), 99–104. https://doi.org/10.1007/s10457-011-9388-1 Rallo, L., Barranco, D., Castro‐García, S., Connor, D. J., Gómez del Campo, M., & Rallo, P. (2013). High‐Density Olive Plantations. In Horticultural Reviews Volume 41 (pp. 303–384). Wiley. https://doi.org/10.1002/9781118707418.ch07 Ramachandran Nair, P. K., Mohan Kumar, B., & Nair, V. D. (2009). Agroforestry as a strategy for carbon sequestration. Journal of Plant Nutrition and Soil Science , 172 (1), 10–23. https://doi.org/10.1002/jpln.200800030 Rosenstock, T. S., Wilkes, A., Jallo, C., Namoi, N., Bulusu, M., Suber, M., Mboi, D., Mulia, R., Simelton, E., Richards, M., Gurwick, N., & Wollenberg, E. (2019). Making trees count: Measurement and reporting of agroforestry in UNFCCC national communications of non-Annex I countries. Agriculture, Ecosystems and Environment , 284 . https://doi.org/10.1016/j.agee.2019.106569 Roupsard, O., Audebert, A., Ndour, A. P., Clermont-Dauphin, C., Agbohessou, Y., Sanou, J., Koala, J., Faye, E., Sambakhe, D., Jourdan, C., le Maire, G., Tall, L., Sanogo, D., Seghieri, J., Cournac, L., & Leroux, L. (2020). How far does the tree affect the crop in agroforestry? New spatial analysis methods in a Faidherbia parkland. Agriculture, Ecosystems and Environment , 296 . https://doi.org/10.1016/j.agee.2020.106928 Saah, D., Johnson, G., Ashmall, B., Tondapu, G., Tenneson, K., Patterson, M., Poortinga, A., Markert, K., Quyen, N. H., San Aung, K., Schlichting, L., Matin, M., Uddin, K., Aryal, R. R., Dilger, J., Lee Ellenburg, W., Flores-Anderson, A. I., Wiell, D., Lindquist, E., … Chishtie, F. (2019). Collect Earth: An online tool for systematic reference data collection in land cover and use applications. Environmental Modelling & Software , 118 , 166–171. https://doi.org/10.1016/j.envsoft.2019.05.004 Sánchez-Navarro, V., Martínez-Martínez, S., Acosta, J. A., Almagro, M., Martínez-Mena, M., Boix-Fayos, C., Díaz-Pereira, E., Temnani, A., Berrios, P., Pérez-Pastor, A., & Zornoza, R. (2023). Soil greenhouse gas emissions and crop production with implementation of alley cropping in a Mediterranean citrus orchard. European Journal of Agronomy , 142 , 126684. https://doi.org/10.1016/j.eja.2022.126684 Sarti, M., Ciolfi, M., Lauteri, M., Paris, P., & Chiocchini, F. (2021). Trees outside forest in Italian agroforestry landscapes: detection and mapping using sentinel-2 imagery. European Journal of Remote Sensing , 54 (1), 609–623. https://doi.org/10.1080/22797254.2021.1986678 Stanimirova, R., Tarrio, K., Turlej, K., McAvoy, K., Stonebrook, S., Hu, K. T., Arévalo, P., Bullock, E. L., Zhang, Y., Woodcock, C. E., Olofsson, P., Zhu, Z., Barber, C. P., Souza, C. M., Chen, S., Wang, J. A., Mensah, F., Calderón-Loor, M., Hadjikakou, M., … Friedl, M. A. (2023). A global land cover training dataset from 1984 to 2020. Scientific Data , 10 (1). https://doi.org/10.1038/s41597-023-02798-5 Story, M., & Congalton, R. G. (1986). Accuracy assessment: a user’s perspective. Photogrammetric Engineering and Remote Sensing , 52 , 397–399. https://api.semanticscholar.org/CorpusID:133251404 Taïbi, A. N., Hannani, M. El, Khalki, Y. El, & Ballouche, A. (2019). The agroforestry parks of Azilal (Morocco): a centuries-old and still living landscape construction. Revue de Géographie Alpine , 107–3 . https://doi.org/10.4000/rga.6612 Terasaki Hart, D. E., Yeo, S., Almaraz, M., Beillouin, D., Cardinael, R., Garcia, E., Kay, S., Lovell, S. T., Rosenstock, T. S., Sprenkle-Hyppolite, S., Stolle, F., Suber, M., Thapa, B., Wood, S., & Cook-Patton, S. C. (2023). Priority science can accelerate agroforestry as a natural climate solution. Nature Climate Change , 13 (11), 1179–1190. https://doi.org/10.1038/s41558-023-01810-5 Thapa, B., Lovell, S., & Wilson, J. (2023). Remote sensing and machine learning applications for aboveground biomass estimation in agroforestry systems: a review. Agroforestry Systems , 97 (6), 1097–1111. https://doi.org/10.1007/s10457-023-00850-2 Torralba, M., Fagerholm, N., Burgess, P. J., Moreno, G., & Plieninger, T. (2016). Do European agroforestry systems enhance biodiversity and ecosystem services? A meta-analysis. Agriculture, Ecosystems & Environment , 230 , 150–161. https://doi.org/10.1016/j.agee.2016.06.002 Vaupel, A., Bednar, Z., Herwig, N., Hommel, B., Moran-Rodas, V. E., & Beule, L. (2023). Tree-distance and tree-species effects on soil biota in a temperate agroforestry system. Plant and Soil , 487 (1–2), 355–372. https://doi.org/10.1007/s11104-023-05932-9 Zayani, I., Ammari, M., Ben Allal, L., & Bouhafa, K. (2023). Agroforestry olive orchards for soil organic carbon storage: Case of Saiss, Morocco. Heliyon , 9 (12), e22910. https://doi.org/10.1016/j.heliyon.2023.e22910 Zhu, X., Liu, W., Chen, J., Bruijnzeel, L. A., Mao, Z., Yang, X., Cardinael, R., Meng, F. R., Sidle, R. C., Seitz, S., Nair, V. D., Nanko, K., Zou, X., Chen, C., & Jiang, X. J. (2020). Reductions in water, soil and nutrient losses and pesticide pollution in agroforestry practices: a review of evidence and processes. Plant and Soil , 453 (1–2), 45–86. https://doi.org/10.1007/s11104-019-04377-3 Additional Declarations No competing interests reported. Supplementary Files AnnexeAReferencedata.xlsx AnnexeBSurveyinstructions.pdf AnnexeCInterpretedsites.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6000362","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":414212845,"identity":"85ffad4c-9a10-4382-81c1-08a7522f13dd","order_by":0,"name":"Ouadya Tahiri","email":"data:image/png;base64,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","orcid":"","institution":"AgroParisTech, UMR CIRED","correspondingAuthor":true,"prefix":"","firstName":"Ouadya","middleName":"","lastName":"Tahiri","suffix":""},{"id":414212846,"identity":"d802d949-8d57-4d67-aea5-d4cc052ccbd8","order_by":1,"name":"Damien Beillouin","email":"","orcid":"","institution":"CIRAD, UPR HORTSYS","correspondingAuthor":false,"prefix":"","firstName":"Damien","middleName":"","lastName":"Beillouin","suffix":""},{"id":414212847,"identity":"151e8c1a-a18b-4e09-956c-597b625054de","order_by":2,"name":"Patrice Dumas","email":"","orcid":"","institution":"CIRAD, UMR CIRED","correspondingAuthor":false,"prefix":"","firstName":"Patrice","middleName":"","lastName":"Dumas","suffix":""},{"id":414212848,"identity":"821a63b1-2aa4-42db-b2ec-410f105064e5","order_by":3,"name":"Rémi Prudhomme","email":"","orcid":"","institution":"CIRAD, UMR CIRED","correspondingAuthor":false,"prefix":"","firstName":"Rémi","middleName":"","lastName":"Prudhomme","suffix":""},{"id":414212849,"identity":"83285c2d-39a6-4ab7-88a1-89aa4b14fba2","order_by":4,"name":"David Makowski","email":"","orcid":"","institution":"Université Paris-Saclay, INRAE, UMR 518 MIA-PS","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Makowski","suffix":""}],"badges":[],"createdAt":"2025-02-10 15:08:25","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6000362/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6000362/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":76648349,"identity":"63c5a3ec-bb98-44d6-a87f-6b4f66b9a0f5","added_by":"auto","created_at":"2025-02-19 09:21:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":286784,"visible":true,"origin":"","legend":"\u003cp\u003eClassification tree to classify 30x30m land units into three types of agroforestry systems (alley cropping, hedgerows, scattered agroforestry) and four types of non-agroforestry systems (orchards, natural trees, cropland without trees, other) based on visual land cover characteristics\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/717b5926b3e8cb7b681fdba0.png"},{"id":76649009,"identity":"54dd2235-84b6-45ac-a5b4-c91a2863e907","added_by":"auto","created_at":"2025-02-19 09:29:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":613564,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of Collect Earth Online platform with an example of image classified by the interpreters. The yellow square represents the classified unit, the red one represents the unit to use to calculate tree density, and the yellow circle represents the human influence radius\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/163e4861a21b68af3b1d2d06.png"},{"id":76648352,"identity":"e687481c-77d7-4417-9182-a04f57a9fad5","added_by":"auto","created_at":"2025-02-19 09:21:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":969208,"visible":true,"origin":"","legend":"\u003cp\u003eMaps illustrating the distribution of the labeled sites across different agroforestry and non-agroforestry classes (a) and distribution of well-classified and misclassified sites resulting from the classification tree implemented by the administrator (interpreter 1) (b)\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/3cc8f5206b54f5f48c60d342.png"},{"id":76648362,"identity":"a4006e1f-4c98-4ff6-93d5-b81eb1e92723","added_by":"auto","created_at":"2025-02-19 09:21:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3077071,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of the classification tree for the six land use classes: (a) confusion matrix showing the number of correctly classified sites (diagonal cells) and misclassified sites (off-diagonal cells); (b) commission and omission errors per class (proportion). The group of sites considered is the group of interpreter 1 (n=267)\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/82ecfdcd34a30f21937efef7.png"},{"id":86159231,"identity":"eefda135-0b6d-4746-ba3b-5a916e07b6bc","added_by":"auto","created_at":"2025-07-07 12:02:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6059896,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/db4f4da3-d806-4ed9-b027-9146de80b302.pdf"},{"id":76648350,"identity":"a1c68802-d55f-46fb-a9c8-06f39f1a2b44","added_by":"auto","created_at":"2025-02-19 09:21:39","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":114080,"visible":true,"origin":"","legend":"","description":"","filename":"AnnexeAReferencedata.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/aa0f853a6b85451e9e8d2294.xlsx"},{"id":76648357,"identity":"5629af07-79cb-4299-bebc-a86d96b0ebd7","added_by":"auto","created_at":"2025-02-19 09:21:40","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":474850,"visible":true,"origin":"","legend":"","description":"","filename":"AnnexeBSurveyinstructions.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/5be41bb4013dbd63b66af010.pdf"},{"id":76649010,"identity":"bcfc3ede-bf92-4d38-b74f-de2a6375e622","added_by":"auto","created_at":"2025-02-19 09:29:39","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":75132,"visible":true,"origin":"","legend":"","description":"","filename":"AnnexeCInterpretedsites.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6000362/v1/d0b9745b3ed0b776843f49b7.