Efficient Argan Tree Deforestation Detection Using Sentinel-2 Time Series and Machine Learning

Abstract

The argan tree (Argania spinosa) is a rare species native to southwestern Morocco, valued for its fruit, which produces argan oil, a highly prized natural product with nutritional, health, and cosmetic benefits. However, increasing deforestation poses a significant threat to its survival. This study monitors changes in an argan forest near Agadir, Morocco, from 2017 to 2023 using Sentinel-2 satellite imagery and advanced image processing algorithms. Various machine learning models were evaluated for argan tree detection, with LightGBM achieving the highest accuracy when trained on a dataset integrating spectral bands, temporal features, and vegetation indices information. The model achieved 100% accuracy on tabular test data and 85% on image-based test data. The generated deforestation maps estimated an approximate forest loss of 2.86% over six years. This study explores methods to enhance detection accuracy, provides valuable statistical data for deforestation mitigation, and highlights the critical role of remote sensing, advanced image processing, and artificial intelligence in environmental monitoring and conservation, particularly in argan forests.

1. Introduction

The argan tree (Argania spinosa (L.) Skeels) is a rare species endemic to southwestern Morocco [1]. Argan trees play a critical role in the environmental, social, and economic landscape. Known for their resilience in arid conditions, the argan trees contribute to biodiversity by supporting a rich network of plant and animal life while also enhancing soil stability [2]. Beyond their ecological significance, they also serve as a source of livelihood and income for local communities through valuable products like argan oil and timber [3]. The argan oil industry significantly contributes to the Moroccan economy, as highlighted in

Table 1. Argan oil market analysis: global and Morocco size and forecast (2019–2030) [4,5].

	Moroccan Market Size	Global Market Size
Revenue, 2019 (US$M)	$37.4	$233.0
Forecast, 2030 (US$M)	$97.0	$712.5

, which presents the market size of Moroccan and global argan oil in 2019, along with forecasts for 2030 [4,5]. The growing global demand for argan oil across the cosmetics, skincare, and culinary sectors further underscores its increasing economic importance.

The argan forest considered in this study is located in the Souss region, centered around the city of Agadir (Morocco), an area distinguished by its agricultural sector, which is a leading economic sector alongside tourism. This argan forest is one of the largest in the country [6] and includes a section that was declared a Biosphere Reserve by UNESCO in 1998, dedicated to protecting this unique habitat [7]. However, accurately assessing the precise distribution and abundance of argan trees in the region remains challenging [8].

At present, the argan forest has been threatened by deforestation due to several factors [9]. On the one hand, agricultural expansion, urban encroachment [6], massive logging for producing firewood and charcoal, not to mention climate change and drought [8,10]. As a result, the argan forests have lost approximately 50% of their total area over the past century (

Table 2. Changes in argan area and density (argan trees per ha): historical data from the early 20th century to 2016 [11].

	Early 20th Century	2016
Area (ha)	1,500,000	829,087
Density (trees/ha)	100	30

), with an estimated annual loss of 600 hectares [11]. This has raised alarm bells regarding the possible extinction of this magnificent tree in the future [12]. It is therefore crucial for resource conservation that significant efforts are made to create accurate maps, essential for the monitoring and proper management of land use and land cover (LULC) in this strategic zone, given its role in food security and Morocco’s environmental and economic development [13].

In recent years, remote sensing technologies and satellite imagery have provided the capacity to collect accurate, comprehensive, and wide-ranging data, facilitating studies related to forest mapping and the accurate monitoring of forest habitat change and deforestation [14,15,16,17]. In particular, multispectral and multitemporal data from the European Space Agency’s Sentinel-2 satellites have contributed to the success of this technology in these applications [18,19,20].

Tree detection studies using Sentinel-2 data fall into two main categories: those relying solely on single-date multispectral information and those incorporating both multispectral and multi-temporal data [21]. Two primary methods are used to assess spectral band significance. The first calculates feature importance scores from classifiers, offering a quantitative approach to select optimal bands, though interpretability and classification performance can sometimes pose challenges [22]. The second method involves individual classification experiments with each spectral feature, providing a clearer understanding of their impact [23,24]. Incorporating multi-temporal data is crucial due to the “same spectra but different objects, same objects but different spectra” phenomenon, caused by tree phenological cycles, which can limit detection accuracy when using single-date data [21,25]. The multi-temporal analysis allows for better tracking of these phenological stages, enhancing detection accuracy [26]. However, cloud contamination can interfere with this process [27]. Integrating multi-source data has been suggested, but this approach requires extensive preprocessing and sensor calibration [28,29]. In summary, using multispectral and multi-temporal Sentinel-2 data significantly improves tree detection accuracy, with results strongly dependent on information selection strategies [21,26].

In studies on tree detection and forest mapping using Sentinel-2 data, various classification algorithms have been employed, encompassing both machine learning and deep learning techniques [18], with supervised classification generally preferred over unsupervised approaches [30]. Before undertaking any remote sensing task, it is essential to identify the most appropriate approach for data analysis, especially with Sentinel-2. There are three primary approaches: pixel-based [31], patch-based [32], and object-based [31], which differ primarily in the basic unit of analysis. Pixel-based and patch-based approaches often employ machine learning algorithms for LULC classification, with Random Forest (RF) and Support Vector Machine (SVM) being the most frequently used. These algorithms have shown promising results, often achieving an Overall Accuracy (OA) exceeding 80% [18]. Other studies with similar objectives have utilized other machine learning algorithms, such as Maximum Likelihood Classification (MLP), Artificial Neural Networks (ANN), Decision Trees, k-Nearest Neighbors (KNN), and Bayes models [18]. It is worth noting that pixel-based methods, while commonly used, can be prone to noise. Patch-based or object-based approaches are generally preferable, especially for change detection tasks [33,34]. Object-based methods, in particular, often take advantage of deep learning algorithms, which were first introduced to remote sensing in 2014 [20]. Since then, deep learning has gained significant attention due to its success in this domain [35]. Among the most commonly used deep learning models are Convolutional Neural Networks (CNN), which have a demonstrated OA exceeding 90% in LULC classification and mapping tasks using Sentinel-2 data [17,36]. Overall, deep learning algorithms outperform traditional machine learning algorithms in terms of accuracy due to their superior feature extraction capabilities, non-linear modeling, and high-level semantic segmentation [37,38]. Nevertheless, the efficiency of machine learning models remains relevant, as they typically require less training data compared to deep learning models, making them well-suited for smaller datasets [39]. Finally, regardless of the type of algorithm chosen for a remote sensing task, factors such as data preprocessing and algorithm parameterization can significantly enhance accuracy [18].

