Skip to content
Frontiers
Search

Original Research

Frontiers in Soil Science ·

Volume 6 - 2026 | https://doi.org/10.3389/fsoil.2026.1864901

Soil texture classification and mapping in semi-arid regions using machine learning and EnMAP hyperspectral data

  • Hassan Mosaid1*corresponding author
  • Ahmed Barakat2
  • Mohamed Bayad1
  • Nora Ezouitine2
  • Soufiane Hajaj2
  • Hassan Ouakhir3
  • Mohammed Hssaisoune4,5,6
  • Khalid El Bahjaouy2
  • Mohamed El garnaoui7
  • Lhoussaine Bouchaou4,6
  • El Houssaine Bouras1
  1. 1Center for Remote Sensing Applications (CRSA), Mohammed VI Polytechnic University (UM6P), Ben Guerir, Morocco
  2. 2Geomatics, Georesources and Environment Laboratory, Faculty of Sciences and Techniques, Sultan Moulay Slimane University, Beni Mellal, Morocco
  3. 3Department of Demography and Environment, High Institute of Nursing Professions and Health Techniques, Beni Mellal, Morocco
  4. 4Applied Geology and Geo-Environment Laboratory, Faculty of Sciences, Ibn Zohr University, Agadir, Morocco
  5. 5Faculty of Applied Sciences, Ibn Zohr University, Ait Melloul, Morocco
  6. 6Mohammed VI Polytechnic University, International Water Research Institute, Ben Guerir, Morocco
  7. 7Center of Urban Systems (CUS), Mohammed VI Polytechnic University (UM6P), Ben Guerir, Morocco
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

Soil texture is a fundamental parameter that governs processes in agriculture, water resources management, environmental conservation, and land use planning. Recent advances in remote sensing technologies, particularly hyperspectral imaging, have significantly enhanced the ability to characterize and map soil texture with high precision. Hyperspectral sensors capture detailed spectral data that reveals subtle variations in soil properties, often undetected by conventional methods. The present study evaluates the potential of EnMAP hyperspectral imagery, combined with machine learning (ML), to predict and map soil textural classes in semiarid environments. Three scenarios were assessed: (i) using hyperspectral data alone, (ii) incorporating spectral indices, and (iii) integrating terrain parameters. Four machine learning algorithms, Random Forest (RF), XGBoost, Support Vector Machines (SVM), and k-Nearest Neighbors (KNN), were evaluated for soil textural classification. Among them, SVM consistently outperformed the others across nearly all scenarios, achieving an overall accuracy (OA) and precision of 75% and 79.2%, respectively, under Scenario 1 (hyperspectral bands only), which is identified as the recommended configuration for operational soil textural class mapping using EnMAP data. RF and XGBoost also demonstrated strong performance, confirming their robustness for soil textural classification and mapping. In contrast, KNN yielded the lowest classification accuracy, particularly when spectral indices were included, suggesting limited effectiveness in this context. The incorporation of terrain features improved classification accuracy for kNN model, highlighting the value of multi-source data integration. The obtained findings demonstrate that EnMAP hyperspectral imagery, with specific ML, offers a powerful tool for precise soil textural classes mapping, critical for sustainable agriculture, erosion control, and water management. This approach can provide policymakers and farmers with valuable insights for making data-driven land-use decisions that promote soil health and food security.

Keywords

1 Introduction

Soil texture is a fundamentally important attribute, as it plays a pivotal role in soil degradation processes and water movement, influencing soil quality and productivity (1). Understanding the spatial variability of soil properties supports more effective environmental and agricultural management and enhances the efficient use of natural resources. However, the qualitative information provided by existing soil maps is often insufficient for precise, site-specific water and fertilizer management (2). Hence, spatial prediction of soil textural class is important for both environmental management and agricultural productivity. In soil science, mapping soil textural class across various spatial scales is efficient for many practical applications. However, obtaining accurate estimates of its spatial variability typically requires the collection and analysis of large numbers of soil samples. Although laboratory-based spectroscopic techniques, such as (Visible and Near-Infrared (Vis-NIR) and mid-infrared (MIR) spectroscopy, have demonstrated high accuracy in predicting soil properties (3), their use is limited to point measurements, making them insufficient for monitoring spatial distribution without extensive sampling campaigns. The complexity and cost of traditional field surveys have prompted researchers to explore alternative indirect methods, particularly those based on reflectance spectroscopy and hyperspectral satellite data (4, 5).

In this context, considerable efforts have been made to use satellite images for estimating soil properties (6–10). However, until recently, most of these attempts relied on data with low spectral resolution. For example, multispectral remote sensing images, such as Sentinel-2 and Landsat-8, are widely used to map soil properties due to their free accessibility and relatively simple data processing requirements (11). Similarly, several studies have highlighted the important role of Sentinel-2 and Landsat satellites in advancing the mapping of soil physical and chemical parameters, providing reliable data for agricultural management (4, 12, 13). However, their broad spectral bands limit the detection of specific absorption features, potentially reducing the accuracy of predictions (2). Hence, hyperspectral satellite imagery is a promising alternative, combining high spectral resolution with broad spatial coverage, making it an effective tool for digital soil mapping. The high spectral resolution of hyperspectral sensors enables the detection of precisely defined absorption features, contributing to the identification of key soil properties (2, 14–16).

To date, several researchers have attempted to exploit the capabilities of hyperspectral satellites for soil texture mapping. For example, Hosseinpour-Zarnaq (17) used hyperspectral PRISMA data to estimate soil texture in both cultivated and non-cultivated lands in Iran. Similarly, Shen (18) developed a new model for surface soil texture using the Advanced Hyperspectral Imager (AHSI) on board the ZiYuan-1-02D satellite (ZY1-02D) to map soil texture in China. Additionally, Dutta (19) demonstrated the feasibility of using AVIRIS-based visible and near-infrared data for analyzing soil texture, providing valuable insights into both the accuracy and the associated uncertainty of estimating soil texture. Gomez (20) used HyMap hyperspectral data in conjunction with Partial Least Squares Regression (PLSR) and found that the studied soil properties were effectively mapped and achieved both high local prediction accuracy and reliable spatial pattern representation. Additionally, Kanning (21) demonstrated that combining AISA hyperspectral data with PLSR provides a cost-effective approach for spatial prediction of soil texture. However, the use of Hyperion data for soil texture mapping remains limited, likely due to its relatively low signal-to-noise ratio, as only a few studies have been conducted in this area (22).

Data from the Environmental Mapping and Analysis Program (EnMAP) satellite represents a major breakthrough in hyperspectral remote sensing, offering enhanced capabilities particularly suited for soil science (23, 24). With its fine spectral resolution and a superior signal-to-noise ratio, EnMAP offers strong potential for soil-related studies, emphasizing its suitability for global-scale mapping and monitoring of soil properties (25). Although several studies have investigated the potential of EnMAP using simulated data (23, 26, 27) and, more recently, real sensor data since its launch in 2022 (16, 28), research specifically focused on soil textural class mapping using EnMAP imagery remains limited. In particular, a noticeable gap in published work assessing the performance of machine learning (ML) algorithms in predicting soil textural class from EnMAP data. ML approaches offer a promising alternative to traditional methods, enabling more efficient and accurate predictions, especially in data-scarce and heterogeneous environments such as semi-arid regions. Leveraging EnMAP hyperspectral imagery in combination with ML algorithms could significantly enhance digital soil mapping (DSM) and support better-informed land management strategies, addressing a critical challenge for DSM in Morocco.

Hence, this study aims to assess the capability of EnMAP-based machine learning models, including Random Forest (RF), Support Vector Machines (SVM), Extreme Gradient Boosting (XGBoost), and k-Nearest Neighbors (kNN), for predicting and mapping soil textural class in the Middle Atlas region of Morocco. Three modeling scenarios are explored to optimize the extraction of spectral features related to soil textural class. Furthermore, model performance is evaluated using statistical metrics, enabling researchers to assess the practical utility of EnMAP for large-scale soil textural class prediction and broader DSM applications.