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Classification of agroforestry systems by photo-interpretation of satellite imagery","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAgroforestry systems, which integrate trees within agricultural landscapes (Nair \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1993\u003c/span\u003e) are increasingly recognized for their multifaceted contributions to biodiversity conservation (Torralba et al. \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), soil health enhancement (Zhu et al. \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), and climate change mitigation (Chapman et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). However, these systems are highly heterogeneous (Nair \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1993\u003c/span\u003e), with diverse spatial arrangements and components (e.g., crops, pasture animals, trees), leading to considerable variability in their ecological and economic performance. In particular, studies have demonstrated that carbon sequestration and yield values can strongly differ across different agroforestry system types (Beillouin et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Cardinael et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Ivezić et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The impacts of agroforestry systems are also highly context-dependent, varying by factors such as soil types, climatic conditions, and local management practices (Kuyah et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The global distribution of these systems remains uneven, with significant regional disparities in their prevalence and adoption. This highlights the need for accurate, high-resolution spatial maps that capture the diversity of agroforestry systems and account for the contextual variability in their environmental impacts. In the absence of such maps, and clear differentiation between agroforestry practices, current global land use models fail to adequately represent these systems, resulting in substantial uncertainties regarding their environmental benefits (Terasaki Hart et al. \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe spatial heterogeneity of agroforestry systems complicates their mapping from satellite imagery (Escobar-L\u0026oacute;pez et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Ndao et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), as some agroforestry systems may be confounded with other land use types, such as orchards or natural vegetation (Bol\u0026iacute;var-Santamar\u0026iacute;a \u0026amp; Reu \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Lesiv et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Given the growing global emphasis on the ambition of agroforestry practices (Rosenstock et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) and its environmental significance, developing reliable and scalable classification systems based on remote sensing techniques is critical. Such classification systems are essential for accurately mapping agroforestry systems, tracking their expansion, monitoring their ecological and economic outcomes, and assessing their role in climate change mitigation across diverse regions, in particular, when coupled with robust measurement, reporting, and verification (MRV) frameworks (Batjes et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003ePhoto interpretation refers to the process of identifying, analyzing, and extracting information from aerial or satellite imagery based on visible features (Lillesand et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). This technique has been extensively applied across various fields, such as forestry (Brovelli et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Lesiv et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), agriculture (Belgiu \u0026amp; Csillik \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Hussain et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), and archeology (Doneus \u0026amp; Doneus \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). A key advantage of this approach is its ability to enable remote monitoring without the need for on-site presence. With the rise of machine learning and the growing accessibility of satellite data, photo interpretation is increasingly used to produce extensive databases, often including large volumes of labeled points across vast geographical areas (Stanimirova et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Such databases can subsequently be used to train and test machine-learning models to generate land-use maps (Brandt \u0026amp; Stolle \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Lesiv et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003ePhoto-interpretation has been widely used to detect trees outside forested areas (Brandt \u0026amp; Stolle, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Sarti et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). However, these approaches are often insufficient to distinguish different types of agroforestry systems, and agroforestry from other tree-dominated land use, such as savannahs. Lesiv et al. (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) used visual interpretation of satellite imagery to distinguish agroforestry from other tree-based ecosystems as part of their global forest management database initiative. While their approach improves georeferenced data availability on agroforestry, the absence of clear criteria defining the agroforestry class reduces the reproducibility of the approach. In addition, in their final classification, the different agroforestry systems were not distinguished. Consequently, identifying agroforestry systems from satellite imagery remains a significant challenge, necessitating the development of an operational and standardized classification framework. Such a framework must address the inherent ambiguity between agroforestry systems with scattered trees (e.g. parklands) and non-agroforestry natural landscapes (e.g., savannahs).\u003c/p\u003e \u003cp\u003eAdditionally, the delineation of agroforestry systems remains poorly defined, particularly when these systems co-exist, or are adjacent to other land use types. Agroforestry units have been delineated in some studies using segmentation techniques, such as object-based image analysis (OBIA) combined with other geospatial data (Ndao et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). However, these techniques can become computationally exhaustive when applied to large-scale studies, as they require significantly longer processing times (Guirado et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study aims to develop a robust and systematic framework for differentiating and classifying agroforestry systems using photo-interpretation by analyzing the structural characteristics of these systems. The approach is based on the premise that distinct agroforestry systems can be characterized by unique spatial configurations of trees and associated crops or pastures. Our specific objectives are threefold: i) identify key land cover characteristics that can reliably differentiate agroforestry from non-agroforestry systems using photo-interpretation, ii) develop an operational classification tree based on these characteristics, enabling the discrimination of various agroforestry systems from other land use types, iii) evaluate this classification tree using a set of satellite images covering multiple land use types and geographical regions. By achieving these goals, we aim to provide a globally applicable and effective tool for large-scale mapping of agroforestry systems.\u003c/p\u003e"},{"header":"2. Materials and methods","content":"\u003cp\u003eWe followed a three-step approach to classify agroforestry systems based on photo-interpretation. In step 1, we conducted a literature review in order to identify key land use characteristics that could serve as reliable indicators to distinguish different agroforestry systems. In step 2, we developed a streamlined and generalizable classification tool by translating these indicators into simple decision rules. In step 3, we validated the resulting classification tool by testing it with different interpreters using 300 land use reference plots.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e\u003cb\u003e2.1. Step 1. Identification of land use characteristics to classify agroforestry and non-agroforestry systems\u003c/b\u003e\u003c/h2\u003e \u003cp\u003eOur objective is to categorize the agroforestry systems, as defined by Nair et al. (2009) and refined by Cardinael et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), based on the spatial arrangement of trees both within-plot and in the surrounding area. Some agroforestry systems, however, were excluded from the scope of this study, particularly those characterized by closed canopy systems (e.g., shaded perennial crops, and multistrata systems), which are difficult to distinguish from forests using image-based techniques. Additionally, sequential agroforestry systems, such as improved fallows, were not considered due to their temporal dynamics (Friday et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e1999\u003c/span\u003e), which makes them unsuitable for image-based classification. Based on these considerations, we focused on three categories of agroforestry systems : (1) Alley cropping, where trees are arranged in parallel rows; (2) Hedgerows, where trees surround agricultural fields; and (3) Scattered agroforestry systems, in which trees are dispersed within pasturelands or croplands (e.g., \u003cem\u003eParklands\u003c/em\u003e, \u003cem\u003eDehesa\u003c/em\u003e, and \u003cem\u003eMontados\u003c/em\u003e).\u003c/p\u003e \u003cp\u003eTree presence and their spatial arrangement alone may be insufficient to reliably differentiate between agroforestry and some other land-use types. This is especially challenging in distinguishing alley cropping from orchards as they often have similar tree arrangements (i.e., parallel rows). To address this challenge, we assumed that tree row spacing is narrower in orchards compared to alley cropping systems, driven by the need for higher fruit yield (Gomez-del-Campo et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Rallo et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAnother potential source of confusion arises when distinguishing between scattered agroforestry systems and savannah or some other forms of dispersed natural tree cover as these systems share the same spatial configuration and other ecological and dynamic characteristics (FAO \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). To tackle this issue, we assumed that natural systems are located further from human structures compared to scattered agroforestry systems, which tend to be situated closer to human settlements (FAO \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). Grazing is one of the most mobile and irregular activities that characterize these agroforestry systems. However, it is difficult to define a spatial threshold as it varies according to several factors (Liao \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). In this case, we start from the assumption that pastoralists would first utilize patches of rangelands with better forage quality and lower travel costs near the settlement (Liao et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe observation scale is critical for classification, and landscape heterogeneity may lead to misclassification. For instance, hedgerows may be confused with large cropland fields surrounded by patches of trees. Beyond a given distance from the hedgerow, a plot should be classified as cropland without trees rather than agroforestry. To avoid misclassification, we cut out the landscape in observation units assumed to be representative for the purpose of agroforestry classification.\u003c/p\u003e \u003cp\u003eTo handle the issues presented in this section, the classification integrates information from three spatial scales:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003e30m x 30m Observation Unit\u003c/b\u003e: This unit serves as the primary level for classifying individual sites. We selected this size because of the radius of influence of agroforestry trees on variables such as soil parameters (Cardinael et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Pardon et al. \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), crop yield (Roupsard et al. \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), and soil organism abundance (Vaupel et al. \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) typically ranges between 10m and 30m. Additionally, this scale aligns with widely used remote sensing imagery data, such as Landsat 8 \u0026amp; 9 with 30m-spatial resolution (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ewww.usgs.gov/landsat-missions\u003c/span\u003e\u003cspan address=\"http://www.usgs.gov/landsat-missions\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e).