Soil erosion creates substantial environmental and economic challenges, especially in areas vulnerable to land degradation. This study investigates the use of machine learning (ML) techniques—namely Support Vector Machines (SVM) and Generalized Linear Models (GLM)—for geospatial modeling of soil erosion susceptibility (SES). By leveraging geospatial data and incorporating a range of factors including hydrological, topographical, and environmental variables, the research aims to improve the accuracy and reliability of SES predictions. Results show that the SVM model predominantly identifies areas as having moderate (40.59%) or low (38.50%) susceptibility, whereas the GLM model allocates a higher proportion to very low (24.55%) and low (38.59%) susceptibility. Both models exhibit high performance, with SVM and GLM achieving accuracies of 87.4% and 87.2%, respectively, though GLM slightly surpasses AUC (0.939 vs. 0.916). GLM places greater emphasis on hydrological factors such as distance to rivers and drainage density, while SVM provides a more balanced assessment across various variables. This study demonstrates that ML-based models can significantly enhance SES assessments, offering a more nuanced and accurate approach than traditional methods. The findings highlight the value of adopting innovative, data-driven techniques in environmental modeling and offer practical insights for land management and conservation practices.
Research Article
Transformation of Geospatial Modelling of Soil Erosion Susceptibility Using Machine Learning
https://doi.org/10.21203/rs.3.rs-4933265/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
Soil Erosion Susceptibility (SES)
Geospatial Modelling
Machine Learning (ML)
Support Vector Machines (SVM)
Generalized Linear Models (GLM)
Soil erosion remains a critical environmental challenge that impacts ecosystems, agriculture, and infrastructure (Olii et al., 2023). The degradation of fertile topsoil, sedimentation of waterways, and loss of vegetation cover are just some of the detrimental effects caused by erosion, which can lead to long-term ecological damage (Arabameri et al., 2019) and economic losses (Almouctar et al., 2021). Predicting and managing soil erosion susceptibility (SES) is therefore essential for sustainable land use and environmental conservation (Kucuker & Cedano Giraldo, 2022). To achieve this, accurate and reliable models that can predict the susceptibility of different areas to erosion are needed. Traditionally, modeling SES has relied on empirical methods that use historical data and simple statistical relationships to predict future erosion patterns (Saini et al., 2015). However, these methods often fail to capture the complexity of the interactions between the many environmental variables that influence erosion processes, such as rainfall intensity, soil type, land use, and topography (Olii, Olii, et al., 2024). As a result, there is a growing need for more sophisticated modeling approaches that can better account for these complexities and provide more accurate predictions (Golijanin et al., 2022; Kucuker & Cedano Giraldo, 2022).
Geospatial modeling, when combined with machine learning techniques, significantly enhances the capability to predict and analyze SES. Machine learning (ML) models such as Random Forest (RF), Decision Tree (DT), Artificial Neural Networks (ANN), Support Vector Machines (SVM), Generalized Linear Models (GLM), etc can handle complex, non-linear relationships between environmental factors and SES, which traditional geospatial models might not fully capture (Al-Bawi et al., 2021; Gayen et al., 2019). By integrating these models with GIS and remote sensing data, the spatial patterns of SES can be more accurately mapped and predicted (Olii, Olii, et al., 2024). This combination allows for a more data-driven approach, where the models can learn from large datasets, adjust to various geographical contexts, and improve prediction accuracy by using classified and weighted factors tailored to local environmental conditions. The synergy between geospatial modeling and ML offers powerful tools for more effective land management and soil erosion prevention strategies.
Most studies utilizing ML for SES modeling rely on raw or normalized continuous data, without prior classification into discrete classes or the assignment of weights based on expert judgment (Golkarian et al., 2023; Huang et al., 2023; Phinzi & Szabó, 2024). This approach can lead to less interpretable models, as the continuous nature of the data might obscure important distinctions between different categories of environmental factors. Additionally, the lack of expert-informed weights may result in the model underestimating or overestimating the significance of certain variables, potentially compromising the accuracy and robustness of predictions. This limitation highlights the need for more sophisticated methods that incorporate domain expertise into the ML modeling process to improve the reliability and practical applicability of SES assessments. This leaves a gap in understanding how the integration of this traditional approach with advanced ML models could enhance prediction accuracy and model interpretability. This study introduces a novel approach by integrating traditional classification and weighting of environmental factors with advanced ML models like SVM and GLM for SES mapping. Unlike previous studies that often apply these models individually, this research emphasizes the innovative combination of pre-classifying factors into discrete classes and assigning weights based on expert knowledge. This methodology not only enhances model interpretability, making the results more accessible to practitioners and decision-makers, but also addresses common machine learning challenges such as data complexity, overfitting, and multicollinearity. Furthermore, the adaptable classification system allows the model to be customized to various geographic settings, increasing its applicability and robustness. This study offers a significant advancement over traditional modeling approaches by improving prediction accuracy and model stability, particularly in regions with unique environmental conditions.