2 Materials and methods

2.1 Description of the study area

This study was conducted in the upstream part of the Oum Er Rbia River basin, specifically in the southwest of the Middle Atlas Mountains. The study area spans approximately between latitudes 32° and 33° north and longitudes 5° and 6° west (Figure 1). The geomorphological aspect of the region is characterized by diverse terrain, including alluvial terraces, closed depressions, and valleys. Geologically, it consists of a mixture of Cretaceous sub-tabular limestone formations, Liasic-aged dolomitic limestone, Triassic-aged doleritic basalt, quartzite, and red clay. This environment has led to the formation of a diverse group of soil types, including vertisols, calcimagnesic, isohumic, and fersialitic soils. Climatically, the region is classified as a mountainous Mediterranean climate, characterized by cold winters and hot, dry summers. Temperatures range from -5 °C in winter to 40 °C in summer (29). The annual rainfall average is approximately 666 mm, with more than 90% of it falling during the period from December to April. The precipitation pattern exhibits a high erosivity of 1444.48 MJ/mm/hectare-hour-year (30), which significantly contributes to soil erosion and texture degradation. Farmers primarily use the study area’s lands for agriculture and livestock farming, especially raising sheep and goats. However, increasing population pressures and poor land management have resulted in the expansion of agricultural and pasture areas at the expense of forest cover. These changes have contributed to the acceleration of soil degradation, the loss of biodiversity, and the disruption of the hydrological cycle, leading to dam siltation (31, 32).

Localization of the study area showing sampling points.
Figure 1. Localization of the study area showing sampling points.

2.2 Data collection and processing

In this work, we relied on a soil database provided by the Geomatics, Georesources, and Environment laboratory of the Faculty of Sciences and Techniques of Béni Mellal, which contains detailed information on the physical properties of soils, including their textural class in the study area. A total of 88 soil samples, distributed throughout the basin, were collected and analyzed. Soil textural class was determined using Robinson’s method in accordance with AFNOR standard NF X31-107 (AFNOR, 1996). This method is based on the different sedimentation velocities of fine and coarse particles. The analyzed samples represent five textural classes: 19 clay, 17 Silty loam, 17 Sandy, 19 Silty clay, and 16 Sandy loam soils. This data provides a solid basis for building prediction models using EnMAP data and ML techniques.

The hyperspectral data used in this study were acquired by the Environmental Mapping and Analysis Program (EnMAP), a German space mission that was successfully launched in April 2022. The imaging system operates on a push-broom prism principle, enabling simultaneous acquisition of high-resolution spectral data across multiple wavelengths. It covers a broad spectral range, spanning from 420 to 1000 nm in the visible and near-infrared (VNIR) region and from 900 to 2450 nm in the shortwave infrared (SWIR) domain. The mean spectral sampling distance is 6.5 nm for VNIR and 10 nm for SWIR, with a spectral oversampling factor of 1.2 (33), which ensures fine spectral detail and reduces aliasing effects. The EnMAP image used in this study was captured on August 1, 2023. It is an atmospherically corrected image that covers the study area with a resolution of 30 meters (EnMAP L2A products). The 30*30 km tile underwent the pre-removal of bad bands as an important step in hyperspectral imagery processing (34, 35). It contains 242 spectral bands in the visible, near-infrared, and short-wave infrared regions. This richness enables the collection of precise information on the texture and composition of soils. The image data is used to extract the spectral signatures of soils and calculate indices related to their textural class. In this study, only 219 bands were retained after removing bad bands. These retained bands cover a broad spectrum and provide a detailed signature, allowing us to differentiate between soil textural classes. In addition, vegetation in the study area was masked using the Modified Normalized Difference Vegetation Index (mNDVI ≥ 0.2). However, to assess the effect of spectral pre-processing, five input configurations were evaluated: raw surface reflectance, and four pre-processing and transformation techniques, namely continuum-removed, Savitzky-Golay smoothing, and 1st and 2nd derivative spectra.A Scenario-based modeling method for classifying soil textural class.

In this study, three different modeling scenarios were designed to evaluate the impact of various sets of input variables on the accuracy of soil textural class predictions (Figure 2). Each scenario incorporated an increasing set of spectral information and additional data relevant to distinguishing soil textural classes. This approach enabled a comprehensive evaluation of the performance and robustness of the selected machine learning algorithms under varying data conditions, thereby revealing the optimal configuration of input features for reliable soil textural class mapping.

Flowchart of the adopted methodology in this study.
Figure 2. Flowchart of the adopted methodology in this study.

2.2.1 Scenario 1: EnMAP hyperspectral imagery only

In this baseline scenario, only the original spectral bands from the EnMAP hyperspectral image are used as input variables. These bands capture soil surface reflectance in the visible to short-wavelength infrared range, allowing for the detection of differences in surface properties and contributing to the discrimination between different soil textural classes. This scenario serves as a baseline for assessing the predictive power of spectral data alone.

2.2.2 Scenario 2: EnMAP hyperspectral imagery + spectral indices

In this scenario, a set of spectral indices derived from the original bands of the EnMAP image are added to the input data (Figures 2, 3). These indices are designed because they may enhance the ability of ML models to detect soil textural classes.

The controlling factors maps (spectral indices).
Figure 3. The controlling factors maps (spectral indices).

2.2.3 Scenario 3: EnMAP hyperspectral imagery + spectral indices + morphological factors

This scenario combines all input features from the previous two scenarios and integrates additional terrain-related variables (Figure 2). These morphological factors can influence soil formation and distribution, and their inclusion may enhance model performance by providing contextual environmental information. Ultimately, the performance of the three scenarios will be compared to determine the most effective combination for predicting soil textural class in this semi-arid area.

2.3 Spectral indices

The spectral indices used in this study were calculated based on the formulas available in the Index database, developed by Henrich (36). This database offers scientifically validated indices for environmental applications and thus facilitates the precise analysis of soil composition and properties:

2.3.1 Clay index

The Clay index is used to estimate the presence and abundance of clay minerals in the soil. It is calculated by analyzing specific absorption features within the shortwave infrared (SWIR) region, typically over the wavelength range of 2145–2365 nm (37), where clay minerals such as kaolinite, smectite, and illite exhibit characteristic absorption. This index is useful in DSM, enhancing the ability to detect clay content based on surface reflectance characteristics. It is calculated using the following Equation 1:

(1) Clay index=[Band 189]−[Band 205][Band 194]*[Band 194]

Where Band 189 is the reflectance around 2165.47 nm, Band 194 represents the reflectance around 2207.86 nm, and Band 205 is the reflectance around 2298.17 nm.

2.3.2 Kaolinite index

The Kaolinite Index (KI) is a spectral ratio used to detect the presence of kaolinite, a common 1:1 clay mineral, by exploiting its distinctive absorption properties in the short-wavelength infrared (SWIR) range. This index is adjusted based on spectral reflectance values in the wavelength ranges 2235–2365 nm and 2145–2185 nm, which correspond to the characteristic spectral absorption regions of kaolinite. It is calculated according to the following Equation 2:

(2) Kaolinite Index=[Band 205] [Band 189]

Band 205 indicates the reflectance at approximately 2298.17 nm, whereas Band 189 indicates the reflectance at roughly 2165.47 nm.

2.3.3 Ferric oxide index

The Ferric Oxide Index (FOI) is used to detect the presence and distribution of iron oxides, such as hematite and goethite, on the soil surface. These minerals exhibit unique spectral properties, particularly in the visible to near-infrared (VNIR) and short-wavelength infrared (SWIR) ranges. In this study it is calculated using the following Equation 3:

(3) Ferric Oxide Index=[Band 154][Band 68]

Band 154 represents the reflectance around 1653.0 nm, while Band 68 represents the reflectance around 808.905 nm.

2.4 Morphological factors

In addition to spectral data, topographic parameters play an essential role in the formation and variation of soil textural classes. In this sense, we have integrated morphological parameters, namely slope, Topographic Weightiness Index (TWI), Plan Curvature, and Profile Curvature (Figure 4), to evaluate their impact on soil textural class prediction. Figure 2 presents a summary of the methodology steps adopted in this study to predict and map soil textural classes within the study area.

The controlling factors maps (Morphological factors).
Figure 4. The controlling factors maps (Morphological factors).