\u003c/span\u003e At this scale, tree presence, shadow patterns, crown shapes, and surrounding land use characteristics are visually assessed to initiate the classification process.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003e10,000m\u0026sup2; Land Unit (1 ha)\u003c/b\u003e: This scale is used to calculate tree density, serving as a threshold for distinguishing agroforestry systems with scattered trees from denser forest systems. A threshold of 50 trees per hectare was applied, in accordance with the recommendations of the French Agroforestry Association (AFAF) (2014) to establish an agroforestry system. The chosen 1 ha area scale allows the interpreter to estimate tree density relatively quickly, by visual inspection and is large enough to avoid measurement errors at low density.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003e2 km Radius\u003c/b\u003e: This broader scale evaluates the proximity of human-made structures to distinguish natural scattered tree systems (e.g., savannah) from agroforestry systems with scattered trees (e.g., parklands). The 2 km distance threshold reflects the average daily walking distance of pastoralists from a \u0026ldquo;base camp\u0026rdquo; to a grazing site (Liao et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). In this context, the \"base camp\" is typically any human-made structure, such as a house, agricultural plot, or road. Roads, in particular, are key indicators, as they enhance accessibility and enable the development of agroforestry systems (Bennett \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). This 2 km distance threshold between the \u0026ldquo;base camp\u0026rdquo; and the observation unit is used to distinguish scattered agroforestry systems and natural ecosystems.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eBased on this analysis, we identify and select the following land use characteristics for classification: (i) the presence or absence of trees; (ii) the presence of geometric patterns indicative of human-managed landscapes; (iii) the spatial arrangement of trees (trees in parallel rows, tree edges, or dispersed trees); (iv) the proximity to of human structures (e.g., houses, roads); and (v) tree density, which helps differentiate between various tree-dominated systems. These characteristics collectively provide a robust foundation for distinguishing agroforestry systems from other land-use types while ensuring generalizability and practical applicability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Step 2. Definition of the classification tree\u003c/h2\u003e \u003cp\u003eTo categorise land use systems into distinct classes, we followed the method proposed by ​Caumont and Ivanaj (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) based on similarity profiles using previously defined variables. This approach focuses on identifying the most discriminative variables and grouping systems according to three principles: i) \u003cem\u003eExclusivity\u003c/em\u003e\u0026mdash;each system is assigned to a single group; ii) \u003cem\u003eHomogeneity\u003c/em\u003e\u0026mdash; systems within each group share similar characteristics, minimizing intra-group variability while maximizing inter-group differences; and iii) \u003cem\u003eParsimony\u003c/em\u003e\u0026mdash;a minimal number of groups is used to summarize the diversity of systems.\u003c/p\u003e \u003cp\u003eThe resulting classification tree (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) uses a hierarchical structure with seven decision nodes, allowing for systematic discrimination between land use types. It classifies 30x30m land units into seven distinct land use types: Alley cropping, Hedgerows, Scattered agroforestry, Natural trees, Cropland without trees, Orchards, and others. The first classification criterion in the classification process is the presence of trees in the 30x30m observation unit. Trees are identified visually based on their shadow, crown shape, and surrounding land use characteristics (Brandt \u0026amp; Stolle \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The next classification steps are determined by six other closed questions (Yes/No) formulated based on the variables identified in the previous step, such as the trees' spatial arrangement, tree density, proximity to human activities, and other visual variables.\u003c/p\u003e \u003cp\u003eUltimately, we get both a user-friendly and globally applicable classification tree. Its simplicity, rooted in a limited set of decision rules, allows non-specialists to systematically and independently classify seven categories of land use types.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Step3. Assessing the classification tree\u003c/h2\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1. Georeferenced sites\u003c/h2\u003e \u003cp\u003eTo assess the accuracy of the decision tree, we compiled a database of georeferenced sites, each labeled and corresponding to a specific land use system. To collect these labeled sites, a search was performed on three databases, Scopus, Web of Science, and Google Scholar, for studies reporting georeferenced data of the selected systems between 2016 and 2024. Additional sources such as reports from funded research projects (e.g., AGFORWARD) were consulted. These searches were conducted using keywords such as alley cropping, tree-based intercropping, hedgerows, and scattered agroforestry. Additionally, we collected data from three pre-existing databases for croplands without trees (Laso Bayas et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Lesiv et al. \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) and natural trees (Lesiv et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The complete list of data sources is available in Annex A. These labeled sites serve as ground-truth data, having been validated by experts.\u003c/p\u003e \u003cp\u003eEach study was reviewed to ensure compliance with the following selection criteria: i) \u003cb\u003eTemporal alignment\u003c/b\u003e - only points collected between 2016 and 2024 were retained to match the aerial imagery used for photo interpretation. ii) \u003cb\u003eSystem specificity\u003c/b\u003e: Studies with vague or ambiguous descriptions of agroforestry systems or those potentially conflating different system types were excluded. iii) \u003cb\u003eGeoreferencing precision\u003c/b\u003e: Sites with imprecise coordinates were discarded unless this information could be corrected based on the annexes of the papers or any other indicator that could help to geolocate the point. Georeferenced points were cross-verified using Google Earth (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://earth.google.com/\u003c/span\u003e\u003cspan address=\"https://earth.google.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e)\u003c/span\u003e and Bing Maps (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.bing.com/maps\u003c/span\u003e\u003cspan address=\"https://www.bing.com/maps\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e)\u003c/span\u003e to ensure that each site met the basic characteristic of the specified system, for instance, we did not include a point as agroforestry or natural trees in the validation dataset when there is no visible tree on the satellite imagery, as this would distort the interpretation and accuracy of our framework.\u003c/p\u003e \u003cp\u003eAs a result, 134 studies were retained and distributed as follows: 45 focused on Alley cropping, 21 on hedgerows, 35 on Scattered AF, and 35 on orchards. From the selected studies, we were able to identify 150 georeferenced agroforestry sites (50 sites per agroforestry system) and 150 georeferenced non-agroforestry sites to evaluate the accuracy of the classification tree. We tried to set a threshold of 50 points per system to guarantee a balanced database and ensure that rare classes are well represented. We supplemented the classes originally with less than 50 sites with additional georeferenced images obtained from AGFORWARD reports (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ewww.agforward.eu\u003c/span\u003e\u003cspan address=\"http://www.agforward.eu\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e)\u003c/span\u003e to reach the threshold.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2. Implementation of the classification tree\u003c/h2\u003e \u003cp\u003eThe 300 georeferenced sites were uploaded to \u003cem\u003eCollect Earth Online (\u003c/em\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ewww.collect.earth\u003c/span\u003e\u003cspan address=\"http://www.collect.earth\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e)\u003c/span\u003e, a user-friendly platform designed for satellite imagery interpretation (Saah et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2019\u003c/span\u003e)​. For each site, a 30x30m observation unit was defined at the center of the georeferenced coordinates. To facilitate the implementation of the classification tree (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), we delineated the three observation units presented in step 1: i) \u003cb\u003e30x30m Square\u003c/b\u003e: The primary unit of classification. ii) \u003cb\u003e100x100m Square\u003c/b\u003e: Used to calculate tree density. iii) \u003cb\u003e2km Radius\u003c/b\u003e: Used to account for proximity to human-made structures.\u003c/p\u003e \u003cp\u003eThe sequential classification rules derived from the classification tree were programmed into the platform. Five interpreters, including one administrator, who are also the authors of this paper, applied these rules to classify the satellite imagery. A comprehensive training guide was provided to introduce the interpreters to the platform (see Annex B) and the classification protocol.\u003c/p\u003e \u003cp\u003eTo ensure unbiased assessments, the interpreters (other than the administrator) did not have access to the true categories of the systems represented in the images and worked independently without communication. Subsequently, the administrator compiled and compared their classifications (Annex C) against reference data for each site.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.3.3. Metrics used for evaluation\u003c/h2\u003e \u003cp\u003eThe performance of the interpreters\u0026rsquo; classifications was evaluated according to four metrics: Overall accuracy, Commission error, Omission error, and Cohen\u0026rsquo;s Kappa (McHugh \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). Additionally, the consistency among the five interpreters was assessed using the Fleiss\u0026rsquo; Kappa metric (Fleiss \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1971\u003c/span\u003e). These metrics were used to quantify both individual (i.e., per interpreter) and collective (aggregated) classification performances.\u003c/p\u003e \u003cp\u003eThe accuracy is defined as the proportion of labeled sites correctly classified. From this, the error rate (%) was derived as: (1-Accuracy)*100.