Finally, it is important to briefly review the role of remote sensing in detecting changes in the same geographical area over time, a key focus of this study [40]. Change detection methods using multispectral satellite data can be categorized into four main types [41]: Algebra-Based Methods [42], Statistics-Based Methods [43,44], Transformation-Based Methods [42], and Deep-Learning-Based Methods [17,45]. For deforestation mapping, two primary approaches are commonly used. The first involves classifying images for each time period into multiple land-cover labels (e.g., trees, bare soil, urban areas, water) using machine learning classifiers [46] or deep learning algorithms [47,48]. By comparing these classification maps, deforestation maps are generated. However, this approach is prone to error propagation from the classification maps, underscoring the need to improve classification accuracy. The second approach directly compares two images from different dates. Deep neural network models, such as Improved UNet++ [49], are specifically designed for this task. These models use multi-temporal data in an end-to-end manner, analyzing spectral, spatial, and structural features to produce a final change map. This method can also be accomplished using machine learning algorithms [50].

Although numerous studies have explored remote sensing data and developed algorithms for forest mapping, few have specifically focused on argan forests, either for general mapping or deforestation detection, largely due to the limited availability of argan tree-related data [51]. Two studies focused on mapping argan trees near Essaouira, close to Agadir, using the NDVI index derived from Sentinel-2 time series data and the Support Vector Machine (SVM) algorithm, achieving overall accuracies of 89.78% [52] and 92.60% [53], respectively. Another study mapped argan using Sentinel-1 time series, Sentinel-2 data, and three machine learning algorithms, in addition to a layer of the missing Shuttle Radar Topography Mission Digital Elevation Model (DEM), achieving the highest accuracy of 93.25% [54]. Building on this foundation, our team researched mapping argan deforestation using deep learning algorithms and a patch-based approach [20,32,55]. This study aims to further enhance these results.

Recognizing the ecological and economic importance of argan trees and the threats they face, this study aims to map argan forests and calculate the deforestation rate using remote sensing and machine learning tools. Specifically, we create argan tree maps for 2017/2018 and 2022/2023, comparing these maps to generate a map and calculate the deforestation rate over this period. Given the limited availability of argan-specific data, machine learning algorithms were selected for their efficiency, faster data analysis capabilities, and minimal computational requirements. This work seeks to provide specialists with valuable insights and statistics to inform strategies for mitigating argan forest deforestation.

2. Materials and Methods

Figure 1 shows a flow chart summarizing the most important steps of the methodology followed, which include:

• Data acquisition and preprocessing.
• Tree detection.
• Change detection.

2.1. Study Area

The study area was the Admine forest located in the Souss region, near the city of Agadir in southern Morocco (Figure 2). It covers approximately 128 square kilometers. This forest is part of the UNESCO-declared biosphere reserve dedicated to the protection of the argan tree. From a geological perspective, the area is situated in the alluvial basin of the Souss valley, which separates the Sahara Desert from the Anti-Atlas Mountains [55]. The region is characterized by hot summers and mild winters, with a dry and semi-arid climate. In recent years, it has experienced a reduction in rainfall, leading to long-term environmental changes in the area. The predominant natural vegetation in the Souss region consists of savanna landscapes, with the argan tree being the dominant species [20]. The height of argan trees ranges from 2 to 10 m, and the canopy area varies from tree to tree, ranging from less than one square meter for small trees to more than 50 square meters for very large trees [56,57]. The study area also comprises buildings, roads, farms, and greenhouses. It is surrounded on all sides by residential areas belonging to Greater Agadir, with Agadir International Airport located centrally. The area was chosen because it has unfortunately experienced significant deforestation in recent years, which may threaten the future disappearance of argan in the region.

2.2. Sentinel-2 Imagery

Sentinel-2 images were used in this study. The mission consists of a constellation of two satellites, Sentinel-2A and -2B, orbiting in a sun-synchronous polar orbit, phased 180 degrees apart from each other. Each Sentinel-2 carries onboard a MultiSpectral Instrument (MSI) with 13 spectral bands in the visible, near-infrared, and short-wave infrared wavelengths and spatial resolutions ranging from 10 m to 60 m. Sentinel-2A and -2B provide images periodically every 10 days at the equator with one satellite and every 5 days with two satellites [18].

Sentinel-2 offers images at two distinct levels of preprocessing: Level 1C (Top-of-Atmosphere reflectance) and Level 2A (Bottom-of-Atmosphere reflectance). While Level 1C data undergo geometric and radiometric corrections, they lack atmospheric adjustments. This type of data is typically used in studies where atmospheric effects are not a primary concern or can be addressed at a later stage. On the other hand, Level 2A data are comprehensively processed, including geometric, radiometric, and atmospheric aspects. This level is preferred for tasks requiring precise surface reflectance. It is worth noting that Level-1C data consist of 13 bands (

Table 3. Sentinel-2 Level 1C spectral bands [59].

Band Name	Description	Wavelength (nm) S2A/S2B	Pixel Size (m)
B1	Aerosols	443.9/442.3	60
B2	Blue	496.6/492.1	10
B3	Green	560/559	10
B4	Red	664.5/665	10
B5	Red Edge 1	703.9/703.8	20
B6	Red Edge 2	740.2/739.1	20
B7	Red Edge 3	782.5/779.7	20
B8	NIR	835.1/833	10
B8A	Red Edge 4	864.8/864	20
B9	Water vapor	945/943.2	60
B10	Cirrus	1373.5/1376.9	60
B11	SWIR 1	1613.7/1610.4	20
B12	SWIR 2	2202.4/2185.7	20

), whereas Level-2A data include only 12 bands. The difference is due to the removal of Band 10, which provides information on cirrus clouds. This band is considered unnecessary for surface reflectance analysis [58]. Therefore, it is removed or excluded from the Level-2A dataset to streamline the product and optimize it for terrestrial applications. All Sentinel-2 bands were resampled to a spatial resolution of 10 m using bilinear interpolation. Sentinel-2 data at both processing levels were used from 1 September 2017 to 31 August 2018 and from 1 September 2022 to 31 August 2023.