2.5 Machine learning techniques

After preparation of all the data used, the next step is model training. This step fits the algorithms to identify the relationship between the extracted variables (such as spectral bands, spectral indices, or morphological factors) and soil textural class. The training is performed using four ML models, such as Random Forest (RF) (38), K-Nearest Neighbors (kNN) (39), Support Vector Machine (SVM) (40), and XGBoost (41). Each model is trained separately for each data scenario. All models were trained using 10-fold cross-validation to ensure reliable results and to prevent overfitting. The modelling process is implemented within the R programming language in the RStudio environment. Table 1 summarizes the hyperparameter tuning strategies adopted for each model according to their implementation in the caret and kernlab packages.

Table 1. Hyperparameter tuning strategies for each machine learning model.

ModelPackageTuning strategyParameter(s) searchedSelection criterion
RFcaret (ranger)10-fold CV; grid search over mtry ∈ {5, 10, 20, 40}, splitrule = gini, min.node.size ∈ {1, 5, 10}; 500 treesmtry, min.node.sizeAccuracy
KNNcaret (knn)10-fold CV; grid search over k ∈ {3, 5, 7, …, 31}; features centred and scaledk (neighbors)Accuracy
XGBoostcaret (xgbTree)10-fold CV; tuneLength = 1 (default parameters only)NoneAccuracy
SVMkernlab (ksvm)Fixed C = 2; RBF kernel; σ estimated via sigest(); cross = 10None (C fixed)Internal CV error

2.5.1 Model performance

Model evaluation is a crucial step in quantifying the quality and reliability of predictions, specifically assessing a model’s ability to recognize different classes using independent test data correctly. In the present study, the common 0.3 test ratio is applied to split labeled data and to evaluate the adopted ML models through several specific indicators. However, several classification-specific indicators are employed. Furthermore, precision is calculated per class, indicating the reliability of positive predictions, and calculates the proportion of correctly assigned samples among all those predicted for that class (minimizing false positives). The Overall Accuracy (OA) measures the overall percentage of correctly classified samples by the model. Recall (or sensitivity), also computed per class, quantifies the ML model’s ability to identify all actual instances of a class, representing the proportion of correctly recognized samples compared with the total actual samples in that class (limiting false negatives). Finally, the Receiver Operating Characteristic (ROC) curve graphically illustrates a classifier’s performance by plotting the True Positive Rate against the False Positive Rate. In contrast, the Area under the curve (AUC) provides a single scalar metric summarizing the model’s overall discriminative ability, with higher values indicating better performance.

3 Results

3.1 The spectral reflectance response of different soil textural classes

Figure 5 shows the mean spectral reflectance curves for all soil samples representing different soil textural classes: clay, sandy-loam, sand, silty-clay, and silty-loam. These curves span the spectral range from visible light (420 nm) to the short-wavelength infrared (SWIR) (2430 nm). The spectral reflectance shows clear variations in spectral behavior between different soil textural class types. Coarse-textured soils, namely sandy, silty-loam, and sandy-loam soils exhibit the highest reflectance values across the spectrum, peaking at over 0.3 from 1000 to around 1900 nm. In contrast, fine textured soils (clay and silty-clay) record the lowest reflectance levels, with silty-clay soils ranking lowest in reflectivity across the spectrum. These differences reflect variations in particle size distribution, specific surface area, and water-holding capacity, which strongly influence scattering and absorption processes in soils. Silty soils exhibit an intermediate spectral behavior, consistent with their mixed particle size composition. In addition, pronounced absorption bands near 1500 and 2000 nm are observed for all soil textural classes and are associated with O-H stretching and bending vibrations related to water (soil moisture) and clay minerals (42). Hence, this spectral differentiation highlights the importance of NIR and SWIR wavelengths as key predictors and confirms the effectiveness of hyperspectral remote sensing in distinguishing between soil textural classes.

Mean EnMAP spectral signature for different soil classes.
Figure 5. Mean EnMAP spectral signature for different soil classes.

Table 2 summarizes the descriptive statistics of soil particle-size fractions for the five textural classes. The dataset is reasonably balanced across classes (16–19 samples each), which is an important prerequisite for unbiased ML classification. Sandy soils are characterized by a very high sand content (mean = 89.05± 2.01%) and minimal clay and silt fractions, confirming their coarse texture. Sandy-loamy soils also exhibit a dominant sand fraction (mean = 74.78 ± 2.62%), but with higher silt and clay contents than sandy soils, reflecting their intermediate texture. However, clay soils have the highest clay content, with mean values exceeding 55%, accompanied by relatively low sand and moderate silt fractions. Silty-clay soils exhibit a balanced composition of clay (mean = 42.42 ± 3.29%) and silt (mean = 49.00 ± 4.16%), with a low sand content, consistent with fine-textured soils exhibiting high water retention capacity. Conversely, silty soils are dominated by silt (mean = 71.23 ± 3.45%) and contain comparatively less clay and sand. Although the standard deviations within each textural class are relatively small, indicating limited intra-class variability and overall consistency in textural class classification, variations in the relative proportions of sand, silt, and clay within a given class may locally influence soil optical properties. Such compositional variability can sometimes perturb the spectral signature of soils belonging to the same textural class, particularly in the NIR-SWIR regions where moisture retention and clay mineral content exert strong control on reflectance behavior.

Table 2. Descriptive statistics of soil particle-size fractions.

Texture fractionTexture class
StatisticClay textureSilty-loam textureSandy textureSilty-clay textureSandy-loam texture
Sand (%)Min16.500.6186.100.3070.10
Max29.8018.5993.5013.8078.20
Mean22.7310.8989.058.5874.78
Std.4.264.462.013.402.62
Clay (%)Min48.0012.700.5038.006.20
Max64.7025.006.6053.0014.10
Mean55.9017.884.0142.4210.01
Std.5.323.651.753.292.57
Silt (%)Min11.8063.810.4042.0010.60
Max30.1074.9013.0061.7023.10
Mean21.3771.236.9449.0015.21
Std.5.733.453.154.163.52

3.2 Machine learning model accuracy in differentiating soil textural classes

3.2.1 Scenario 1

Table 3 presents the classification performance of four machine learning models applied to five (5) different soil textural classes under scenario 1 (S1). Across soil textural classes, SVM and XGBoost models demonstrated superior or competitive performance compared to RF and kNN models. For the clay texture, all models achieved high recall values, with kNN, SVM, and XGBoost reaching perfect recall (1.00). In addition, SVM and XGBoost algorithms achieved the best accuracy (0.75), followed by RF (0.71), while kNN showed lower accuracy (0.63). Similar trends were observed for the sandy-loam textural class, where SVM and XGBoost again achieved the highest accuracy values ​​(0.75), followed by RF 0.71, while and kNN obtained lower accuracy of 0.63. For the sandy textural class, SVM and XGBoost maintained higher accuracy (0.75) compared to kNN (0.63). For silty-clay texture, SVM and XGBoost continued to demonstrate the highest accuracy (0.75), while RF and kNN recorded lower accuracies of 0.71 and 0.63, respectively. All models achieved recall values of 0.80, while kNN and SVM reached perfect precision (1.00), indicating reliable discrimination of this class despite its spectral complexity. A similar pattern was observed for the silty-loam textural class, with SVM and XGBoost achieving the highest accuracy values (0.75), outperforming and kNN (0.63). However, RF and SVM achieved perfect precision (1.00), although recall values remained moderate (0.60), suggesting partial misclassification with other textural classes. Thus, for this scenario, SVM and XGBoost provided stable and excellent performance in terms of accuracy across all textural classes using hyperspectral data.

Table 3. Model performance by soil texture class under Scenario 1.