\u003c/p\u003e \u003cp\u003eCohen\u0026rsquo;s Kappa (k) was used to measure the agreement between each interpreter\u0026rsquo;s classification and the true labeled classes, accounting for chance agreement. It is calculated as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:k\\:=\\:\\frac{{p}_{o\\:}-\\:{p}_{e}}{1\\:-\\:{p}_{e}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{o\\:}\\)\u003c/span\u003e\u003c/span\u003eis the observed agreement, representing the proportion of points where the interpreter's classification matched the true label\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:{p}_{o}\\:=\\:\\frac{{\\sum\\:}_{c=1}^{C}T{P}_{c}}{N}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewith \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:T{P}_{c}\\)\u003c/span\u003e\u003c/span\u003e the number of true positives for class \u003cem\u003ec\u003c/em\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:C\\:\\)\u003c/span\u003e\u003c/span\u003e is the total number of classes (here, \u003cem\u003eC\u003c/em\u003e\u0026thinsp;=\u0026thinsp;6), and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:N\\:\\)\u003c/span\u003e\u003c/span\u003e the total number of data points.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{e}\\)\u003c/span\u003e \u003c/span\u003e is the expected agreement by chance, calculated as:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{p}_{e}\\:=\\:{\\sum\\:}_{c=1}^{C}\\left(\\frac{Number\\:of\\:points\\:classified\\:in\\:class\\:c\\:*\\:True\\:number\\:of\\:point\\:in\\:class\\:c}{{N}^{2}}\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eFleiss\u0026rsquo;s kappa, used for assessing consistency among the five interpreters, modifies the calculations of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{o}\\:\\)\u003c/span\u003e\u003c/span\u003eand \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{e}\\:\\)\u003c/span\u003e\u003c/span\u003eaccounts for multiple raters:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:{p}_{o}\\:=\\frac{1}{N\\:\\times\\:R\\:\\left(R\\:-1\\right)}{\\sum\\:}_{i\\:=\\:1}^{N}{\\sum\\:}_{c=1}^{C}{n}_{ic}\\left({n}_{ic}-1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:{p}_{e}\\:=\\:{\\sum\\:}_{c={1}_{\\:}^{\\:}}^{C}{p}_{c}^{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:{p}_{c}\\:=\\:\\frac{1}{N\\times\\:R}{\\sum\\:}_{i=1}^{N}{n}_{ic}\\)\u003c/span\u003e\u003c/span\u003e is the proportion of all assignments to class c.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:R\\:\\)\u003c/span\u003e \u003c/span\u003e: is the number of interpreters (here, \u003cem\u003eR\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5).\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{n}_{ic}\\)\u003c/span\u003e \u003c/span\u003e : is the number of interpreters who assigned point i to class c.\u003c/p\u003e \u003cp\u003eWhile Kappa values provide an overall performance measure, they do not describe the error levels associated with the different classes. To identify strengths and weaknesses in classification for individual classes, we calculated the following metrics:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eOmission Error (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Eo\\)\u003c/span\u003e\u003c/span\u003e) represents the proportion of sites belonging to a given class \u003cem\u003ec\u003c/em\u003e incorrectly classified in another class:\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\:Eo\\:=1-(\\:F{N}_{c}\\:/(T{P}_{c}\\:+F{N}_{c})\\)\u003c/span\u003e \u003c/span\u003e)\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eCommission Error (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Ec\\)\u003c/span\u003e\u003c/span\u003e) represents the proportion of sites incorrectly allocated to a class \u003cem\u003ec\u003c/em\u003e:\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equf\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:Ec\\:=1-(T{P}_{c}\\:/\\:(\\:T{P}_{c}\\:+F{P}_{c}\\left)\\right)$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003eFor instance, misclassifying an alley cropping point as an orchard results in an omission error for the \"Alley cropping\" class and a commission error for the \"Orchard\" class.\u003c/p\u003e \u003cp\u003eThe metrics and their 95% confidence intervals were computed based on bootstrap techniques to account for uncertainty and variability in performance measures.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.3.4. Interpreters\u0026rsquo; errors vs. Classification tree errors\u003c/h2\u003e \u003cp\u003eTo disentangle sources of misclassification, two error types were distinguished:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eInterpreter\u0026rsquo;s Error\u003c/b\u003e: This type of error occurs when a site that was misclassified by at least one interpreter (excluding the administrator) is correctly classified by the administrator (interpreter 1) after a careful implementation of the classification tree. To be as certain as possible that the application of the tree by the interpreter was incorrect, the administrator goes through the tree and checks each decision node to determine the precise step at which the interpreter made a mistake, in order to objectivize the error. The administrator can also use the reported site type to understand why the interpreter made a mistake.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eClassification Tree Error\u003c/b\u003e: This type of error occurs when all interpreters (including the administrator) misclassify a site even though they implement the classification tree correctly.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThis differentiation allowed us to isolate errors attributable to the classification tree itself from those resulting from interpreters\u0026rsquo; skill variability and thus calculate the potential accuracy of the decision tree. The metrics described above were recalculated for each error type, for the interpreter errors and the classification tree errors, separately.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Geographical distribution of the labeled sites\u003c/h2\u003e \u003cp\u003eThe 300 labeled sites used to evaluate our classification method provide comprehensive coverage across all continents, ensuring a broad and balanced geographical distribution (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). Although Europe accounts for the highest proportion (35%) of sites, all agroforestry systems \u0026mdash; alley cropping, hedgerows, and scattered agroforestry\u0026mdash; and tree non-agroforestry systems \u0026mdash;orchards, croplands without trees, and natural trees \u0026mdash; are represented in all continents. Some land use classes are more frequently represented in some regions than others. For instance, in Russia, the \u0026ldquo;Natural trees\u0026rdquo; class is predominant as forests cover a large share of this country (FAO, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Accuracy of the interpreters\u0026rsquo; classifications\u003c/h2\u003e \u003cp\u003eMisclassifications are geographically dispersed, with no clear regional pattern, as shown in the map presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb. This map depicts the distribution of well-classified and misclassified sites by Interpreter 1, based on photo-interpretation using the Collect Earth Online platform. Overall, the classification accuracy ranges from 0.75 to 0.81, depending on the interpreters, with an average of 0.79 (Table.1). These results indicate that the least accurate interpreter was able to classify 75% (95%CI=[70%, 80%]) of sites correctly, while the most accurate achieved 81% (95%CI=[76%, 85%]) accuracy. The coefficient of variation (CV) across the five interpreters is 2.89%, reflecting low dispersion around the mean accuracy and low variability among interpreters. Part of the between interpreters variation can be attributed to certain interpreters not assessing all labeled sites due to difficulties in accessing or visualizing the images. Consequently, the samples of sites evaluated by the five interpreters are slightly different. While accuracy requirements depend on the context (Story \u0026amp; Congalton \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e1986\u003c/span\u003e), an accuracy of 79% is considered sufficient for large-scale analyses focusing on broad patterns, as suggested by Charles and Olson (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). Additionally, all interpreters\u0026rsquo; accuracy values are at least 4.5 times higher than the 17% accuracy expected from random classification with six classes. This high classification performance is further confirmed by the values of Cohen\u0026rsquo;s kappa (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) ranging from 0.70 (95%CI= [0.64, 0.76]) to 0.77 (95%CI= [0.71, 0.81]). According to Landis and Koch (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e1977b\u003c/span\u003e), these values reflect a \u0026ldquo;\u003cem\u003esubstantial\u0026rdquo;\u003c/em\u003e agreement between the interpreters and the reference data. Consistent with the CV of accuracy values, the coefficient of variation of the kappa values among the five interpreters is lower than 5% (3.53%). Additionally, Fleiss' kappa of 0.86 reveals a strong overall agreement among all interpreters (Landis \u0026amp; Koch \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e1977a\u003c/span\u003e).\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\u003eOverall accuracy and Cohen\u0026rsquo;s kappa with their 95% confidence intervals for each interpreter, along with their average (Avg) and coefficient of variation (CV). Values of n indicate the number of sites assessed by each interpreter\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=\"char\" char=\".\" 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=\"char\" char=\".\" 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\u003eInterpreter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95% CI interval\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCohen's kappa\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95% interval\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1 (n\u0026thinsp;=\u0026thinsp;283)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.76\u0026ndash;0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.71\u0026ndash;0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2 (n\u0026thinsp;=\u0026thinsp;269)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.76\u0026ndash;0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.71\u0026ndash;0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3 (n\u0026thinsp;=\u0026thinsp;263)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.70\u0026ndash;0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.64\u0026ndash;0.76\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4 (n\u0026thinsp;=\u0026thinsp;264)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.75\u0026ndash;0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7\u0026ndash;0.81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5 (n\u0026thinsp;=\u0026thinsp;263)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.73\u0026ndash;0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.68\u0026ndash;0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAvg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\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\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the classification errors for each land-use class. Orchards exhibit the highest error rate, averaging 53.3% across interpreters, with individual rates ranging from 48\u0026ndash;66%. Scattered agroforestry has the second highest error rate, averaging 38%, followed by Alley cropping with an average error rate of 25.5%. In contrast, Cropland without trees, Hedgerows, and Natural trees show much lower average error rates of 10.1%, 2.8%, and 2.0% respectively. While error rates vary among interpreters, no single interpreter achieves the lowest error across all classes. For example, interpreter 1 minimizes error rates for Alley cropping, Scattered agroforestry, Croplands without trees, and Natural trees, whereas interpreter 5 performs best for Hedgerows and Orchards.