2.3. Google Earth Engine

Several cloud computing platforms provide remote sensing data [60,61]: Sentinel Hub [61,62], Open Data Cube [63,64], SEPAL [65,66], JEODPP [67,68], pipsCloud [69,70], OpenEO [71,72], and Google Earth Engine (GEE) [73,74], with GEE being among the most widely used [75].

GEE [73,74] is a powerful online platform widely used for Earth observation and environmental research. It provides an intuitive and user-friendly interface, including a code editor and graphical user tools, making it accessible to researchers across disciplines such as geography, geology, and remote sensing. It hosts an archival database of images from various satellites, such as Sentinel, Landsat, and MODIS, along with geospatial data. This facilitates the creation of diverse maps and offers users the ability to download data, tools for processing, capabilities to sort and filter, mask clouds and their shadows, and data analysis [73,74]. Its popularity is attributed to key features such as being free to use and offering efficient parallel processing capabilities without the need to download data [76]. It enables easy progress using data from different spatial and temporal ranges [77,78], thereby saving a significant amount of time and processing large volumes of data [79].

GEE is widely used for tasks such as forest and deforestation mapping using machine learning algorithms, such as Random Forest [50,80,81]. A portion of this study was conducted using GEE. This included the collection of spectral and time series data from Sentinel-2, the calculation and addition of spectral indices, and the labeling of the data. Finally, the processed data were downloaded in tabular format for training the machine learning models, along with the images to which these models were applied [82].

2.4. Data Acquisition and Pre-Processing

The pre-processing consisted of three basic steps: calculating different spectral indices, creating time series for each life cycle of the argan tree (2017/2018 and 2022/2023), and labeling the data based on ground truth.

2.4.1. Spectral Indices

In the field of remote sensing, spectral indices serve as key tools within multispectral transformation processes [83]. These indices transform satellite sensor reflectance readings into environmentally significant figures. Thanks to the multispectral capabilities of satellite imagery, these transformations provide insights into various phenomena’s current conditions. Vegetation indices, in particular, play a crucial role in both the identification and monitoring of vegetation patterns and in calculating key biophysical attributes of plant covers, including biomass, the leaf area index, and the proportion of absorbed photosynthetically active radiation [83]. Therefore, we decided to include various spectral indices available in the Sentinel hub platform (

Table 4. Spectral indices selected [84].

Abbreviation	Name	Formula
ARVI	Atmospherically Resistant Vegetation Index	${\displaystyle \frac{B8A-B4-y \times (B4-B2)}{B8A+B4-y \times (B4-B2)}}$
EVI	Enhanced Vegetation Index	${\displaystyle \frac{2.5 \times (B8-B4)}{(B8+6.0 \times B4-7.5 \times B2)+1.0}}$
GNDVI	Green Normalized Difference Vegetation Index	${\displaystyle \frac{B8-B3}{B8+B3}}$
NDVI	Normalized Difference Vegetation Index	${\displaystyle \frac{B8-B4}{B8+B4}}$
NDWI	Normalized Difference Water Index	${\displaystyle \frac{B3-B8}{B3+B8}}$
PVI	Perpendicular Vegetation Index	${\displaystyle \frac{B8 -a \ast B4 -b}{\sqrt{a^2+1}}}$
SAVI	Soil Adjusted Vegetation Index	${\displaystyle \frac{(B8-B4) \times (1+L)}{B8+B4+L}}$
SCI	Soil Composition Index	${\displaystyle \frac{B11-B8}{B11+B8}}$

y: quotient derived from the components of atmospheric reflectance in the blue and red channel (commonly set to 0.069) [85,86,87]. a: slope of the soil line (commonly set to 0.149) [85,88,89]. b: gradient of the soil line (commonly set to 0.735) [85,88,89]. L: amount of green vegetation cover (commonly set to 0.725) [85,88,90].

) [84]. Throughout the study, tests were carried out to examine the importance of each index.

These indices were selected based on studies aimed at forest mapping [91,92]. Their purpose in this study was to enhance argan tree detection and forest mapping. The Normalized Difference Vegetation Index (NDVI) is widely used to distinguish vegetation from non-vegetation by comparing red and near-infrared reflectance [93]. The Enhanced Vegetation Index (EVI) accounts for background canopy effects and atmospheric disturbances, increasing sensitivity in regions with high biomass [94]. The Atmospherically Resistant Vegetation Index (ARVI) reduces atmospheric scattering effects by incorporating the blue reflectance band [95]. Meanwhile, the Green Normalized Difference Vegetation Index (GNDVI) provides greater sensitivity to chlorophyll content by leveraging green reflectance instead of red [96]. The Normalized Difference Water Index (NDWI) highlights vegetation water content by utilizing green and near-infrared reflectance [97]. The Perpendicular Vegetation Index (PVI) minimizes soil background influence through a geometric approach in spectral space [98]. The Soil-Adjusted Vegetation Index (SAVI) compensates for soil brightness variations by incorporating a correction factor into the NDVI formula [99]. Finally, the Soil Composition Index (SCI) emphasizes spectral variations due to differences in soil constituents and thus helps to isolate soil reflectance from vegetation signals, improving forest/non-forest discrimination in heterogeneous landscapes [100]. While NDVI, EVI, ARVI, GNDVI, and SAVI mainly differ in their sensitivity to soil background and atmospheric effects, NDWI provides additional insights into vegetation and soil water content, while SCI helps assess soil composition. Collectively, these indices enable a more accurate and robust classification of forested areas [101].

2.4.2. Time Series Construction

The second pre-processing step was creating the time series for each life cycle of the argan tree (phenology). The remotely sensed spectral signature of argan trees varies over time due to several factors. A primary factor is the dependency of the Sentinel-2 imagery on sunlight; changes in sunlight intensity result in variations in spectral information. Additionally, alterations in the phenology contribute to spectral changes. To address this challenge, 24 image capture dates were selected for each year, representing various phenological stages in the argan tree’s life cycle (

Table 5. The image capture days selected to create the time series for the years 2017/2018 and 2022/2023, where two images were chosen from each month that met the criteria, one from the first half and one from the second half.