ScenarioTexture classModelPrecisionRecallAccuracy
S1Clay textureRF model0.570.800.71
KNN model0.621.000.63
SVM model0.621.000.75
XGBoost model0.831.000.75
Sandy loam textureRF model0.670.500.71
KNN model0.330.500.63
SVM model0.670.500.75
XGBoost model0.500.500.75
Sandy textureRF model0.670.800.71
KNN model0.670.400.63
SVM model0.670.800.75
XGBoost model0.800.800.75
Silty clay textureRF model0.800.800.71
KNN model1.000.800.63
SVM model1.000.800.75
XGBoost model0.800.800.75
Silty loam textureRF model1.000.600.71
KNN model0.670.400.63
SVM model1.000.600.75
XGBoost model0.750.600.75

3.2.2 Scenario 2

Table 4 displays the classification performance across five (5) soil textural classes under scenario 2. For clay texture, SVM model achieved the highest accuracy (0.71), followed by RF and XGBoost models with an accuracy of 0.63, whereas kNN showed lower accuracy of 0.58. Although kNN reached perfect recall (1.00), its lower accuracy reflects reduced precision and a higher rate of misclassification in this scenario. In the case of sandy-loam texture, SVM again led the performance, achieving a high accuracy of 0.75 followed by RF (0.71). In contrast, both XGBoost and kNN yielded lower accuracy values (0.63 and 0.58, respectively), with kNN showing particularly weak performance in terms of both precision and recall (0.25), highlighting the spectral complexity of this textural class. However, for sandy soil texture, SVM outperformed the other models achieving the highest accuracy (0.75) along with strong precision (0.80) and recall (0.80). RF followed with an accuracy of 0.71, while XGBoost and kNN recorded lower accuracies of 0.63 and 0.58, respectively. Regarding the silty-clay texture, SVM once again achieved the highest accuracy (0.75), followed by RF (0.71). Both kNN and XGBoost showed reduced accuracy (0.58 and 0.63, respectively), despite kNN and SVM reaching perfect precision (1.00). This suggests that while some models can correctly identify silty-clay texture, overall classification consistency remains limited due to spectral overlap. Regarding the silty-loam texture, SVM achieved the highest accuracy (0.75), outperforming RF (0.71), while XGBoost and kNN recorded lower accuracy values of 0.63 and 0.58, respectively. Although RF showed high precision, its moderate recall value (0.60) indicate partial confusion with neighboring textural classes.

Table 4. Model performance by soil texture class under Scenario 2.

ScenarioTexture classModelPrecisionRecallAccuracy
S2Clay textureRF model0.570.800.71
KNN model0.631.000.58
SVM model0.631.000.75
XGBoost model0.600.600.63
Sandy loam textureRF model0.670.500.71
KNN model0.250.250.58
SVM model0.670.500.75
XGBoost model0.670.500.63
Sandy textureRF model0.670.800.71
KNN model0.500.400.58
SVM model0.800.800.75
XGBoost model0.570.800.63
Silty clay textureRF model0.800.800.71
KNN model1.000.800.58
SVM model1.000.800.75
XGBoost model0.600.600.63
Silty loam textureRF model1.000.600.71
KNN model0.500.400.58
SVM model0.750.600.75
XGBoost model0.750.600.63

3.2.3 Scenario 3

Under scenario 3 (Table 5), classification performance across soil texture classes shows more limited discrimination capacity compared to the previous scenarios. For the clay textural class, RF, kNN, and SVM achieved identical accuracy values (0.71), while XGBoost showed lower performance with an accuracy of 0.58. Although kNN reached perfect recall (1.00), its accuracy remained moderate, indicating increased misclassification despite good detection of clay texture. For the sandy-loam textural class, RF, kNN, and SVM again achieved similar accuracy values (0.71), whereas XGBoost exhibited lower accuracy (0.58). Precision and recall values were generally moderate across models, reflecting greater spectral confusion for this intermediate textural class and reduced model separability under this scenario. For the sandy texture, RF, kNN, and SVM achieved comparable accuracy values (0.71), with SVM providing slightly higher precision (0.80) and recall (0.80). XGBoost again showed weaker performance, with an accuracy of 0.58. These results suggest that sandy soils remain relatively easier to classify, although overall performance is constrained under Scenario 3. Regarding the silty-clay texture, RF, kNN, and SVM maintained consistent accuracy levels (0.71), while XGBoost recorded lower accuracy (0.58). Notably, kNN achieved perfect precision (1.00) with high recall (0.80), indicating the rule of adding morphological factors and spectal indices. Finally, for the silty-loam texture, RF, kNN, and SVM again reached identical accuracy values (0.71), whereas XGBoost showed reduced performance (0.58). Although RF and kNN achieved perfect precision (1.00), recall values remained moderate (0.60) compared to previous scenarios, indicating partial confusion with neighboring textural classes. These findings suggest that under Scenario 3, model discrimination capacity is limited, highlighting the sensitivity of soil textural class classification performance to input feature configuration and reinforcing the importance of scenario selection when applying hyperspectral data for soil textural class mapping.

Table 5. Model performance by soil texture class under Scenario 3.

ScenarioTexture classModelPrecisionRecallAccuracy
S3Clay textureRF model0.570.800.71
KNN model0.711.000.71
SVM model0.570.800.71
XGBoost model0.500.600.58
Sandy loam textureRF model0.670.500.71
KNN model0.400.500.71
SVM model0.670.500.71
XGBoost model0.500.250.58
Sandy textureRF model0.670.800.71
KNN model0.600.600.71
SVM model0.800.800.71
XGBoost model0.670.800.58
Silty clay textureRF model0.800.800.71
KNN model1.000.800.71
SVM model0.800.800.71
XGBoost model0.600.600.58
Silty loam textureRF model1.000.600.71
KNN model1.000.600.71
SVM model0.750.600.71
XGBoost model0.600.600.58

These results confirm that SVM and RF maintain strong, balanced performance across all soil textural classes, particularly in silty-clay and silty-loam soils. The kNN model, while adequate in some cases, continues to show limitations in more complex textures, such as silty-loam and sandy-loam soils. However, XGBoost shows competitive precision but occasionally shows reduced accuracy, emphasizing the importance of balancing metrics in model evaluation for hyperspectral soil textural class classification.

3.3 Model performance and evaluation

3.3.1 Relative importance of predictors across prediction classes

The relative importance of the predictors differed across the three strata of the model (Figure 6). In Figure 6a (high importance), the model’s performance was primarily determined by a small subset of spectral predictors, with the most significant contributions coming from the SWIR region (2100–2300 nm). The bands around 2200 nm, in particular band 194 on the EnMAP image which is associated with clay mineral indices, showed the highest values ​​of importance. Furthermore, the kaolinitic index was among the most influential predictors, supporting the interpretation that mineralogical composition, and particularly clay-related characteristics, plays a central role in predicting soil textural classes. The contributions of the visible bands were comparatively limited, suggesting reduced sensitivity to surface color variability or iron oxides. In addition, the medium stratum (Figure 6b) exhibited a more balanced importance distribution across predictors. The prominence of the Clay Index suggests sensitivity to textural variability, although no single predictor dominated the model. In contrast, the lower stratum (Figure 6c) showed a less important pattern that was distributed more evenly over a large number of spectral bands. This stratified importance model highlights the value of combining spectral bands with terrain-derived predictors and spectral indices in environmental modeling. However, in each study, the statistical metrics will confirm the impact of each predictor.

Relative importance of predictors: **(a)** high importance, **(b)** medium importance, and **(c)** low importance.
Figure 6. Relative importance of predictors: (a) high importance, (b) medium importance, and (c) low importance.

3.3.2 Average performance by models and scenarios

This section concerns the overall evaluation of the models used in textural class prediction. The SVM model achieved the highest performance under Scenario 1, with a mean accuracy of 75%, precision of 79.2%, and recall of 74% (Figure 7a). The XGBoost model followed with a competitive mean accuracy, precision and recall (Figure 6a). Additionally, the comparison between raw surface reflectance and pre-processing and transformation techniques confirmed that raw reflectance after bad band removal yields superior classification performance across all models (Supplementary Figure 4), supporting its adoption as the preferred spectral input for soil textural class mapping. In all three scenarios, the SVM model consistently outperformed the other models, achieving the highest levels of accuracy, precision, and recall, especially in the other cases. XGBoost and RF are competitive with robust and balanced performance especially RF. KNN was the least effective model, especially in scenario 2 (Figure 7b), where its performance decreased significantly. In scenario 3 (Figure 7c), performance differed from that observed in scenarios 1 and 2. In particular, the kNN model exhibited high performance, achieving accuracy, precision, and recall values comparable to those of the SVM and RF models. This improvement suggests that, under the conditions defined in scenario 3, the neighborhood-based local learning of the KNN model was well-suited to its structure. In contrast, XGBoost performed inferiorly in scenario 3, with a notable reduction in accuracy, precision, and recall compared to the other models.