\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\u003eError rate ((1-Accuracy) *100) per land-use class and interpreter, along with their average (Avg) and coefficient of variation (CV). Values of n indicate the number of sites assessed by each interpreter or for each land use class\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" 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=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterpreter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAlley cropping\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHedgerows\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eScattered AF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOrchards\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCroplands WT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNatural trees\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1 (n\u0026thinsp;=\u0026thinsp;283)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e28.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e65.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2 (n\u0026thinsp;=\u0026thinsp;269)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e43.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e48.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3 (n\u0026thinsp;=\u0026thinsp;263)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e42.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e46.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e48.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e6.12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4 (n\u0026thinsp;=\u0026thinsp;264)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e19.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e37.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e54.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e10.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5 (n\u0026thinsp;=\u0026thinsp;263)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e25.64\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\u003e34.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e47.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e23.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAvg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e25.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e37.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e53.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e10.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e35.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e64.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e110.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Potential accuracy of the classification tree\u003c/h2\u003e \u003cp\u003eThe accuracy metrics were calculated a second time on all labeled sites after removing interpretation errors. This second run isolates errors attributable solely to the tree's limitations in distinguishing between classes, removing the influence of interpreters' misinterpretations. Consequently, the resulting error rates represent the theoretical lower bounds of classification errors achievable with perfect implementation free from interpreter-related inaccuracies.\u003c/p\u003e \u003cp\u003eAfter the removal of interpreters\u0026rsquo; errors, the classification tree demonstrates improved performance, achieving an average accuracy of 0.86 and a Cohen\u0026rsquo;s kappa of 0.83 (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Nevertheless, approximately 15% of classification errors persist, reflecting inherent limitations of the decision tree rather than errors stemming from interpretation. The small fluctuations of the accuracy and Cohen\u0026rsquo;s kappa between the five groups of sites (coefficient of variation\u0026thinsp;\u0026lt;\u0026thinsp;1.6%) are not due to differences between experts but from variations in the sets of images assessed. As mentioned above, some interpreters did not evaluate all 300 labeled sites, leading to slight fluctuations in the sample sizes and associated metrics.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAccuracy and Cohen\u0026rsquo;s kappa with 95% confidence intervals, along with their average (Avg) and coefficient of variation (CV) after excluding the interpreter-related error. The samples of sites considered in Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e were carefully checked by Interpreter 1 to assess the residual error\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=\"char\" char=\".\" 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=\"char\" char=\".\" 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\u003eSample of sites\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95% interval\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCohen's kappa\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95% interval\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1 (n\u0026thinsp;=\u0026thinsp;267)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.81\u0026ndash;0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.77\u0026ndash;0.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2 (n\u0026thinsp;=\u0026thinsp;250)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.83\u0026ndash;0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.79\u0026ndash;0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3 (n\u0026thinsp;=\u0026thinsp;235)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.79\u0026ndash;0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.75\u0026ndash;0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4 (n\u0026thinsp;=\u0026thinsp;244)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.82\u0026ndash;0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.79\u0026ndash;0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5 (n\u0026thinsp;=\u0026thinsp;237)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.82\u0026ndash;0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.79\u0026ndash;0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAvg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\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\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the performance of the classification tree for each land use class after removing the interpreters-related errors. The off-diagonal cells of the confusion matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea) reveal two primary sources of misclassification, i.e. confusion between Orchards and Alley Cropping, and between Scattered AF and Natural trees. The commission and omission errors (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb) provide further insight into these misclassifications. Specifically, 36% of the sites are misclassified as Alley cropping, predominantly from Orchards class (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea), while 58% of the Orchard sites are misclassified into other classes, primarily Alley cropping. In contrast, Alley Cropping shows lower omissions errors (14%), and the confusion matrix indicates that the misclassification of Alley Cropping as Orchards occurs less frequently than the reverse. This pattern indicates a higher propensity for Orchards to be misclassified compared to Alley Cropping, even when interpreter-related errors are eliminated.\u003c/p\u003e \u003cp\u003eConcerning the confusion between Scattered AF and Natural Trees, Scattered AF has an omission error of 23% and a commission error of 10%, whereas Natural Trees has no omission error and a commission error of 18%. These results indicate that nearly one out of four sites of Scattered AF is misclassified, while all Natural trees sites are correctly identified. This pattern suggests that the confusion mainly arises from misclassifying Scattered trees as Natural trees, rather than the opposite.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMerging the Alley Cropping and Orchards classes results in a notable improvement in the classification tree's performance. The overall accuracy increased from 0.86 to 0.94 (reducing the error by 6%), while Cohen\u0026rsquo;s kappa rises from 0.83 to 0.92, whichLandis and Koch (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e1977b\u003c/span\u003e) described as an \u0026ldquo;almost perfect\u0026rdquo; agreement. The omission error of the new combined class Alley Cropping/Orchard drops to 4%, and the commission error is reduced to 1.7%.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eWe developed a classification tree to identify three agroforestry systems - \u003cem\u003eHedgerows\u003c/em\u003e, \u003cem\u003eScattered\u003c/em\u003e agroforestry, and \u003cem\u003eAlley cropping\u003c/em\u003e - and three non-agroforestry systems \u0026mdash; \u003cem\u003eCropland without trees\u003c/em\u003e, \u003cem\u003eNatural trees\u003c/em\u003e, and \u003cem\u003eOrchards\u003c/em\u003e\u0026mdash; using a photo-interpretation approach. This novel method relies on straightforward classification rules, making the proposed approach accessible to interpreters with basic knowledge of agroforestry and agriculture. Moreover, our classification tree can be easily applied using the free and open-source \u003cem\u003eCollect Earth Online\u003c/em\u003e platform (Saah et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Despite its simplicity, the classification tree demonstrates high performance, achieving robust results in distinguishing between agroforestry and non-agroforestry systems. Indeed, based on a dataset of 300 labeled sites representing agroforestry and non-agroforestry systems across all continents, we demonstrated that the lower limit of classification error rate achievable with perfectly trained users would be 14%. By merging the classes \u0026ldquo;Alley Cropping\u0026rdquo; and \u0026ldquo;Orchards,\u0026rdquo; we further reduced the error rate to approximately 10%. This balance between ease of use and reliability makes it an innovative tool for large-scale agroforestry mapping and monitoring.\u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Sources of error and perspectives for improvement\u003c/h2\u003e \u003cp\u003eThe classification error rate with non-expert ranged from 19\u0026ndash;25%, depending on the interpreter. Errors associated with interpreters are manageable and could be minimized through targeted training. Previous studies on photo interpretation have shown intensive training sessions can significantly reduce photo interpretation errors (Lesiv et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Importantly, interpreters' errors often vary across individuals and images, suggesting that a collaborative, consensus-based approach could further enhance classification consistency. In contrast, classification errors inherent to the decision tree itself are systemic, arising from its structural limitations. A low rate of residual errors persists despite eliminating interpretation errors and reflects the intrinsic difficulty of distinguishing between certain land-use classes.