2017/2018		2022/2023
The First Half	The Second Half	The First Half	The Second Half
02/09/2017	19/09/2017	08/09/2022	28/09/2022
02/10/2017	22/10/2017	06/10/2022	21/10/2022
01/11/2017	21/11/2017	07/11/2022	22/11/2022
08/12/2017	18/12/2017	10/12/2022	20/12/2022
02/01/2018	17/01/2018	01/01/2023	16/01/2023
01/02/2018	19/02/2018	03/02/2023	18/02/2023
11/03/2018	21/03/2018	05/03/2023	20/03/2023
10/04/2018	27/04/2018	09/04/2023	24/04/2023
10/05/2018	15/05/2018	09/05/2023	19/05/2023
06/06/2018	19/06/2018	15/06/2023	25/06/2023
04/07/2018	24/07/2018	05/07/2023	18/07/2023
08/08/2018	18/08/2018	02/08/2023	17/08/2023

). The inhabitants of the study area harvest argan fruits in the summer, during the months of July and August. We therefore considered that the cycle starts at the beginning of September and ends at the end of August. The dates were carefully and precisely selected based on specific criteria, such as choosing two images per month and ensuring that each image is ideal for the study, as shown in Table 5. The study years were selected because 2017 marks the first release of Sentinel-2 Level-2A data, while 2023 corresponds to the year in which the data used in this study were collected [102].

2.4.3. Ground Truth Data

Ground truth data were meticulously collected using tools from the GEE platform. To ensure the accuracy required for a comprehensive analysis, argan trees were precisely geolocated and marked (using a GEE tool) through a detailed examination of high-resolution Google Earth imagery. This identification was based on prior knowledge of the study area and the distinctive characteristics of the argan tree, such as crown size, color, and shadow patterns, which allowed us to distinguish them from other tree species present in the images. Data validation was carried out via field visits to the study area, enhancing the quality and credibility of data (Figure 3). Once the argan trees were identified, the data were downloaded using a custom GEE script. Data collection was conducted in late 2023. It is important to note that we did not have argan data for 2017/2018; therefore, we assumed that the trees present in 2022/2023 were also present in 2017/2018, as no reforestation efforts had occurred between these periods.

The ground truth data were divided into two categories: the Argan, representing argan trees, and the Non-Argan, representing other objects such as buildings, farms, soil, etc. Since accurate identification of argan trees is crucial for the success of this study, two types of data were used: tabular data and image data.

The tabular data (data for training and testing) consisted of spectral information for each selected pixel, along with spectral indices. These data were divided into two categories: single-temporal and multi-temporal tabular data. Single-temporal tabular data were used to train machine learning models on data from a single day, whereas multi-temporal tabular data were used to train models on time-series data, with 24 days from each life cycle selected to form the time series. To ensure a balanced representation, an equal number of samples were selected for each category, with 695 pixels chosen for the Argan category and 695 pixels for the Non-Argan. This balanced sample size ensures that both categories receive equal attention during training, preventing bias toward the dominant class. The specific number of pixels was determined by the maximum amount of data that could be collected within the constraints of the available time frame.

For the imagery data (data for testing only), the aim was to identify the optimal scenario for detecting and mapping argan trees with high accuracy. A small region was completely labeled to create a reference image (Figure 4), enabling the evaluation of different scenarios. The reference image has dimensions of 65 by 90 pixels and includes 688 pixels representing the Argan category and 5162 pixels representing the Non-Argan category, accurately reflecting the 2023 distribution of argan trees.

2.5. Machine Learning Models

The argan tree is a large species with thorny, horizontally spreading branches, resulting in a wide, sprawling appearance. In Sentinel-2 imagery, which has a resolution of 10 m, each argan tree typically corresponds to one to four pixels (see Figure 5). This pixel representation allows the application of machine learning algorithms to classify pixels into two previously defined categories: “Argan” (containing argan trees) or “Non-Argan” (without argan trees).

Eight machine learning algorithms were selected for classifying the study area and detecting argan trees, with the objective of identifying the most effective models for this task: Support Vector Machine (SVM) [103], Decision Tree (DT) [104], Random Forest (RF) [104], LightGBM [105], XGBoost [106], Gaussian Naive Bayes [107], K-Nearest Neighbors (KNN) [108], and Artificial Neural Network (ANN) [108]. All algorithms were supervised and trained on various dataset scenarios. Each model had tunable parameters to optimize performance. Initially, all models were tested with their default parameter settings. Subsequently, parameter adjustments were made to improve the results.

Table 6. Most significant parameters for each model and the tested values.

Machine Learning Model	Basic Parameters	Tested Values
Support Vector Machine	kernel	rbf, poly, linear, sigmoid
Decision Tree	criterion	gini, entropy
	splitter	best, random
Random Forest	criterion	gini, entropy, log_loss
LightGBM	boosting_type	gbdt, dart, goss
XGBoost	booster	gbtree, gblinear, dart
K-Nearest Neighbors	n_neighbors	2, 3, 4, 5, 6, 7, 8, 9, 10
Artificial Neural Network	units (Neurons in hidden layer)	26/24, 18, 13/12, 10

highlights the key parameters for each model and the tested values, except for Gaussian Naive Bayes, which is typically used with its default settings and required no modifications. It is worth noting that other parameters were tested, but they were disregarded as they did not provide any performance improvement.

The SVM, RF, KNN, and ANN algorithms are widely used in tree detection and mapping due to their proven effectiveness in handling spatial and spectral data [18], making them suitable candidates for this study. Additionally, DT, LightGBM, XGBoost, and Gaussian Naive Bayes were selected for their strong classification capabilities, as demonstrated in various remote sensing applications.

Support Vector Machine (SVM) uses the principle of maximizing the margin between classes in a high-dimensional feature space, leading to robust performance, especially in high-dimensional or small-sample scenarios [109]. Decision Tree (DT) constructs a hierarchical structure of if-then rules by recursively partitioning the feature space, offering ease of interpretation due to its rule-based nature [110]. Random Forest (RF) is an ensemble method that aggregates multiple decision trees grown on bootstrap samples, enhancing predictive accuracy and reducing overfitting [111]. LightGBM relies on a gradient boosting framework and employs a leaf-wise tree growth strategy with histogram-based decision rules, enabling faster training on large-scale datasets [112]. Similarly, XGBoost also uses gradient boosting but incorporates second-order gradients and advanced regularization, often achieving high accuracy and efficiency in competition settings [113]. Gaussian Naive Bayes applies Bayes’ theorem under an assumption of feature independence and normal distributions, making it computationally efficient and effective with limited data [114]. K-Nearest Neighbors (KNN) is an instance-based learner that classifies a point based on the majority class among its nearest neighbors, remaining straightforward to implement with minimal explicit training [115]. Finally, artificial neural networks (ANNs) utilize layers of interconnected units or “neurons” that learn complex, often non-linear mappings from inputs to outputs, excelling in tasks involving intricate feature interactions [116].