Average performance by model and scenario, **(a)** scenario 1; **(b)** scenario 2, **(c)** scenario 3.
Figure 7. Average performance by model and scenario, (a) scenario 1; (b) scenario 2, (c) scenario 3.

3.4 Digital soil textural class mapping

Following the evaluation of statistical metrics of all the models (RF, SVM, kNN, and XGBoost) used in this study, Figure 8 shows the spatial distribution map of soil textural classes within the study area using the SVM model, which achieved the highest performance in previous model comparisons. The map displays five (5) soil textural classes: clay, silty-loam, sandy, silty-clay, and sandy-loam. The clay textural class is predominantly concentrated in the central and southeastern parts of the study area around Ait Ishaq, covering an area of 244.37 km2. The silty-loam textural class occurs throughout the area (195.28 km2) but appears more fragmented, indicating transitional zones between clay and sandy textures. Sandy textural class is widespread, covering the northern and western parts of the study area, where they are often associated with highly eroded and well-drained areas. However, silty-clay textural class appears scattered but is often located near hydrographic networks (257.55 km2), indicating depositional environments with finer sediment accumulation. Sandy-loam soils tend to be concentrated in mid-slope areas (102.86 km2), representing areas with moderate texture mixing due to erosion and sediment deposition. This outcome provides important inputs for land use planning, agricultural development, and improved soil management, especially given the influence of hydrological and human factors on soil physical properties (43, 44). Predictive maps of other scenarios and their ROC-AUC are represented in Supplementary Materials.

SVM predictive map under scenario 1.
Figure 8. SVM predictive map under scenario 1.

To assess the spatial reliability of the SVM predictions, Shannon entropy was computed from the posterior class probability maps (Figure 9). The uncertainty map reveals that prediction confidence is highest in the central clay-dominated areas and in the sandy areas of the northern and western portions of the study area. Higher uncertainty is concentrated along transitional boundaries between textural classes, particularly between silty-clay and clay soils, and between sandy-loam and silty-loam areas. These transitional zones coincide with complex geomorphological settings (mid-slope positions, alluvial terrace margins) where spectral mixing between textural classes is expected.

Shannon entropy uncertainty map for SVM predictions under Scenario 1.
Figure 9. Shannon entropy uncertainty map for SVM predictions under Scenario 1.

4 Discussion

The obtained results demonstrate that hyperspectral imaging unlocks unprecedented precision in soil textural class mapping by capturing subtle spectral signatures imperceptible to conventional methods capabilities, particularly transformative for semi-arid regions like the Middle Atlas of Morocco, where soil heterogeneity challenges traditional surveys. The capability of RF, SVM, and XGBoost methods confirms their ability to decode complex spectral-terrain relationships, while KNN’s struggles underscore the importance of algorithm selection when integrating multi-source geodata. These advancements, validated against global benchmarks (45, 46), align with findings from Mulder (22) and Zhang (47), who highlighted that hyperspectral datasets, when paired with robust ML classifiers, provide consistent improvements in soil property mapping. This improvement establishes a replicable framework for sustainable land management, bridging the gap between hyperspectral RS innovation and on-ground agricultural decision-making.

Classification findings using RF, SVM, and XGBoost models effectively exploit hyperspectral images, achieving the highest accuracy, with precision, recall, and AUC values often exceeding 80% (Figure 10). This demonstrates their strong capability in soil textural class classification based on spectral data. Comparable performance was reported by Vibhute (48), who achieved an average accuracy of 71.18% using a hyperspectral data-based SVM model to extract five various soil types. In addition, Song (49) evaluated multiple models, including PLSR, RF, and XGBoost, and found that XGBoost achieved the highest prediction accuracy for soil texture estimation, followed closely by Random Forest. These results confirm the strong effectiveness of ensemble tree-based methods for soil textural class mapping using UAV-mounted hyperspectral imagery. The RF model performs comparably to SVM and XGBoost, though with slightly lower metrics. Nevertheless, RF remains reliable and stable with the used data, as also noted by Vaudour (50), who emphasized RF’s robustness with heterogeneous soil datasets. SVM is widely recognized for its effectiveness in multidimensional hyperspectral imagery, particularly in LULC and geological monitoring (28, 51, 52). Recent studies have further confirmed SVM’s robustness in handling the curse of dimensionality inherent in hyperspectral datasets, with the radial basis function (RBF) kernel showing particular effectiveness in capturing complex, non-linear spectral relationship (53). However, kNN model performed less well, especially with the integration of spectral indices, highlighting its fundamental limitations in managing high-dimensional and heterogeneous hyperspectral data. Recent studies has documented that kNN may be less effective when data contains many noisy features due to their non-informative influence in calculating distance, with performance particularly compromised in hyperspectral imagery where mixed pixels and spectral noise are prevalent (54, 55). The integration of morphological parameters, such as slope and terrain orientation, further enhances model performance, consistent with the findings of McBratney (56) regarding the influence of terrain on soil reflectance.

ROC AUC measurements for the used models.
Figure 10. ROC AUC measurements for the used models.

In addition, SOM absorbs strongly in the VIS–NIR range and may partially mask textural spectral features, particularly in organically rich vertisols present in the study area. Soil moisture attenuates overall reflectance and shifts absorption feature positions, introducing variability unrelated to textural class. Iron oxide mineralogy, captured by the Ferric Oxide Index, co-varies with soil formation processes rather than directly with textural class, and thus acts as a correlated but indirect predictor. In this study, the use of an August 2023 image (peak dry season in the Middle Atlas) was a deliberate choice to minimize surface moisture effects. While organic matter content was not explicitly controlled for, its spatial variability in these degraded semi-arid soils is expected to be low relative to texture-driven spectral variation.

Moreover, this reliability in detecting soil textural class is particularly valuable in heterogeneous landscapes like the study area, where soil texture varies due to relief, slope, and lithology. As documented by Ben-Dor (57), these environmental factors strongly influence soil formation and surface reflectance. The spatial variability of texture reinforces the necessity of using high-resolution hyperspectral data combined with advanced ML models for precise soil mapping. The discrimination of soil textural classes using EnMAP hyperspectral imagery is fundamentally governed by mineralogical absorption features captured across the VNIR-SWIR spectral range (420–2450 nm). These results align with the findings of Meng (58), who demonstrated importance of SWIR bands in soil texture inversion. In addition, the most influential spectral regions were located in the shortwave infrared domain, particularly around 2000 nm, where Al-OH absorption features associated with clay minerals. These absorption features are well documented as diagnostic indicators of clay mineralogy and soil texture variability (59). In the near-infrared region (900 nm), clear absorption peaks linked to iron oxides further supported discrimination, consistent with previous findings on iron-related absorption features in semiarid soils (60). This spectral behavior supports treating the ferric oxide index as a stable but indirect proxy for pedogenic iron mineralogy rather than a direct textural indicator, reinforcing the interpretation already advanced above. In addition, the rationale for acquiring imagery during the dry season is further reinforced by the physical relationship between soil moisture and spectral band depth: reflectance at 1400 nm and 1900 nm water absorption bands deepens near-linearly with increasing gravimetric moisture content (61), meaning that even modest residual soil water could attenuate the clay feature used for textural discrimination.

The hyperspectral capability of EnMAP provides a clear advantage over multispectral systems for soil mapping. Narrow contiguous bands enable the detection of subtle mineralogical absorption features that are note resolvable with broadband sensors like Sentinel or Landsat. Hyperspectral remote sensing has consistently demonstrated improved prediction accuracy for soil texture and clay content compared to multispectral approaches (62). From an operational standpoint, hyperspectral data offer scalable and repeatable mapping solutions over large areas, reducing dependence on labor-intensive field surveys. While ground sampling remains essential for calibration and validation, satellite based approaches significantly improve spatial coverage and cost efficiency, particularly in hard-to-reach areas. For the future, integrating additional soil attributes (e.g., moisture, chemical composition, organic matter) and climate variables (e.g., temperature, precipitation) could further refine predictions. Moreover, advanced deep learning architectures, particularly convolutional neural networks (CNNs), show promise for capturing complex, non-linear relationships in hyperspectral datasets (63), potentially enhancing accuracy beyond current ML methods.