\u003c/p\u003e \u003cp\u003eA major source of error involved the frequent misclassification of orchards as alley cropping (more frequently than the reverse), likely due to the similarity between their spatial arrangement as both feature trees planted in parallel rows. This similarity stems from shared agronomic principles, such as optimal tree spacing for maximizing productivity and facilitating mechanization (Gomez-del-Campo et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Menzel \u0026amp; Le Lagadec \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Furthermore, orchards and alley cropping can correspond to different stages of land use change. For instance, adding crops to orchards can create alley cropping systems, while removing crops can revert alley cropping to orchards (S\u0026aacute;nchez-Navarro et al. \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This dynamic nature underlines the necessity for high temporal and spatial resolution imagery to accurately monitor land-use transitions. However, widely used platforms like Google Earth or Microsoft Bing are limited by the infrequency of image updates. It reduces their utility for tracking rapid land-use change, as they capture images over extended intervals, limiting their ability to monitor fine temporal variations. In addition, there might be some issues in the labeling of alley cropping and orchard because some studies may refer to alley cropping as orchard agroforestry (Zayani et al. \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) making it unclear whether to attribute a given plot to orchard or alley cropping.\u003c/p\u003e \u003cp\u003eAnother challenge emerged in distinguishing scattered agroforestry from natural forests, particularly in regions where dense tree cover coincides with human activity, such as silvopastoral systems beneath tree canopies (Coble et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; \u0026Ouml;llerer et al. \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), This is confirmed by the commission error in the 'Natural Trees' class, where 19% of the points (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb) classified as natural trees were actually scattered agroforestry. Even though the tree density criterion prevents a correct classification in those cases, this criterion remains important, as scattered agroforestry systems are often found in transformed forest landscapes influenced by human activity, with low tree density, such as the Parklands in Africa (Ta\u0026iuml;bi et al. \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), the Spanish Dehesas (Gaspar et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), and the Montados in Portugal (Pinto-Correia et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). In addition to the tree\u0026rsquo;s density, our classification tree relies on proximity to human settlements to differentiate agroforestry from natural forests by defining the 2km distance radius. This approach aligns with the method of Lesiv et al. (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), who define natural forests as areas without detectable disturbances within a specific spatial distance (with a radius of approximately 0.5km).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Implications for future applications\u003c/h2\u003e \u003cp\u003eCollecting validation data remains a challenging step, largely due to a shortage of labeled and georeferenced data. Detailed information on agroforestry systems is often lacking, even in scientific publications. Moreover, precise coordinates are not always provided, reducing the number of usable validation plots. Our classification tree offers a new operational tool for future applications, such as collecting georeferenced and labeled data in areas including one or several types of agroforestry systems. Such data are essential for assessing the impact of agroforestry on biomass, crop production, biodiversity, and carbon stocks. It constitutes an interesting and less costly alternative to on-site data observation.\u003c/p\u003e \u003cp\u003eWhile remote sensing and classification trees are increasingly used for large-scale agroforestry mapping and monitoring, they cannot fully replace the need for detailed field-based observations, especially in complex agroforestry systems. Closed-canopy systems, such as shaded perennial crop systems (e.g., cacao, coffee, and tropical tree crop agroforestry), present particular challenges for remote sensing due to their complex vertical structures and the dynamic interactions between trees and crops (Bol\u0026iacute;var-Santamar\u0026iacute;a \u0026amp; Reu \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These systems, with their dense canopies, often elude accurate classification through traditional photo-interpretation techniques, particularly when distinguishing between different crops or when the canopy structure conceals underlying biodiversity.\u003c/p\u003e \u003cp\u003eTo address these limitations, future applications would benefit from an integrated approach that combines remote sensing with on-the-ground validation efforts. Advanced technologies such as LiDAR (Light Detection and Ranging) can significantly improve the characterization of complex canopy structures by providing detailed three-dimensional data on vegetation height, density, and stratification (Thapa et al. \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). When combined with high-resolution satellite imagery, LiDAR enables more accurate differentiation between agroforestry systems and other land uses, especially in landscapes with dense and complex vegetation. This hybrid approach is particularly advantageous for systems like shaded agroforestry, where subtle distinctions between trees and crops can be difficult to capture with optical imagery alone. LiDAR\u0026rsquo;s ability to detect vertical structure and fine-scale vegetation characteristics offers a more comprehensive view of these systems, which are often inadequately represented through traditional remote sensing methods (Dagar et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAnother promising direction for enhancing agroforestry monitoring is the use of crowdsourced data. Citizen science initiatives, where local communities contribute ground-based observations on land-use practices and environmental conditions, can play a pivotal role in validating satellite-based classifications and expanding the geographic coverage of agroforestry data collection (L.C. \u0026amp; L. 2023). By incorporating local knowledge into global datasets, these initiatives offer the potential for more granular insights into agroforestry systems, particularly in remote or understudied areas. Additionally, crowdsourcing is a scalable, cost-effective strategy for data collection, particularly in regions with limited access to high-end remote sensing technologies. It allows for broader participation and facilitates the rapid accumulation of valuable data, enhancing the accuracy of agroforestry mapping.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eDespite advancements in photo-interpretation platforms and the availability of high-resolution satellite imagery, operational systems are still lacking for accurately classifying agroforestry systems. Here, we developed a simple method to classify agroforestry systems into three categories\u0026mdash;alley cropping, scattered agroforestry, and hedgerows\u0026mdash;and to distinguish these systems from non-agroforestry land uses, i.e. cropland without trees, natural trees, and orchards. The proposed method relies on a classification tree that can be easily implemented by photo-interpretation of satellite imagery. Based on 300 observations of agroforestry and non-agroforestry plots covering all continents, we showed that its classification error rate ranged from 19 to 25%, depending on the interpreter. After eliminating the interpreter\u0026rsquo;s errors, the classification error rate was reduced to 14%. The main source of error stemmed from confusion between alley cropping and orchards and, to a lower extent, between scattered trees and natural trees. The fusion of these pairs of categories decreases the classification error to 6%. Looking ahead, the proposed classification tree offers significant potential for large-scale mapping of agroforestry systems. By improving the accuracy and scalability of agroforestry monitoring, this tool could contribute to more effective assessments of the ecosystem services and climate adaptation benefits provided by agroforestry systems, especially in the context of global climate change.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe authors have no conflicts of interest to declare that are relevant to the content of this article.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis work benefited from the French state aid managed by the ANR under the \"Investissements d'avenir\" programme with the reference ANR-16-CONV-0003 as a partial stipend alongside with a partial stipend from CIRAD. The funding sources had no role in the study design, data collection, analysis, or decision to publish.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors wrote and reviewed the main manuscript textO.T. analyzed data and prepared figures\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData is provided within the manuscript or supplementary information files\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAFAF (2014) L\u0026rsquo;AGROFORESTERIE EN 10 QUESTIONS. https://www.agroforesterie.fr/wp-content/uploads/2022/07/afaf-agroforesterie-en-10-questions.pdf .Accessed June 2024\u003c/li\u003e\n\u003cli\u003eBatjes, N. H., Ceschia, E., Heuvelink, G. B. M., Demenois, J., le Maire, G., Cardinael, R., Arias-Navarro, C., \u0026amp; van Egmond, F. (2024). Towards a modular, multi-ecosystem monitoring, reporting and verification (MRV) framework for soil organic carbon stock change assessment. \u003cem\u003eCarbon Management\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(1). https://doi.org/10.1080/17583004.2024.2410812\u003c/li\u003e\n\u003cli\u003eBeillouin, D., Ben-Ari, T., \u0026amp; Makowski, D. (2020). Erratum: Evidence map of crop diversification strategies at the global scale (Environmental Research Letters (2019) (14) (123001) DOI: 10.1088/1748-9326/ab4449). In \u003cem\u003eEnvironmental Research Letters\u003c/em\u003e (Vol. 15, Issue 1). Institute of Physics Publishing. https://doi.org/10.1088/1748-9326/ab5ffb\u003c/li\u003e\n\u003cli\u003eBelgiu, M., \u0026amp; Csillik, O. (2018). Sentinel-2 cropland mapping using pixel-based and object-based time-weighted dynamic time warping analysis. \u003cem\u003eRemote Sensing of Environment\u003c/em\u003e, \u003cem\u003e204\u003c/em\u003e, 509\u0026ndash;523. https://doi.org/10.1016/j.rse.2017.10.005\u003c/li\u003e\n\u003cli\u003eBennett, V. J. (2017). Effects of Road Density and Pattern on the Conservation of Species and Biodiversity. \u003cem\u003eCurrent Landscape Ecology Reports\u003c/em\u003e, \u003cem\u003e2\u003c/em\u003e(1), 1\u0026ndash;11. https://doi.org/10.1007/s40823-017-0020-6\u003c/li\u003e\n\u003cli\u003eBol\u0026iacute;var-Santamar\u0026iacute;a, S., \u0026amp; Reu, B. (2021). Detection and characterization of agroforestry systems in the Colombian Andes using sentinel-2 imagery. \u003cem\u003eAgroforestry Systems\u003c/em\u003e, \u003cem\u003e95\u003c/em\u003e(3), 499\u0026ndash;514. https://doi.org/10.1007/s10457-021-00597-8\u003c/li\u003e\n\u003cli\u003eBol\u0026iacute;var-Santamar\u0026iacute;a, S., \u0026amp; Reu, B. (2024). Assessing