Differences among these algorithms revolve around their assumptions (e.g., independence in Naive Bayes), interpretability (e.g., Decision Trees and Random Forests), computational efficiency (e.g., LightGBM and XGBoost), and capability to capture non-linear relationships (e.g., ANNs and SVMs), making each algorithm more or less suitable depending on data characteristics and application requirements.

All algorithms were implemented using Python v3.9. For the algorithms SVM, Decision Tree, Random Forest, Gaussian Naive Bayes, and K-Nearest Neighbors, specific functions for each were utilized from the ‘sklearn’ package [117,118]. The ‘lightgbm’ and ‘xgboost’ packages were used for LightGBM and XGBoost algorithms. Lightgbm is a gradient boosting framework developed by Microsoft, utilizing tree-based learning algorithms designed for speed and efficiency [112,119]. Xgboost is a separate library providing a highly efficient implementation of gradient boosting [113,120]. For the ANN algorithm, the TensorFlow package was employed [121,122]. TensorFlow is an open-source machine learning framework developed by Google, primarily utilized for deep learning tasks.

2.6. Accuracy Assessment

Evaluating the accuracy of results and comparing them provides insights into the model’s and data’s effectiveness [123]. Evaluation is done by comparing classification results with ground truth data [53].

Table 7. Tabular Data Split.

Data	Data Split	Categories	Number of Samples
Single-temporal data	Training Data Set (90%)	Argan	626
		Non-Argan	626
	Testing Data Set (10%)	Argan	69
		Non-Argan	69
Multi-temporal data	Training Data Set (80%)	Argan	15,024
		Non-Argan	15,024
	Testing Data Set (20%)	Argan	1656
		Non-Argan	1656

details the data split used for training and testing. All data originate from the study area but were collected from various locations within it, ensuring diversity and offering a more comprehensive representation of the region. In general, a larger portion of the data was used for training to provide a sufficiently large sample size for effectively training the model. The remaining data were used for testing to evaluate model performance. For fair comparisons between models, it was critical to train them on the same training samples and evaluate them using identical test samples. We used the most commonly employed metrics in similar studies (

Table 8. Formulas of the metrics [108].

Metric Name	Formula
Overall accuracy	${\displaystyle \frac{TP + TN}{TP + FP + TN + FN}}$
Precision (Argan)	${\displaystyle \frac{TP}{TP + FP}}$
Precision (Non-Argan)	${\displaystyle \frac{TN}{TN + FN}}$
Kappa	${\displaystyle \frac{p_0 - p_e}{1 - p_e}}$

TP (True Positive): correctly identified Argan (Argan tree is present, and the model predicts Argan). TN (True Negative): correctly identified Non-Argan (Argan tree is absent, and the model predicts Non-Argan). FP (False Positive): incorrectly identified as Argan (Argan tree is absent, but the model predicts Argan). FN (False Negative): incorrectly identified as Non-Argan (Argan tree is present, but the model predicts Non-Argan). $p_0={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}},p_e={\displaystyle \frac{(TP+FN)\times (TP+FP)+(FP+TN)\times (FN+TN)}{{(TP+TN+FP+FN)}^2}}$.

) [108]: Overall accuracy (OA), precision, and the kappa coefficient. These metrics were key to determining the best-performing model.

The OA is a primary indicator of the likelihood of correct classification, calculated from the sum of the confusion matrix’s diagonal elements (Figure 6). Precision assesses the accuracy of each class individually. The kappa coefficient, a more complex metric, assesses observed accuracy against random chance, considering not just the diagonal but also the off-diagonal elements of the confusion matrix. The “sklearn” package facilitated the evaluation by providing functions for each of the used metrics.

During the final testing phase of argan tree detection, the fully labeled reference image was used to evaluate the performance of the most robust models. The best-performing model was determined based on its ability to correctly classify this reference image, achieving the highest number of accurately classified pixels compared to other classifiers.

2.7. Argan Tree Detection Method

After processing the data, the goal was to detect argan trees in the study area for 2017/2018 and 2022/2023. The approach involved testing various scenarios to minimize mapping errors and improve accuracy. This phase relied on tabular data and included the following steps:

• Model evaluation: machine learning models were tested using original spectral bands and derived spectral indices (single temporal data) to identify the best-performing model for argan tree detection.
• S2 Product Levels: the suitability of Sentinel-2 Level 1C and Level 2A products for training was compared.
• Optimal detection time: the best time of year for detecting argan trees was identified to ensure model accuracy for both periods.
• Impact of spectral information elements: the effect of individual spectral bands and indices on detection accuracy was analyzed by training the best model on each element separately.
• Optimal spectral information combination: the best combination of spectral bands and indices was determined, testing next scenarios:
  • 10-m resolution spectral bands.
  • 10- and 20-m resolution spectral bands.
  • All spectral bands.
  • The top 10 most influential spectral features.
  • All spectral bands and indices.
• Optimal temporal information combination: the impact of temporal data combinations was assessed, including:
  • Single-date data for the detection day.
  • Single-date data from both periods.
  • Multi-temporal data from one year.
  • Multi-temporal data from both periods.
• Final model training and mapping: The final model was used to classify pixels from the 2017/2018 and 2022/2023 images of the study area into two categories: Argan and Non-Argan. This process resulted in the generation of two argan distribution maps.

After evaluating the tests, the best model was trained on the optimal data to ensure robustness. Two optimal Sentinel-2 images were then selected, one for 2017/2018 and another for 2022/2023. Each pixel in these images was classified as “contains Argan” or “does not contain Argan”, resulting in two detailed maps of argan tree distribution for the study area in each period.

2.8. Change Detection Method

To create a deforestation map of argan forests in the study area between 2017/2018 and 2022/2023, four scenarios were considered:

• Detected in both 2017 and 2022.
• Not detected in either year.
• Detected in 2017 but not in 2022 (deforestation).
• Detected in 2022 but not in 2017 (regeneration or new plantations).