5 Limitations

Despite the methodological safeguards applied (70/30 stratified split and 10-fold cross-validation), the relatively limited sample size represents a recognized constraint of the present study. In high-dimensional hyperspectral settings, where feature space dimensionality (219 bands) far exceeds sample count, the risk of overfitting cannot be fully excluded. This concerns is consistent with Ng (64) who demonstrated that training samples size directly governs the accuracy and stability of models predicting soil properties from spectral data, and with Wang (65), who reviewed the broader challenge of hyperspectral image classification under small sample conditions and its tendency to inflate apparent performance while limiting generalization. The balanced class distribution across the five textural classes (16–19 samples per class) partially mitigates this risk by preventing class-dominated learning and ensuring that no single texture class disproportionately influences model training. However, the sample set was designed to comprehensively cover all soil textural classes identified within the study area, rather than being restricted to a subset of dominant textures, which strengthens the internal validity of the classification framework for the landscape sampled. In addition, a comparable approach was adopted by Yan (66), who successfully mapped soil organic matter from UAV hyperspectral imagery using a small calibration dataset (20 samples), showing that carefully designed sampling can yield reliable spectral-soil models even when sample numbers are constrained. However, all samples originate from a relatively homogeneous semi-arid ecosystem within the Middle Atlas of Morocco, sharing similar climatic and land management conditions. Consequently, transferability of the trained models to other semi-arid regions would require recalibration with locally collected samples, and is not recommended without such adaptation. It is important to note that this study provides an initial benchmark for the use of EnMAP hyperspectral imagery in soil texture classification. By evaluating a range of commonly used ML algorithms under a consistent framework, the results establish a valuable reference for upcoming investigations by researches. Subsequent studies may build upon this baseline by exploring more advanced statistical, ML, and deep learning approaches to further enhance classification performance and fully exploit the spectral richness of EnMAP data.

6 Conclusion

This research demonstrated that EnMAP hyperspectral data are very successful in predicting soil textural class, since they provide accurate information on soil composition. Combining EnMAP hyperspectral imagery with appropriate machine learning (ML) approaches facilitates the reliable classification of diverse soil textural classes, such as clay, silty loam, sandy, silty clay, and sandy loam, through the analysis of texture-specific spectral responses. Three modeling scenarios were evaluated to enhance prediction accuracy. The first one used only the spectral bands of the EnMAP images. The second incorporated targeted spectral indices designed to capture specific soil properties. The third scenario further integrated topographic variables to account for terrain morphology. The inclusion of EnMAP spectral features consistently improved model accuracy across most algorithms, underscoring their strong predictive relevance. SVM and XGBoost achieved average accuracy above 75%, demonstrating an excellent ability to differentiate textural classes. The RF model also achieved good accuracy and precision, with values of 80.8% and 74.1%, respectively, and demonstrated stable performance across all scenarios. The physically interpretable basis of the classification, rooted in diagnostic absorption features of clay minerals around 2000 nm, iron oxides in the VIS-NIR region, and bound water absorption near 1500 nm, reinforces the scientific robustness of the hyperspectral approach. These findings suggest that although topographic information offers valuable contextual insights, its successful exploitation is contingent upon the model’s ability to manage complex spaces. Based on the comparative evaluation of three modeling scenarios and five ML algorithms, we recommend Scenario 1 (EnMAP spectral bands alone) combined with the SVM model as the most operationally efficient and accurate configuration for soil textural class mapping in semi-arid environments. However, it would also be interesting to test similar strategies in other areas with varying characteristics to see how well they perform. Finally, we could better manage our agricultural and environmental resources if we built automated systems that could swiftly and frequently map soil textural classes.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used Claude AI to assist with language editing and improving clarity of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published paper.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fsoil.2026.1864901/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Hamdi A Zurqani, University of Arkansas at Monticello, United States

Reviewed by: Tarik Mitran, Indian Space Research Organisation, India

Anush Kumar Kasaragod, University of Mississippi, United States

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

The author(s) declared that financial support was received for this work and/or its publication. Research Funding program (OCP nutricrops), University Mohamed VI Polytechnic.

Data availability

The datasets presented in this article are not readily available because The hyperspectral image data used in this study were provided by the German Aerospace Center (DLR) under license. The data are available at https://www.enmap.org/data_access/ (accessed on 03 October 2025). The soil dataset is available from the corresponding author upon reasonable request. Requests to access the datasets should be directed to Hassan.MOSAID-EXT@um6p.ma.

Author contributions

HM: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. AB: Conceptualization, Investigation, Methodology, Supervision, Validation, Writing – review & editing. MB: Conceptualization, Methodology, Validation, Writing – review & editing. NE: Investigation, Writing – review & editing. SH: Methodology, Writing – review & editing. HO: Writing – review & editing. MH: Writing – review & editing. KE: Writing – review & editing. ME: Writing – review & editing. LB: Writing – review & editing. EB: Conceptualization, Funding acquisition, Investigation, Writing – review & editing, Validation.

Acknowledgements

The authors sincerely thank the German Space Agency at the German Aerospace Center (DLR) for providing access to the EnMAP hyperspectral dataset. The authors also gratefully acknowledge the financial support provided by the OCP nutricrops.