canopy structure in Andean (agro)forests using 3D UAV remote sensing. \u003cem\u003eAgroforestry Systems\u003c/em\u003e, \u003cem\u003e98\u003c/em\u003e(5), 1225\u0026ndash;1241. https://doi.org/10.1007/s10457-023-00865-9\u003c/li\u003e\n\u003cli\u003eBrandt, J., \u0026amp; Stolle, F. (2021). A global method to identify trees outside of closed-canopy forests with medium-resolution satellite imagery. \u003cem\u003eInternational Journal of Remote Sensing\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e(5), 1713\u0026ndash;1737. https://doi.org/10.1080/01431161.2020.1841324\u003c/li\u003e\n\u003cli\u003eBrovelli, M. A., Sun, Y., \u0026amp; Yordanov, V. (2020). Monitoring Forest Change in the Amazon Using Multi-Temporal Remote Sensing Data and Machine Learning Classification on Google Earth Engine. \u003cem\u003eISPRS International Journal of Geo-Information\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(10), 580. https://doi.org/10.3390/ijgi9100580\u003c/li\u003e\n\u003cli\u003eCardinael, R., Chevallier, T., Barth\u0026egrave;s, B. G., Saby, N. P. A., Parent, T., Dupraz, C., Bernoux, M., \u0026amp; Chenu, C. (2015). Impact of alley cropping agroforestry on stocks, forms and spatial distribution of soil organic carbon \u0026mdash; A case study in a Mediterranean context. \u003cem\u003eGeoderma\u003c/em\u003e, \u003cem\u003e259\u0026ndash;260\u003c/em\u003e, 288\u0026ndash;299. https://doi.org/10.1016/j.geoderma.2015.06.015\u003c/li\u003e\n\u003cli\u003eCardinael, R., Umulisa, V., Toudert, A., Olivier, A., Bockel, L., \u0026amp; Bernoux, M. (2018). Revisiting IPCC Tier 1 coefficients for soil organic and biomass carbon storage in agroforestry systems. \u003cem\u003eEnvironmental Research Letters\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(12). https://doi.org/10.1088/1748-9326/aaeb5f\u003c/li\u003e\n\u003cli\u003eCaumont, D., \u0026amp; Ivanaj, S. (2017). Chapitre 5. Comment construire et valider une typologie. In \u003cem\u003eAnalyse des donn\u0026eacute;es\u003c/em\u003e (pp. 161\u0026ndash;198). Dunod. https://shs.cairn.info/analyse-des-donnees--9782100758517-page-161?lang=fr\u003c/li\u003e\n\u003cli\u003eChapman, M., Walker, W. S., Cook-Patton, S. C., Ellis, P. W., Farina, M., Griscom, B. W., \u0026amp; Baccini, A. (2020). Large climate mitigation potential from adding trees to agricultural lands. \u003cem\u003eGlobal Change Biology\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(8), 4357\u0026ndash;4365. https://doi.org/10.1111/gcb.15121\u003c/li\u003e\n\u003cli\u003eCharles, E., \u0026amp; Olson, J. (2008). IS 80% ACCURACY GOOD ENOUGH? \u003cem\u003ePecora 17 \u0026ndash; The Future of Land Imaging\u0026hellip;Going Operational\u003c/em\u003e. https://www.asprs.org/a/publications/proceedings/pecora17/0026.pdf Accessed November 2024\u003c/li\u003e\n\u003cli\u003eCoble, A. P., Contosta, A. R., Smith, R. G., Siegert, N. W., Vadeboncoeur, M., Jennings, K. A., Stewart, A. J., \u0026amp; Asbjornsen, H. (2020). Influence of forest-to-silvopasture conversion and drought on components of evapotranspiration. \u003cem\u003eAgriculture, Ecosystems \u0026amp; Environment\u003c/em\u003e, \u003cem\u003e295\u003c/em\u003e, 106916. https://doi.org/10.1016/j.agee.2020.106916\u003c/li\u003e\n\u003cli\u003eDagar, J. C., Gupta, S. R., \u0026amp; Dimobe, K. (2024). Agroforestry from a Global Perspective: Recent Developments, Technological Advancements, and Emerging Research Trends. In \u003cem\u003eAgroforestry Solutions for Climate Change and Environmental Restoration\u003c/em\u003e (pp. 3\u0026ndash;32). Springer Nature Singapore. https://doi.org/10.1007/978-981-97-5004-7_1\u003c/li\u003e\n\u003cli\u003eDoneus, N., \u0026amp; Doneus, M. (2024). Intensively Cultivated Roman Villae Estates: Case Study of Medulin Bay (Istria, Croatia). \u003cem\u003eArchaeological Prospection\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(2), 75\u0026ndash;97. https://doi.org/10.1002/arp.1933\u003c/li\u003e\n\u003cli\u003eEscobar-L\u0026oacute;pez, A., Castillo-Santiago, M. \u0026Aacute;., Mas, J. F., Hern\u0026aacute;ndez-Stefanoni, J. L., \u0026amp; L\u0026oacute;pez-Mart\u0026iacute;nez, J. O. (2024). Identification of coffee agroforestry systems using remote sensing data: a review of methods and sensor data. \u003cem\u003eGeocarto International\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e(1). https://doi.org/10.1080/10106049.2023.2297555\u003c/li\u003e\n\u003cli\u003eFAO. (1999). CHAPTER I: AGROFORESTRY PARKLAND SYSTEMS. In \u003cem\u003eAgroforestry parklands in sub-Saharan Africa (FAO conservation guide)\u003c/em\u003e. Food \u0026amp; Agriculture Organization of the United Nations (FAO). https://www.fao.org/4/x3940e/X3940E11.htm \u003cem\u003eAccessed 2024\u003c/em\u003e\u003c/li\u003e\n\u003cli\u003eFAO. (2020). \u003cem\u003eGlobal Forest Resources Assessment 2020\u003c/em\u003e. https://www.fao.org/interactive/forest-resources-assessment/2020/en/. Accessed November 2024\u003c/li\u003e\n\u003cli\u003eFleiss, J. L. (1971). Measuring nominal scale agreement among many raters. \u003cem\u003ePsychological Bulletin\u003c/em\u003e, \u003cem\u003e76\u003c/em\u003e(5), 378\u0026ndash;382. https://doi.org/10.1037/h0031619\u003c/li\u003e\n\u003cli\u003eFriday, K. S. ., Drilling, M. Elmo., \u0026amp; Garrity, D. P. . (1999). \u003cem\u003eImperata grassland rehabilitation using agroforestry and assisted natural regeneration\u003c/em\u003e. International Centre for Research in Agroforestry, Southeast Asian Regional Research Programme.\u003c/li\u003e\n\u003cli\u003eGaspar, P., Mes\u0026iacute;as, F. J., Escribano, M., \u0026amp; Pulido, F. (2009). Sustainability in Spanish Extensive Farms (Dehesas): An Economic and Management Indicator-Based Evaluation. \u003cem\u003eRangeland Ecology \u0026amp; Management\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e(2), 153\u0026ndash;162. https://doi.org/10.2111/07-135.1\u003c/li\u003e\n\u003cli\u003eGomez-del-Campo, M., Connor, D. J., \u0026amp; Trentacoste, E. R. (2017). Long-term Effect of Intra-Row Spacing on Growth and Productivity of Super-High Density Hedgerow Olive Orchards (cv. Arbequina). \u003cem\u003eFrontiers in Plant Science\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e. https://doi.org/10.3389/fpls.2017.01790\u003c/li\u003e\n\u003cli\u003eGuirado, E., Tabik, S., Alcaraz-Segura, D., Cabello, J., \u0026amp; Herrera, F. (2017). Deep-learning Versus OBIA for Scattered Shrub Detection with Google Earth Imagery: Ziziphus lotus as Case Study. \u003cem\u003eRemote Sensing\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(12), 1220. https://doi.org/10.3390/rs9121220\u003c/li\u003e\n\u003cli\u003eHussain, M. H., Abuhani, D. A., Khan, J., ElMohandes, M., Zualkernan, I., \u0026amp; Ali, T. (2023). A Light-Weight Cropland Mapping Model Using Satellite Imagery. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(15), 6729. https://doi.org/10.3390/s23156729\u003c/li\u003e\n\u003cli\u003eIvezić, V., Yu, Y., \u0026amp; Werf, W. van der. (2021). Crop Yields in European Agroforestry Systems: A Meta-Analysis. In \u003cem\u003eFrontiers in Sustainable Food Systems\u003c/em\u003e (Vol. 5). Frontiers Media S.A. https://doi.org/10.3389/fsufs.2021.606631\u003c/li\u003e\n\u003cli\u003eKuyah, S., Whitney, C. W., Jonsson, M., Sileshi, G. W., \u0026Ouml;born, I., Muthuri, C. W., \u0026amp; Luedeling, E. (2019). Agroforestry delivers a win-win solution for ecosystem services in sub-Saharan Africa. A meta-analysis. \u003cem\u003eAgronomy for Sustainable Development\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e(5). https://doi.org/10.1007/s13593-019-0589-8\u003c/li\u003e\n\u003cli\u003eLandis, J. R., \u0026amp; Koch, G. G. (1977a). An Application of Hierarchical Kappa-type Statistics in the Assessment of Majority Agreement among Multiple Observers. \u003cem\u003eBiometrics\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(2), 363. https://doi.org/10.2307/2529786\u003c/li\u003e\n\u003cli\u003eLandis, J. R., \u0026amp; Koch, G. G. (1977b). The Measurement of Observer Agreement for Categorical Data. \u003cem\u003eBiometrics\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(1), 159\u0026ndash;174. https://doi.org/10.2307/2529310\u003c/li\u003e\n\u003cli\u003eLaso Bayas, J. C., Lesiv, M., Waldner, F., Schucknecht, A., Duerauer, M., See, L., Fritz, S., Fraisl, D., Moorthy, I., McCallum, I., Perger, C., Danylo, O., Defourny, P., Gallego, J., Gilliams, S., Akhtar, I. ul H., Baishya, S. J., Baruah, M., Bungnamei, K., \u0026hellip; Wilson, J. W. (2017). A global reference database of crowdsourced cropland data collected using the Geo-Wiki platform. \u003cem\u003eScientific Data\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(1), 170136. https://doi.org/10.1038/sdata.2017.136\u003c/li\u003e\n\u003cli\u003eL.C., V. D., \u0026amp; L., F. (2023). \u003cem\u003eOne Million Voices Global Review: A review and analysis of existing citizen science initiatives and projects supporting agroecology and agroecological transitions\u003c/em\u003e. Center for International Forestry Research (CIFOR). https://doi.org/10.17528/cifor-icraf/008921\u003c/li\u003e\n\u003cli\u003eLesiv, M., Laso Bayas, J. C., See, L., Duerauer, M., Dahlia, D., Durando, N., Hazarika, R., Kumar Sahariah, P., Vakolyuk, M., Blyshchyk, V., Bilous, A., Perez‐Hoyos, A., Gengler, S., Prestele, R., Bilous, S., Akhtar, I. ul H., Singha, K., Choudhury, S. B., Chetri, T., \u0026hellip; Fritz, S. (2019). Estimating the global distribution of field size using crowdsourcing. \u003cem\u003eGlobal Change Biology\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e(1), 174\u0026ndash;186. https://doi.org/10.1111/gcb.14492\u003c/li\u003e\n\u003cli\u003eLesiv, M., Schepaschenko, D., Buchhorn, M., See, L., D\u0026uuml;rauer, M., Georgieva, I., Jung, M., Hofhansl, F., Schulze, K., Bilous, A., Blyshchyk, V., Mukhortova, L., Brenes, C. L. M., Krivobokov, L., Ntie, S., Tsogt, K., Pietsch, S. A., Tikhonova, E., Kim, M., \u0026hellip; Fritz, S. (2022). Global forest management data for 2015 at a 100 m resolution. \u003cem\u003eScientific Data\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(1). https://doi.org/10.1038/s41597-022-01332-3\u003c/li\u003e\n\u003cli\u003eLiao, C. (2018). Quantifying multi-scale pastoral mobility: Developing a metrics system and using GPS-Tracking data for evaluation. \u003cem\u003eJournal of Arid Environments\u003c/em\u003e, \u003cem\u003e153\u003c/em\u003e, 88\u0026ndash;97. https://doi.org/10.1016/j.jaridenv.2017.12.012\u003c/li\u003e\n\u003cli\u003eLiao, C., Clark, P. E., DeGloria, S. D., \u0026amp; Barrett, C. B. (2017). Complexity in the spatial utilization of rangelands: Pastoral mobility in the Horn of Africa. \u003cem\u003eApplied Geography\u003c/em\u003e, \u003cem\u003e86\u003c/em\u003e, 208\u0026ndash;219. https://doi.org/10.1016/j.apgeog.2017.07.003\u003c/li\u003e\n\u003cli\u003eLillesand, T., W. Kiefer, R., \u0026amp; Chipman, J. (2015). Concepts and Foundations of Remote Sensing 1. In WILEY (Ed.), \u003cem\u003eRemote Sensing and Image Interpretation\u003c/em\u003e (7th Edition, pp. 59\u0026ndash;84).