The deforestation map was generated by comparing each pixel in the 2017/2018 map to its corresponding location in the 2022/2023 map using a patch-based approach. Due to the 10-m resolution of Sentinel-2, individual argan trees could not be identified, necessitating the use of spatial kernels to improve classification accuracy. Five kernel sizes were evaluated to balance detail preservation and noise reduction. Smaller kernels provide finer details but are prone to increased classification errors. Larger kernels smooth classification results, reducing noise but potentially overlooking small-scale deforestation events [124]. For sparse and heterogeneous forests like argan, medium to large kernels are generally preferred to enhance stability while retaining relevant patterns [125]. The tested kernel sizes were:

• Patch 1 × 1 (Pixel-based): Direct comparison of individual pixels between the two maps.
• Patch 3 × 3: Each pixel in one map was compared to all pixels within a 3 × 3 kernel centered on the corresponding pixel in the other map.
• Patch 5 × 5, 7 × 7, 9 × 9: Similar to the 3 × 3 kernel but using larger sizes.

After generating the deforestation maps, the deforestation rate of argan forests was calculated for each kernel size (Equation (1)).

$$rate = {\displaystyle \frac{Argan\_2017 }{Argan\_17\_22 + Argan\_2017 + Argan\_2022}},$$

(1)

where Argan_17_22 is the number of pixels classified as argan trees in both 2017 and 2022, Argan_2017 is the number of pixels classified as argan trees in 2017 but not in 2022 (deforestation) and Argan_2022 is the number of pixels classified as argan trees in 2022 but not in 2017 (regeneration).

The optimal patch size was determined based on:

• Visual assessment of classification accuracy and deforestation pattern preservation.
• Comparison with deforestation rates from previous studies and official reports.

These methodological choices were crucial in ensuring a reliable estimation of deforestation, despite the limitations of Sentinel-2 imagery and the absence of ground truth data for 2017/2018.

3. Results

3.1. Statistics and Analysis

3.1.1. Optimizing Models and S2 Data Level to Identify the Best Time for Argan Tree Detection

In the initial phase, all selected machine learning models were trained on Sentinel-2 bands (both Levels 1C and 2A) and remote sensing indices (single-temporal data).

Table 9. Results of the machine learning models for detecting argan trees in 2017/2018.

Algorithm	Data Level	Overall Accuracy	Argan Precision	Non-Argan Precision	Kappa
Support Vector Machine	1C	0.964	0.96	0.97	0.928
	2A	0.957	0.97	0.94	0.913
Decision Tree	1C	0.978	0.96	1.00	0.957
	2A	0.978	0.96	1.00	0.957
Random Forest	1C	0.978	0.96	1.00	0.957
	2A	0.978	0.96	1.00	0.957
LightGBM	1C	0.986	0.97	1.00	0.971
	2A	0.986	0.97	1.00	0.971
XGBoost	1C	0.978	0.96	1.00	0.957
	2A	0.971	0.95	1.00	0.942
Gaussian Naive Bayes	1C	0.877	0.91	0.85	0.754
	2A	0.870	0.92	0.83	0.739
K-Nearest Neighbors	1C	0.957	0.93	0.98	0.913
	2A	0.964	0.96	0.97	0.928
Artificial Neural Network	1C	0.964	0.94	0.98	0.928
	2A	0.949	0.94	0.96	0.899

and

Table 10. Results of the machine learning models for detecting Argan trees in 2022/2023.

Algorithm	Data Level	Overall Accuracy	Argan Precision	Non-Argan Precision	Kappa
Support Vector Machine	1C	0.964	0.97	0.96	0.928
	2A	0.971	0.97	0.97	0.942
Decision Tree	1C	0.986	0.97	1.00	0.971
	2A	0.986	0.97	1.00	0.971
Random Forest	1C	0.986	0.97	1.00	0.971
	2A	0.986	0.97	1.00	0.971
LightGBM	1C	0.993	0.99	1.00	0.986
	2A	0.986	0.97	1.00	0.971
XGBoost	1C	0.986	0.97	1.00	0.971
	2A	0.986	0.97	1.00	0.971
Gaussian Naive Bayes	1C	0.884	0.92	0.85	0.768
	2A	0.884	0.91	0.86	0.768
K-Nearest Neighbors	1C	0.971	0.96	0.99	0.942
	2A	0.971	0.97	0.97	0.942
Artificial Neural Network	1C	0.971	0.96	0.99	0.942
	2A	0.964	0.96	0.97	0.928

display the results obtained. The first observation noted is that the OA and kappa rise and fall simultaneously, making it sufficient to rely on one of them. The results clearly demonstrate the superiority of DT, RF, LightGBM, and XGBoost over the other models. LightGBM, in particular, showed outstanding results, with an OA of approximately 98.0% for both levels in 2017/2018 and of 98.0% for Level 2A and 99.3% for Level 1C in 2022/2023. The Gaussian Naive Bayes model yielded the lowest OA. Based on those results, LightGBM will be the primary model used for argan tree detection. Regarding Levels 1C and 2A, the results are very close. Specifically, for LightGBM, the results are equal for 2017/2018, with a slight advantage for Level 1C over Level 2A in 2022/2023. Consequently, Level 1C data will be used in the remaining stages of this study.

Having selected the most suitable machine learning model and data level, the focus now shifts to determining the time that yields the best results. Figure 7 presents the classification outcomes for argan detection, assessed using OA, with the LightGBM model and single-temporal Level 1C data. The results demonstrate notable variability in performance across different dates, highlighting the importance of selecting the optimal time for argan tree detection. For the 2017/2018 period, the best results were achieved in early October, early November, and late February, whereas for 2022/2023, the highest OA was observed in late December. To ensure consistency, the selected detection time for both years had to fall within a period that yielded strong results for each. Based on this criterion, late July was chosen, specifically 24 July 2018 and 18 July 2023.

3.1.2. Impact of Spectral Information Elements

Figure 8 shows the results of argan classification using individual spectral information elements (Sentinel-2 bands and spectral indices) with single temporal data. The results varied across different days, with the best outcome for each element selected (OA between 0.631 and 0.855). While high-resolution Sentinel-2 bands (10 m) generally showed greater importance, all elements provided acceptable results (OA > 60%). Notably, despite B1 having a lower resolution (60 m), it still proved relevant for detection. The lowest performance was observed with spectral band B10, which also explains the minimal differences observed between Level 1C and Level 2A data. These findings suggest that the choice of spectral information is somewhat flexible, as most elements lead to similar detection accuracy, supporting the idea that combining various spectral data may increase the overall model performance.

3.1.3. Optimal Spectral Information Combination

After studying the importance of each spectral information element individually, it is essential to determine the best combination of these elements that yields the best results for argan detection. It should be noted that combination 4 consists of the top 10 pieces of information for argan detection. Based on results from Figure 8, these elements are: B1, B2, B3, B4, B6, B8, ARVI, NDVI, PVI, and SAVI.