References

  1. 1 Hillel D. Soil physics explores the hidden turmoil in the field. In: Hillel D, editor. Applications of Soil Physics. New York: Academic Press (1980). p. 1 – 2.
  2. 2 Castaldi F, Palombo A, Santini F, Pascucci S, Pignatti S, Casa R. Evaluation of the potential of the current and forthcoming multispectral and hyperspectral imagers to estimate soil texture and organic carbon. Remote Sens Environ. (2016) 179: 54 – 65. doi: doi:10.1016/j.rse.2016.03.025
  3. 3 Ben-Dor E, Chabrillat S, Demattê JAM, Taylor GR, Hill J, Whiting ML. Using imaging spectroscopy to study soil properties. Remote Sens Environ. (2009) 113: S38–55. doi: doi:10.1016/j.rse.2008.09.019
  4. 4 Gholizadeh A, Žižala D, Saberioon M, Borůvka L. Soil organic carbon and texture retrieving and mapping using proximal, airborne and Sentinel-2 spectral imaging. Remote Sens Environ. (2018) 218: 89 – 103. doi: doi:10.1016/j.rse.2018.09.015
  5. 5 Casa R, Castaldi F, Pascucci S, Palombo A, Pignatti S. A comparison of sensor resolution and calibration strategies for soil texture estimation from hyperspectral remote sensing. Geoderma. (2013) 197-198: 17 – 26. doi: doi:10.1016/j.geoderma.2012.12.016
  6. 6 Silvero NEQ, Demattê JAM, Vieira JDS, Mello FADO, Amorim MTA, Poppiel RR. Soil property maps with satellite images at multiple scales and its impact on management and classification. Geoderma. (2021) 397. doi: doi:10.1016/j.geoderma.2021.115089
  7. 7 Gallo BC, Demattê JAM, Rizzo R, Safanelli JL, Mendes WDS, Lepsch IF. Multi-temporal satellite images on topsoil attribute quantification and the relationship with soil classes and geology. Remote Sens. (2018) 10. doi: doi:10.3390/rs10101571
  8. 8 Ghazali MF, Wikantika K, Harto AB, Kondoh A. Generating soil salinity, soil moisture, soil pH from satellite imagery and its analysis. Inf Process Agric. (2020) 7: 294 – 306. doi: doi:10.1016/j.inpa.2019.08.003
  9. 9 Bayad M, Gerard B, Chehbouni A, Hawkesford MJ, Chau HW, Bouray M. Spatial distribution and drivers of soil organic carbon content in croplands of Morocco. Pedosphere. (2025). doi: doi:10.1016/j.pedsph.2025.04.007
  10. 10 Mosaid H, Barakat A, John K, Faouzi E, Bustillo V, El Garnaoui M. Improved soil carbon stock spatial prediction in a Mediterranean soil erosion site through robust machine learning techniques. Environ Monit Assess. (2024) 196: 130. doi: doi:10.1007/s10661-024-12294-x
  11. 11 Castaldi F. Sentinel-2 and Landsat-8 multi-temporal series to estimate topsoil properties on croplands. Remote Sens. (2021) 13. doi: doi:10.3390/rs13173345
  12. 12 El Bahjaouy K, Barakat A, Oussilkane A, Hilali A, Amrani N, Mosaid H. Spatial mapping of soil salinity in a semiarid region using a machine learning model based on spectral indices and ground data. Model Earth Syst Environ. (2025) 11. doi: doi:10.1007/s40808-025-02444-6
  13. 13 Gasmi A, Gomez C, Lagacherie P, Zouari H, Laamrani A, Chehbouni A. Mean spectral reflectance from bare soil pixels along a Landsat-TM time series to increase both the prediction accuracy of soil clay content and mapping coverage. Geoderma. (2021) 388. doi: doi:10.1016/j.geoderma.2020.114864
  14. 14 Jia S, Li H, Wu X, Li Q. Laboratory-based hyperspectral image analysis for the classification of soil texture. J Appl Remote Sens. (2019) 13. doi: doi:10.1117/1.JRS.13.046508
  15. 15 Misbah K, Laamrani A, Casa R, Voroney P, Dhiba D, Ezzahar J. Spatial prediction of soil attributes from PRISMA hyperspectral imagery using wrapper feature selection and ensemble modeling. PFG- J Photogramm Remote Sens Geoinf Sci. (2024) 93: 197 – 215. doi: doi:10.1007/s41064-024-00323-w
  16. 16 Bouslihim Y, Bouasria A. Potential of EnMAP hyperspectral imagery for regional-scale soil organic matter mapping. Remote Sens. (2025) 17. doi: doi:10.3390/rs17091600
  17. 17 Hosseinpour-Zarnaq M, Omid M, Sarmadian F, Ghasemi-Mobtaker H, Alimardani R, Bohlol P. Exploring the capabilities of hyperspectral remote sensing for soil texture evaluation. Ecol Inf. (2025) 90. doi: doi:10.1016/j.ecoinf.2025.103336
  18. 18 Shen Q, Shang K, Xiao C, Tang H, Wu T, Wang C. A novel hyperspectral remote sensing estimation model for surface soil texture using AHSI/ZY1-02D satellite image. Int J Appl Earth Obs Geoinf. (2025) 138. doi: doi:10.1016/j.jag.2025.104453
  19. 19 Dutta D, Goodwell AE, Kumar P, Garvey JE, Darmody RG, Berretta DP. On the feasibility of characterizing soil properties from AVIRIS data. IEEE Trans Geosci Remote Sens. (2015) 53: 5133–47. doi: doi:10.1109/TGRS.2015.2417547
  20. 20 Gomez C, Lagacherie P, Coulouma G. Regional predictions of eight common soil properties and their spatial structures from hyperspectral Vis–NIR data. Geoderma. (2012) 189-190: 176–85. doi: doi:10.1016/j.geoderma.2012.05.023
  21. 21 Kanning M, Siegmann B, Jarmer T. Regionalization of uncovered agricultural soils based on organic carbon and soil texture estimations. Remote Sens. (2016) 8. doi: doi:10.3390/rs8110927
  22. 22 Mulder VL, de Bruin S, Schaepman ME, Mayr TR. The use of remote sensing in soil and terrain mapping — a review. Geoderma. (2011) 162: 1 – 19. doi: doi:10.1016/j.geoderma.2010.12.018
  23. 23 Steinberg A, Chabrillat S, Stevens A, Segl K, Foerster S. Prediction of common surface soil properties based on Vis-NIR airborne and simulated EnMAP imaging spectroscopy data: Prediction accuracy and influence of spatial resolution. Remote Sens. (2016) 8. doi: doi:10.3390/rs8070613
  24. 24 Guanter L, Kaufmann H, Segl K, Foerster S, Rogass C, Chabrillat S. The EnMAP spaceborne imaging spectroscopy mission for Earth observation. Remote Sens. (2015) 7: 8830–57. doi: doi:10.3390/rs70708830
  25. 25 Milewski R, Chabrillat S, Tsakiridis NL, Tziolas N, Schmid T. (2024). “ Evaluation of EnMAP imagery for accurate topsoil estimation in Mediterranean agricultural regions ”, in: IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, Piscataway, NJ: IEEE 7 – 12.
  26. 26 Chabrillat S, Foerster S, Steinberg A, Segl K. (2014). “ Prediction of common surface soil properties using airborne and simulated EnMAP hyperspectral images: Impact of soil algorithm and sensor characteristic ”, in: 2014 IEEE Geoscience and Remote Sensing Symposium, Piscataway, NJ: IEEE 13–8.
  27. 27 Ward KJ, Chabrillat S, Brell M, Castaldi F, Spengler D, Foerster S. Mapping soil organic carbon for airborne and simulated EnMAP imagery using the LUCAS soil database and a local PLSR. Remote Sens. (2020) 12. doi: doi:10.3390/rs12203451
  28. 28 Hajaj S, El Harti A, Pour AB, Khandouch Y, Üstüner M, Amiri MM. Balancing hyperspectral dimensionality reduction and information preservation for machine learning-based lithological classification using EnMAP hyperspectral imagery. Remote Sens Applications: Soc Environ. (2025) 38. doi: doi:10.1016/j.rsase.2025.101618
  29. 29 Faouzi E, Arioua A, Hssaisoune M, Boudhar A, Elaloui A, Karaoui I. Sensitivity analysis of CN using SCS-CN approach, rain gauges and TRMM satellite data assessment into HEC-HMS hydrological model in the upper basin of Oum Er Rbia, Morocco. Model Earth Syst Environ. (2022) 8: 4707–29. doi: doi:10.1007/s40808-022-01404-8
  30. 30 Mosaid H, Barakat A, Bustillo V, Rais J. Modeling and mapping of soil water erosion risks in the Srou Basin (Middle Atlas, Morocco) using the EPM model, GIS and magnetic susceptibility. J Landscape Ecol. (2022) 15: 126–47. doi: doi:10.2478/jlecol-2022-0007
  31. 31 Eloudi H, Reddad H, Hssaisoune M, Estrany J, Krimissa S, Elaloui A. Assessing the performance of MCDM, statistical and machine learning ensemble models for gully sensitivity mapping in a semi-arid context. Geocarto Int. (2022) 37: 17435–64. doi: doi:10.1080/10106049.2022.2129818
  32. 32 Mosaid H, Barakat A, Bouras EH, Ismaili M, El Garnaoui M, Abdelrahman K. Dam siltation in the Mediterranean region under climate change: A case study of Ahmed El Hansali Dam, Morocco. Water. (2024) 16: 3108. doi: doi:10.3390/w16213108
  33. 33 Chabrillat S, Foerster S, Segl K, Beamish A, Brell M, Asadzadeh S. The EnMAP spaceborne imaging spectroscopy mission: Initial scientific results two years after launch. Remote Sens Environ. (2024) 315. doi: doi:10.1016/j.rse.2024.114379
  34. 34 Pal M, Rasmussen T, Porwal A. Optimized lithological mapping from multispectral and hyperspectral remote sensing images using fused multi-classifiers. Remote Sens. (2020) 12. doi: doi:10.3390/rs12010177
  35. 35 Ji L, Wang L, Geng X. An automatic bad band pre-removal method for hyperspectral imagery. IEEE J Sel Top Appl Earth Obs Remote Sens. (2019) 12: 4985–94. doi: doi:10.1109/JSTARS.2019.2944930
  36. 36 Henrich C, Krauss V, Götze G, Sandow C. (2012). “ IDB-index-database; development of a database for remote sensing indices ”, in: Crop Science Colloquium, Univ of Bonn.
  37. 37 Henrich V, Krauss G, Götze C, Sandow C. IDB - index-database: development of a database for remote sensing indices. Crop Sci Colloquium. (2011).
  38. 38 Breiman L. Random forests. Mach Learn. (2001) 45: 5 – 32. doi: doi:10.1023/a:1010933404324
  39. 39 Fix E, Hodges JL. Discriminatory analysis. Nonparametric discrimination: Consistency properties. Int Stat Rev / Rev Internationale Statistique. (1989) 57. doi: doi:10.2307/1403797
  40. 40 Sain SR. The nature of statistical learning theory. Technometrics. (1996) 38: 409. doi: doi:10.1080/00401706.1996.10484565
  41. 41 Chen T, Guestrin C. (2016). “ XGBoost: A scalable tree boosting system ”, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, CA: Association for Computing Machinery), 785–94.
  42. 42 Viscarra Rossel RA, Walvoort DJJ, McBratney AB, Janik LJ, Skjemstad JO. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma. (2006) 131: 59 – 75. doi: doi:10.1016/j.geoderma.2005.03.007
  43. 43 Barakat A, Rafai M, Mosaid H, Islam MS, Saeed S. Mapping of water-induced soil erosion using machine learning models: A case study of Oum Er Rbia Basin (Morocco). Earth Syst Environ. (2022) 7: 151–70. doi: doi:10.1007/s41748-022-00317-x
  44. 44 Ouakhir H, Ennaji N, Goumih M, Ghadbane O, Chakir M, El Ghachi M. Assessment of suspended sediment transport and associated sediment patterns in the Mkhdach catchment (Middle Atlas Mountainous/Morocco). Model Earth Syst Environ. (2025) 11. doi: doi:10.1007/s40808-025-02475-z
  45. 45 Khosravi I. Towards sustainable agriculture in Iran using a machine learning-driven crop mapping framework. Eur J Remote Sens. (2025) 58. doi: doi:10.1080/22797254.2025.2490787
  46. 46 Rosero-Vlasova OA, Vlassova L, Pérez-Cabello F, Montorio R, Nadal-Romero E. Modeling soil organic matter and texture from satellite data in areas affected by wildfires and cropland abandonment in Aragón, Northern Spain. J Appl Remote Sens. (2018) 12: 42803. doi: doi:10.1117/1.JRS.12.042803
  47. 47 Zhang W, Gu X, Tang L, Yin Y, Liu D, Zhang Y. Application of machine learning, deep learning and optimization algorithms in geoengineering and geoscience: Comprehensive review and future challenge. Gondwana Res. (2022) 109: 1 – 17. doi: doi:10.1016/j.gr.2022.03.015
  48. 48 Vibhute AD, Kale KV, Dhumal RK, Mehrotra SC. (2015). “ Soil type classification and mapping using hyperspectral remote sensing data ”, in: 2015 International Conference on Man and Machine Interfacing (MAMI), Piscataway, NJ: IEEE 1 – 4.
  49. 49 Song Q, Gao X, Song Y, Li Q, Chen Z, Li R. Estimation and mapping of soil texture content based on unmanned aerial vehicle hyperspectral imaging. Sci Rep. (2023) 13: 14097. doi: doi:10.1038/s41598-023-40384-2
  50. 50 Vaudour E, Gomez C, Fouad Y, Lagacherie P. Sentinel-2 image capacities to predict common topsoil properties of temperate and Mediterranean agroecosystems. Remote Sens Environ. (2019) 223: 21 – 33. doi: doi:10.1016/j.rse.2019.01.006
  51. 51 Nhaila H, Elmaizi A, Sarhrouni E, Hammouch A. Supervised classification methods applied to airborne hyperspectral images: comparative study using mutual information. Proc Comput Sci. (2019) 148: 97 – 106. doi: doi:10.1016/j.procs.2019.01.013
  52. 52 Ustuner M. (2024). “ Randomized principal component analysis for hyperspectral image classification ”, in: 2024 IEEE Mediterranean and Middle-East Geoscience and Remote Sensing Symposium (M2GARSS). Piscataway, NJ: IEEE.
  53. 53 Pathak DK, Kalita SK, Bhattacharya DK. Hyperspectral image classification using support vector machine: a spectral spatial feature based approach. Evol Intell. (2021) 15: 1809–23. doi: doi:10.1007/s12065-021-00591-0
  54. 54 Lu J, Gweon H. Random k conditional nearest neighbor for high-dimensional data. PeerJ Comput Sci. (2025) 11: e2497. doi: doi:10.7717/peerj-cs.2497
  55. 55 Lin D, Yang Z, Wang M, Cheng Y, Pan Q. (2020). “ Collaborative-representation-based nearest neighbor classifier for hyperspectral image classification combined with superpixel and loopy belief propagation ”, in: Advances in Brain Inspired Cognitive Systems (Cham: Springer International Publishing).
  56. 56 McBratney AB, Mendonça Santos ML, Minasny B. On digital soil mapping. Geoderma. (2003) 117: 3 – 52. doi: doi:10.1016/S0016-7061(03)00223-4
  57. 57 Ben-Dor E, Chabrillat S, Demattê JAM. Characterization of soil properties using reflectance spectroscopy, Boca Raton, FL: CRC Press 2nd Edition. (2018). pp. 187 – 247. doi:. doi:10.1201/9781315164151-8
  58. 58 Meng J, Chu N, Luo C, Liu H, Li X. High-resolution mapping of topsoil sand content in planosol regions using temporal and spectral feature optimization. Remote Sens. (2025) 17. doi: doi:10.3390/rs17030553
  59. 59 Ben-Dor E, Banin A. Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties. Soil Sci Soc Am J. (1995) 59: 364–72. doi: doi:10.2136/sssaj1995.03615995005900020014x
  60. 60 Qi Y, Gan S, Yuan X, Hu L, Hu J, Zhao H. Research on the quantitative inversion of soil iron oxide content using hyperspectral remote sensing and machine learning algorithms in the lufeng annular structural area of Yunnan, China. Sensors (Basel). (2024) 24. doi: doi:10.3390/s24217039
  61. 61 McGuirk SL, Cairns IH. Relationships between soil moisture and visible–NIR soil reflectance: A review presenting new analyses and data to fill the gaps. Geotechnics. (2024) 4: 78 – 108. doi: doi:10.3390/geotechnics4010005
  62. 62 Dutta A, Lall B, Sharma S. Potential of satellite hyperspectral imaging technology in soil health analysis: A step towards environmental sustainability. Environ Monit Assess. (2025) 197: 314. doi: doi:10.1007/s10661-025-13728-w
  63. 63 Zhang H, Li Y, Zhang Y, Shen Q. Spectral-spatial classification of hyperspectral imagery using a dual-channel convolutional neural network. Remote Sens Lett. (2017) 8: 438–47. doi: doi:10.1080/2150704X.2017.1280200
  64. 64 Ng W, Minasny B, Mendes WDS, Demattê JAM. The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data. SOIL. (2020) 6: 565–78. doi: doi:10.5194/soil-6-565-2020
  65. 65 Wang X, Liu J, Chi W, Wang W, Ni Y. Advances in hyperspectral image classification methods with small samples: A review. Remote Sens. (2023) 15. doi: doi:10.3390/rs15153795
  66. 66 Yan Y, Yang J, Li B, Qin C, Ji W, Xu Y. High-resolution mapping of soil organic matter at the field scale using UAV hyperspectral images with a small calibration dataset. Remote Sens. (2023) 15. doi: doi:10.3390/rs15051433