\u003c/li\u003e\n\u003cli\u003eMcHugh, M. L. (2012). Interrater reliability: the kappa statistic. \u003cem\u003eBiochemia Medica\u003c/em\u003e, 276\u0026ndash;282. https://doi.org/10.11613/BM.2012.031\u003c/li\u003e\n\u003cli\u003eMenzel, C. M., \u0026amp; Le Lagadec, M. D. (2017). Can the productivity of mango orchards be increased by using high-density plantings? \u003cem\u003eScientia Horticulturae\u003c/em\u003e, \u003cem\u003e219\u003c/em\u003e, 222\u0026ndash;263. https://doi.org/10.1016/j.scienta.2016.11.041\u003c/li\u003e\n\u003cli\u003eNair, P. K. R. . (1993). An introduction to agroforestry. Kluwer Academic Publishers in cooperation with International Centre for Research in Agroforestry. https://apps.worldagroforestry.org/Units/Library/Books/PDFs/32_An_introduction_to_agroforestry.pdf?n Accessed 2024\u003c/li\u003e\n\u003cli\u003eNdao, B., Leroux, L., Gaetano, R., Diouf, A. A., Soti, V., B\u0026eacute;gu\u0026eacute;, A., Mbow, C., \u0026amp; Sambou, B. (2021). Landscape heterogeneity analysis using geospatial techniques and a priori knowledge in Sahelian agroforestry systems of Senegal. \u003cem\u003eEcological Indicators\u003c/em\u003e, \u003cem\u003e125\u003c/em\u003e. https://doi.org/10.1016/j.ecolind.2021.107481\u003c/li\u003e\n\u003cli\u003e\u0026Ouml;llerer, K., Varga, A., Kirby, K., Demeter, L., Bir\u0026oacute;, M., B\u0026ouml;l\u0026ouml;ni, J., \u0026amp; Moln\u0026aacute;r, Z. (2019). Beyond the obvious impact of domestic livestock grazing on temperate forest vegetation \u0026ndash; A global review. \u003cem\u003eBiological Conservation\u003c/em\u003e, \u003cem\u003e237\u003c/em\u003e, 209\u0026ndash;219. https://doi.org/10.1016/j.biocon.2019.07.007\u003c/li\u003e\n\u003cli\u003ePardon, P., Reubens, B., Reheul, D., Mertens, J., De Frenne, P., Coussement, T., Janssens, P., \u0026amp; Verheyen, K. (2017). Trees increase soil organic carbon and nutrient availability in temperate agroforestry systems. \u003cem\u003eAgriculture, Ecosystems \u0026amp; Environment\u003c/em\u003e, \u003cem\u003e247\u003c/em\u003e, 98\u0026ndash;111. https://doi.org/10.1016/j.agee.2017.06.018\u003c/li\u003e\n\u003cli\u003ePinto-Correia, T., Ribeiro, N., \u0026amp; S\u0026aacute;-Sousa, P. (2011). Introducing the montado, the cork and holm oak agroforestry system of Southern Portugal. \u003cem\u003eAgroforestry Systems\u003c/em\u003e, \u003cem\u003e82\u003c/em\u003e(2), 99\u0026ndash;104. https://doi.org/10.1007/s10457-011-9388-1\u003c/li\u003e\n\u003cli\u003eRallo, L., Barranco, D., Castro‐Garc\u0026iacute;a, S., Connor, D. J., G\u0026oacute;mez del Campo, M., \u0026amp; Rallo, P. (2013). High‐Density Olive Plantations. In \u003cem\u003eHorticultural Reviews Volume 41\u003c/em\u003e (pp. 303\u0026ndash;384). Wiley. https://doi.org/10.1002/9781118707418.ch07\u003c/li\u003e\n\u003cli\u003eRamachandran Nair, P. K., Mohan Kumar, B., \u0026amp; Nair, V. D. (2009). Agroforestry as a strategy for carbon sequestration. \u003cem\u003eJournal of Plant Nutrition and Soil Science\u003c/em\u003e, \u003cem\u003e172\u003c/em\u003e(1), 10\u0026ndash;23. https://doi.org/10.1002/jpln.200800030\u003c/li\u003e\n\u003cli\u003eRosenstock, T. S., Wilkes, A., Jallo, C., Namoi, N., Bulusu, M., Suber, M., Mboi, D., Mulia, R., Simelton, E., Richards, M., Gurwick, N., \u0026amp; Wollenberg, E. (2019). Making trees count: Measurement and reporting of agroforestry in UNFCCC national communications of non-Annex I countries. \u003cem\u003eAgriculture, Ecosystems and Environment\u003c/em\u003e, \u003cem\u003e284\u003c/em\u003e. https://doi.org/10.1016/j.agee.2019.106569\u003c/li\u003e\n\u003cli\u003eRoupsard, O., Audebert, A., Ndour, A. P., Clermont-Dauphin, C., Agbohessou, Y., Sanou, J., Koala, J., Faye, E., Sambakhe, D., Jourdan, C., le Maire, G., Tall, L., Sanogo, D., Seghieri, J., Cournac, L., \u0026amp; Leroux, L. (2020). How far does the tree affect the crop in agroforestry? New spatial analysis methods in a Faidherbia parkland. \u003cem\u003eAgriculture, Ecosystems and Environment\u003c/em\u003e, \u003cem\u003e296\u003c/em\u003e. https://doi.org/10.1016/j.agee.2020.106928\u003c/li\u003e\n\u003cli\u003eSaah, D., Johnson, G., Ashmall, B., Tondapu, G., Tenneson, K., Patterson, M., Poortinga, A., Markert, K., Quyen, N. H., San Aung, K., Schlichting, L., Matin, M., Uddin, K., Aryal, R. R., Dilger, J., Lee Ellenburg, W., Flores-Anderson, A. I., Wiell, D., Lindquist, E., \u0026hellip; Chishtie, F. (2019). Collect Earth: An online tool for systematic reference data collection in land cover and use applications. \u003cem\u003eEnvironmental Modelling \u0026amp; Software\u003c/em\u003e, \u003cem\u003e118\u003c/em\u003e, 166\u0026ndash;171. https://doi.org/10.1016/j.envsoft.2019.05.004\u003c/li\u003e\n\u003cli\u003eS\u0026aacute;nchez-Navarro, V., Mart\u0026iacute;nez-Mart\u0026iacute;nez, S., Acosta, J. A., Almagro, M., Mart\u0026iacute;nez-Mena, M., Boix-Fayos, C., D\u0026iacute;az-Pereira, E., Temnani, A., Berrios, P., P\u0026eacute;rez-Pastor, A., \u0026amp; Zornoza, R. (2023). Soil greenhouse gas emissions and crop production with implementation of alley cropping in a Mediterranean citrus orchard. \u003cem\u003eEuropean Journal of Agronomy\u003c/em\u003e, \u003cem\u003e142\u003c/em\u003e, 126684. https://doi.org/10.1016/j.eja.2022.126684\u003c/li\u003e\n\u003cli\u003eSarti, M., Ciolfi, M., Lauteri, M., Paris, P., \u0026amp; Chiocchini, F. (2021). Trees outside forest in Italian agroforestry landscapes: detection and mapping using sentinel-2 imagery. \u003cem\u003eEuropean Journal of Remote Sensing\u003c/em\u003e, \u003cem\u003e54\u003c/em\u003e(1), 609\u0026ndash;623. https://doi.org/10.1080/22797254.2021.1986678\u003c/li\u003e\n\u003cli\u003eStanimirova, R., Tarrio, K., Turlej, K., McAvoy, K., Stonebrook, S., Hu, K. T., Ar\u0026eacute;valo, P., Bullock, E. L., Zhang, Y., Woodcock, C. E., Olofsson, P., Zhu, Z., Barber, C. P., Souza, C. M., Chen, S., Wang, J. A., Mensah, F., Calder\u0026oacute;n-Loor, M., Hadjikakou, M., \u0026hellip; Friedl, M. A. (2023). A global land cover training dataset from 1984 to 2020. \u003cem\u003eScientific Data\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1). https://doi.org/10.1038/s41597-023-02798-5\u003c/li\u003e\n\u003cli\u003eStory, M., \u0026amp; Congalton, R. G. (1986). Accuracy assessment: a user\u0026rsquo;s perspective. \u003cem\u003ePhotogrammetric Engineering and Remote Sensing\u003c/em\u003e, \u003cem\u003e52\u003c/em\u003e, 397\u0026ndash;399. https://api.semanticscholar.org/CorpusID:133251404\u003c/li\u003e\n\u003cli\u003eTa\u0026iuml;bi, A. N., Hannani, M. El, Khalki, Y. El, \u0026amp; Ballouche, A. (2019). The agroforestry parks of Azilal (Morocco): a centuries-old and still living landscape construction. \u003cem\u003eRevue de G\u0026eacute;ographie Alpine\u003c/em\u003e, \u003cem\u003e107\u0026ndash;3\u003c/em\u003e. https://doi.org/10.4000/rga.6612\u003c/li\u003e\n\u003cli\u003eTerasaki Hart, D. E., Yeo, S., Almaraz, M., Beillouin, D., Cardinael, R., Garcia, E., Kay, S., Lovell, S. T., Rosenstock, T. S., Sprenkle-Hyppolite, S., Stolle, F., Suber, M., Thapa, B., Wood, S., \u0026amp; Cook-Patton, S. C. (2023). Priority science can accelerate agroforestry as a natural climate solution. \u003cem\u003eNature Climate Change\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(11), 1179\u0026ndash;1190. https://doi.org/10.1038/s41558-023-01810-5\u003c/li\u003e\n\u003cli\u003eThapa, B., Lovell, S., \u0026amp; Wilson, J. (2023). Remote sensing and machine learning applications for aboveground biomass estimation in agroforestry systems: a review. \u003cem\u003eAgroforestry Systems\u003c/em\u003e, \u003cem\u003e97\u003c/em\u003e(6), 1097\u0026ndash;1111. https://doi.org/10.1007/s10457-023-00850-2\u003c/li\u003e\n\u003cli\u003eTorralba, M., Fagerholm, N., Burgess, P. J., Moreno, G., \u0026amp; Plieninger, T. (2016). Do European agroforestry systems enhance biodiversity and ecosystem services? A meta-analysis. \u003cem\u003eAgriculture, Ecosystems \u0026amp; Environment\u003c/em\u003e, \u003cem\u003e230\u003c/em\u003e, 150\u0026ndash;161. https://doi.org/10.1016/j.agee.2016.06.002\u003c/li\u003e\n\u003cli\u003eVaupel, A., Bednar, Z., Herwig, N., Hommel, B., Moran-Rodas, V. E., \u0026amp; Beule, L. (2023). Tree-distance and tree-species effects on soil biota in a temperate agroforestry system. \u003cem\u003ePlant and Soil\u003c/em\u003e, \u003cem\u003e487\u003c/em\u003e(1\u0026ndash;2), 355\u0026ndash;372. https://doi.org/10.1007/s11104-023-05932-9\u003c/li\u003e\n\u003cli\u003eZayani, I., Ammari, M., Ben Allal, L., \u0026amp; Bouhafa, K. (2023). Agroforestry olive orchards for soil organic carbon storage: Case of Saiss, Morocco. \u003cem\u003eHeliyon\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(12), e22910. https://doi.org/10.1016/j.heliyon.2023.e22910\u003c/li\u003e\n\u003cli\u003eZhu, X., Liu, W., Chen, J., Bruijnzeel, L. A., Mao, Z., Yang, X., Cardinael, R., Meng, F. R., Sidle, R. C., Seitz, S., Nair, V. D., Nanko, K., Zou, X., Chen, C., \u0026amp; Jiang, X. J. (2020). Reductions in water, soil and nutrient losses and pesticide pollution in agroforestry practices: a review of evidence and processes. \u003cem\u003ePlant and Soil\u003c/em\u003e, \u003cem\u003e453\u003c/em\u003e(1\u0026ndash;2), 45\u0026ndash;86. https://doi.org/10.1007/s11104-019-04377-3\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"agroforestry, classification tree, classification errors, photo-interpretation","lastPublishedDoi":"10.21203/rs.3.rs-6000362/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6000362/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEffective and large-scale monitoring of agroforestry (AF) systems is essential to assess the environmental benefits of agroforestry and support sustainable land management strategies. However, a standardized method for classifying these systems using satellite imagery is still missing. Here, we present a novel operational framework to classify agroforestry systems into three categories\u0026mdash;Alley cropping, Scattered agroforestry, and Hedgerows\u0026mdash;and to distinguish these systems from Cropland without trees, Natural trees, and Orchards. The proposed procedure relies on a classification tree based on photo-interpretation of satellite imagery. The accuracy and robustness of this classification tree were evaluated by five interpreters across 300 agroforestry and non-agroforestry plots spanning all continents. Results show that the classification tree accurately distinguishes agroforestry categories from one another and from non-agroforestry systems, with an overall accuracy ranging from 0.75 to 0.81 depending on the interpreter. After eliminating the interpreters\u0026rsquo; errors, the potential classification accuracy increases to 0.86. While hedgerows were accurately classified in most cases with an omission error of 2% and no commission error (0%), the study revealed challenges in differentiating between Alley cropping and Orchards which were frequently confounded. Similarly, plots with Scattered agroforestry were also misclassified as Natural trees leading to a commission error of 19% for this class. Despite these limitations, the proposed classification tree represents a valuable tool for large-scale monitoring of agroforestry systems. Future adaptations of this framework could address regional specificities, further improving its applicability and accuracy.\u003c/p\u003e","manuscriptTitle":"Classification of agroforestry systems by photo-interpretation of satellite imagery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-02-19 09:21:35","doi":"10.21203/rs.3.rs-6000362/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e048a4cf-cced-4297-ac38-c4e9df33ffed","owner":[],"postedDate":"February 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-07-07T11:53:49+00:00","versionOfRecord":[],"versionCreatedAt":"2025-02-19 09:21:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6000362","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6000362","identity":"rs-6000362","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}