Table 11. Performance of multi-spectral data combinations (Test data: tabular).

Data	Life Cycle	Overall Accuracy	Argan Precision	Non-Argan Precision	Kappa
Combination (a). (10-m bands)	2017/2018	0.978	0.96	1.00	0.957
	2022/2023	0.986	0.97	1.00	0.971
Combination (b). (10- and 20-m bands)	2017/2018	0.978	0.96	1.00	0.957
	2022/2023	0.986	0.97	1.00	0.971
Combination (c). (All spectral bands)	2017/2018	0.957	0.94	0.97	0.913
	2022/2023	0.978	0.97	0.99	0.957
Combination (d) (B1, B2, B3, B4, B6, B8, ARVI, NDVI, PVI, and SAVI)	2017/2018	0.971	0.95	1.00	0.942
	2022/2023	0.971	0.96	0.99	0.942
Combination (e). (All spectral bands and indices)	2017/2018	0.986	0.97	1.00	0.971
	2022/2023	0.993	0.99	1.00	0.986

shows the results of all combinations, where combination 5 outperforms the others, confirming the importance of all elements of spectral information. Therefore, we will rely on all Sentinel-2 bands in addition to remote sensing indices.

3.1.4. Optimal Temporal Information Combination

To select the optimal combination of multi-temporal information, the four groups shown in

Table 12. Performance of multi-temporal data combinations (Test data: reference image).

Group	Data	Overall Accuracy	Argan Precision	Non-Argan Precision	Kappa
a	Single-date (2022)	0.823	0.43	0.88	0.759
b	Single-date (2017 and 2022)	0.813	0.47	0.86	0.745
c	Multi-temporal (2022)	0.848	0.65	0.87	0.739
d	Multi-temporal (2017 and 2022)	0.842	0.55	0.88	0.752

were considered and trained separately with the LightGBM algorithm. The models were validated in a real context using the reference image. Unlike all previous results, OA and Kappa do not rise and fall simultaneously, making it insufficient to rely on only one of them. Since the main objective of these models was to detect argan trees, the Argan Precision metric was selected as the deciding factor in choosing which group to use for the remainder of the study. With this in mind, Group C (Table 12) was selected as it showed the highest performance compared to the other groups. Therefore, we will employ two models: the first, trained with 2017/2018 multi-temporal data, will generate an argan tree map for 2017/2018, and the second, trained with 2022/2023 multi-temporal data, will generate a map for the 2022/2023 period.

3.2. Argan Tree Detection Results

Previous experiments (Section 3.1) helped reduce errors and map the argan areas as accurately as possible. Figure 9 shows the resulting maps of the argan zones, one for the year 2017/2018 and another for 2022/2023. Both maps efficiently show the presence of argan trees. The clear differences observed visually confirm that changes have occurred in the study area between the two periods. Figure 10 further validates the quality of the results, showing their strong alignment with ground truth data while also acknowledging the presence of some errors.

3.3. Deforestation Maps of the Argan Forest

Figure 11 illustrates the deforestation map for the 9 × 9 patch. The smaller patches (1 × 1 or 3 × 3) generated deforestation maps with fine details but a significant amount of noise, whereas the larger patches (7 × 7 and 9 × 9) yielded deforestation maps that were more contextually enriched, exhibiting less noise and fewer details. Also,

Table 13. The rate of deforestation in the Admine forest.

Patch	Deforestation Rate
1$\times$1	32.70%
3$\times$3	19.47%
5$\times$5	8.81%
7$\times$7	4.74%
9$\times$9	2.86%

Note: The deforestation rate values include an estimated error of ±0.07%.

presents the estimated deforestation rate for the argan forests in the study area (Admine Forest), showing an initial value of 2.86% ± 0.07%.

4. Discussion

This study aimed to map the Admine argan forest located in the Souss region of Morocco for the periods 2017/2018 and 2022/2023, calculate the deforestation rate over this six-year period, and provide specialists with accurate insights to inform mitigation strategies for this UNESCO World Biosphere Reserve habitat. Using remote sensing and machine learning tools, we addressed the challenge of limited argan-specific data through efficient algorithms capable of analyzing large datasets with minimal computational requirements. Additionally, we sought to enhance the precision of these tools by conducting extensive experiments to identify the optimal scenarios for mapping and deforestation analysis.

The experiments provided valuable insights into the performance of different machine-learning algorithms, data levels, and spectral information. The Decision Tree, Random Forest, and XGBoost machine learning algorithms consistently performed well, in line with prior studies [18], but the LightGBM was selected due to the outstanding results, with OAs greater than 98.0% in all cases. Additionally, the comparison between Sentinel-2 Levels 1C and 2A demonstrated minimal differences, primarily attributed to the negligible contribution of Band 10, the key differentiator between these data levels. These results suggest that both data levels are suitable for this task, offering flexibility in data selection for future studies.

The superior classification results for 2022/2023 compared to 2017/2018 can be attributed to the lack of available ground truth data for 2017/2018. Consequently, the 2022/2023 dataset is inherently more reliable than the 2017/2018 dataset.

Experiments also highlighted the importance of selecting optimal temporal windows for Argan tree detection (Figure 7). Variations in classification results were observed depending on the observation date, underscoring the need for a precise temporal analysis when mapping vegetation in dynamic ecosystems [126,127]. An analysis of classification performance throughout the year, as illustrated in Figure 7, reveals some periods with higher accuracy; however, the difference between the average OA for each season is minimal (

Table 14. The average overall accuracy for each season during 2017/2018 and 2022/2023 periods.

Season	Average OA 2017/2018	Average OA 2022/2023
Autumn (September to November)	0.960	0.976
Winter (December to February)	0.958	0.979
Spring (March to May)	0.957	0.967
Summer (June to August)	0.950	0.975

), suggesting that seasonal variations have a limited impact on classification performance and argan tree detection accuracy. This can be attributed to the severe seven-year drought that Morocco is facing, which has affected the study area [128]. The decreasing rainfall and its seasonal irregularity (Figure 12), as well as the relatively stable temperature throughout the year, which results in limited seasonal fluctuations in the region [129]. On the other hand, the experiments demonstrated that combining spectral and temporal data significantly improves OA (Table 12). Specifically, models trained on time-series data consistently outperformed those trained on single-date data, indicating that temporal information helps the models better learn the characteristics of Argan trees, ultimately leading to improved classification accuracy.