Citation

Mosaid, H., et al. (2026). Soil texture classification and mapping in semi-arid regions using machine learning and EnMAP hyperspectral data. Frontiers in Soil Science, Volume 6. https://doi.org/
Received
Accepted
Published
Volume
6 - 2026

Copyright

Copyright © 2026 Mosaid, Barakat, Bayad, Ezouitine, Hajaj, Ouakhir, Hssaisoune, El Bahjaouy, El garnaoui, Bouchaou and Bouras.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Hassan Mosaid

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

Original ResearchSiderophores and siderophore-associated metabolites from Antarctic Pseudomonas sp. ANT_H12B contribute to iron and phosphorus mobilization in soils with contrasting propertiesXiaomeng Chen, Marcin Musiałowski, Klaudia Dębiec-Andrzejewska, Robert Stasiuk, Luisella Celi, Michela Schiavon, Maria Martin, Silvia CellettiPerspectiveSoil contaminants at the environment-health interface require function-centered risk assessmentSylvester Chibueze Izah, Ebinyo Rebecca Aseibai, Vivian Ibienebakabobo Promise, Matthew Chidozie OgwuOriginal ResearchSoil organic carbon storage across Mediterranean mountain forests: a comparative assessment of forest ecosystem diversity and geological substrate in Ifrane and Talassemtane National Parks, MoroccoMohamed El Mderssa, Meysara ELMALKI, Fouad Malki, Taha Abdelhakim, Younes Abbas, Jamal Ibijbijen