The argan tree is distinguished by its medium-density distribution, adaptation to semi-arid regions, and typical growth in areas surrounded by bare soil. Given these conditions, it was initially expected that incorporating additional spectral information, such as spectral indices, could enhance accuracy—provided an appropriate resampling method was applied. In this study, all Sentinel-2 bands were resampled to a 10-m resolution. While resampling higher-resolution bands does not introduce new information to lower-resolution data, the interpolation method used minimized data distortion, preserved spatial patterns, and enabled the effective use of certain indices (e.g., ARVI and SCI), which combine bands of different native resolutions (10 m, 20 m, and 60 m). Standardizing spatial resolution across all input features benefits machine learning models, as they tend to perform better when all variables share the same resolution, reducing potential biases during training and classification. While resampling inherently has limitations, in this study, it proved advantageous by enhancing model consistency, improving classification performance, and enabling a more effective integration of spectral information.

In this study, OA was assessed using tabular test data and image-based test data, which provide a more realistic representation of real-world complexities. This resulted in a clear gap between the OA achieved with tabular data (100%) and that obtained with image-based data (85%). This discrepancy is primarily due to key classification challenges inherent in image data, including tree edge overlap, which leads to multiple land cover types within a single pixel, the loss of spatial contextual information when data are represented in tabular form, and the impact of mixed pixels, which complicates classification. Nevertheless, achieving perfect accuracy with tabular data does not imply that the task was easy or unreliable; rather, it reflects the quality of data preprocessing and the selection of relevant features. Since the models were trained on tabular data, it is natural for them to perform better on this format than on image-based data. Furthermore, applying the model to an entire scene—characterized by its vastness, class diversity, and factors such as terrain and spatial context—naturally leads to a decrease in accuracy compared to the training phase. This is expected in binary classification, where Argan tree samples were carefully selected to simulate their spectral characteristics.

Pixel-based classification has significant limitations in argan tree detection, leading to errors in deforestation mapping. Factors already discussed, such as the 10-m resolution of Sentinel-2, the presence of mixed pixels, and daily spectral variations, contributed to the classification uncertainties. In an attempt to resolve these problems, a patch-based approach was applied.

Since the study area is located in the plain of Souss argan orchards, which has an average density of 10 trees/ha [6], and that the argan tree may be shrubby or reach up to 10 m, occasionally 20 m [130], we conclude that an argan pixel can be located in a window larger than 3 × 3 pixels (such as a 5, 7, or 9 pixels square). In this case, the most appropriate patch size would be 7 × 7 or even 9 × 9 (close to one hectare) for a sure probability of argan presence. However, determining that optimal patch size (for deforestation detection) was particularly challenging due to the lack of reference data for 2017/2018, which prevented direct numerical validation. Therefore, the two criteria described in Section 2.8 were adopted to evaluate patch performance.

The visual assessment (first criterion) revealed that 1 × 1 and 3 × 3 were the least effective patch sizes. They produced excessive noise and classification errors, making it unreliable. In Figure 13 (Study area—Region 1), which shows a newly constructed road in 2022/2023 that did not exist in 2017/2018, the 1 × 1 and 3 × 3 patches captured this change most clearly. However, in Figure 14 (Study area—Region 2), a zone with no significant changes between 2017/2018 and 2022/2023, the 7 × 7 and 9 × 9 patches best preserved this stability, with 9 × 9 maintaining more relevant spatial details.

The second criterion involved comparing the deforestation rates obtained for each patch size with official reports and previous studies conducted in the same area using Sentinel-2 data and deep learning methodologies, which estimated deforestation rates between 2% and 5% [20,32,131,132]. For instance, a study employing Convolutional Neural Networks (CNN) with 32 × 32 patches estimated a deforestation rate of 2.56% between 2015 and 2020 [20], while another study using U-Net with 16 × 16 and 32 × 32 patches reported a 2.59% deforestation rate between 2015 and 2022 [32]. The 9 × 9 patch aligned most closely with this range, producing results that balanced noise reduction and fine-scale deforestation detection. The 1 × 1 and 3 × 3 patches significantly overestimated deforestation, making them unreliable. Therefore, the 9 × 9 patch size provided the best overall performance, preserving detail while maintaining classification stability (Figure 15). This study highlights the importance of optimizing patch selection strategies to enhance deforestation detection, particularly in regions where ground truth data is limited.

Despite the progress made, several challenges remain. The first challenge arises from the model’s occasional difficulty in distinguishing between argan trees and soil, leading to errors in argan detection. Argan trees, native to arid and semi-arid regions of southwestern Morocco, have small, oval, leathery leaves spaced to conserve water in harsh conditions. The result is less dense foliage than trees from temperate or tropical climates, which tend to have larger and denser leaves. In addition, the gnarled, twisted, and widely spreading branches of the argan contribute to this effect. The low canopy density allows sunlight to reach the ground (Figure 2), which makes the spectral signatures of the argan tree and the ground similar, making it challenging for machine learning models to differentiate between the two. The model’s ability to accurately detect argan trees could be improved by using images with higher spatial resolution than those provided by Sentinel-2, ideally with a resolution of less than one meter, such as drone or airborne data. Alternatively, experimenting with other models, especially deep learning models, may yield better results.

The second challenge lies in the phenomenon of “same spectra but different objects, same objects but different spectra”, which limits classification accuracy when relying on single-date data [21,25]. This study mitigated this issue by employing multi-temporal data, demonstrating the value of temporal information in improving classification results. While challenges remain, our findings provide a solid foundation for further advancements in the detection and management of argan forests.

5. Conclusions

This study demonstrated the feasibility of using Sentinel-2 imagery and machine learning tools for accurate argan forest mapping and deforestation analysis. Extensive experiments were conducted to identify the best machine-learning algorithms, data configurations, and methods for this task, achieving highly accurate results (OA > 85%). This study underscores the critical role of artificial intelligence and technology in offering precise tools and data for decision-makers. By enabling informed and timely interventions, these tools support effective argan forest conservation efforts and help minimize potential ecological impacts. However, further research is needed to address current limitations, such as spectral confusion and the challenges of single-date classification. Future work will focus on improving data quality and exploring new approaches, such as deep learning models and higher-resolution imagery, to enhance the accuracy of both argan tree and deforestation maps.