This study aims to address this research gap by exploring the potential for integrating GLM and SVM in the geospatial modeling of SES. The research will involve a systematic comparison of these models to evaluate their effectiveness in predicting SES across different spatial scales and environmental conditions based on prior classification into discrete classes. Additionally, the study will investigate how these models can be optimized to handle the complexities of spatial data, including the influence of diverse and non-linear environmental variables such as climate, topography, and land use. By integrating GLM and SVM, this research seeks to develop a more robust and comprehensive framework for predicting erosion susceptibility, which could enhance the accuracy and reliability of SES assessments. The findings of this study have the potential to contribute significantly to the field of environmental modeling, offering new insights into the strengths and limitations of GLM and SVM and providing a basis for future research on the integration of statistical and ML approaches in geospatial modeling.
2.1 Study Area
The Saddang Watershed is located in the southwestern part of Sulawesi Island, within coordinates ranging from 2° 43' 42.4992" S to 3° 34' 51.4992" S latitude and 119° 14' 49.4988" E to 120° 3' 43.4988" E longitude. Covering an area of 4,909 km², as shown in Fig. 1, this watershed spans across the South Sulawesi and West Sulawesi provinces. The primary river within the watershed is the Saddang River, which traverses the Enrekang, Tana Toraja, and North Toraja districts in South Sulawesi, and extends into Polewali in West Sulawesi. The river flows into the Makassar Strait through two estuaries: Barbana and Paria. The Saddang Watershed is crucial for both irrigation and energy production. The Benteng Dam provides irrigation for over 94,000 hectares of agricultural land, while the Bakaru Hydroelectric Power Plant, located downstream, has a capacity of 128 MW and plays a key role in meeting the region's energy needs. Additionally, the watershed has significant groundwater potential, estimated at around 1.354 million m3/year, which further supports local water supply and agriculture. The region's geomorphology features a variety of landforms, including fallen deposits, mountain/hill volcanoes, strongly incised folded mountains and hills, and karst hills. Land use within the watershed is varied, comprising settlements, rice fields, plantations, grasslands, swamps, water bodies, and areas designated for mixed dryland agriculture and forests. This mix of land uses supports a wide range of economic activities, including agriculture and forestry, which are vital to the local communities. The topography of the study area is notably diverse, with elevations ranging from 44 m to 2,880 m and an average elevation of 1,277 m. The climate in the Saddang Watershed is characterized by an average annual temperature of around 23°C. October is the warmest month with an average temperature of 26°C, while June is the coolest, averaging 22°C. Rainfall is substantial, with an annual average of 2,500 mm. The month of May receives the highest average rainfall of 387 mm, whereas September receives the least at 68 mm.
2.2 Overview of methodological framework
The methodological framework of this study involves several key steps to assess SES (Fig. 2). It begins with data collection, followed by the mapping of soil erosion inventory. Key factors influencing soil erosion are then selected for analysis. ML models are used to predict SES areas, and their performance is rigorously evaluated. The results are then normalized to ensure consistency and comparability. This structured approach ensures a thorough and accurate assessment of SES.
2.2.1 Collection Data
The data utilized in this study includes a diverse range of geospatial and environmental datasets, each sourced from specific platforms and offering distinct resolutions. The SRTM data, with a resolution of 30 x 30 m2, provides detailed elevation information crucial for topographic analysis and is accessible through the USGS Earth Explorer website (https://earthexplorer.usgs.gov/). Similarly, Landsat 9 OLI/TIRS imagery, also with a 30 x 30 m2 resolution, offers high-quality satellite imagery useful for quantifying vegetation greenness, available from the same USGS platform (https://earthexplorer.usgs.gov/). For soil characteristics, the SoilGrids website (https://soilgrids.org/) supplies maps with a 250 x 250 m2 resolution detailing soil texture, organic carbon content, and bulk density, essential for understanding soil properties and their implications on land use and agriculture. Rainfall data, with a finer resolution of 0.25° x 0.25°, is provided by the NASA POWER Data Access Viewer (https://power.larc.nasa.gov/data-access-viewer/), which helps in analyzing precipitation patterns and their impacts. SAS Planet and Google Earth Images are highly valuable for identifying and monitoring soil erosion. The detailed imagery enables precise visualization of land surface changes, allowing researchers to detect soil erosion patterns. By providing current and historical views of the landscape, these images support a comprehensive analysis of soil erosion processes. Lastly, administrative boundary data, provided in shapefile format, is available from the GADM website (https://gadm.org/), which is instrumental for geographic and spatial analysis, delineating various administrative regions. Each dataset plays a critical role in providing comprehensive insights into the study area, supporting a wide range of analyses from elevation and land cover to soil properties and precipitation.
2.3.2 Soil Erosion Inventory Mapping
The soil erosion inventory map is a key component in developing the SES model, serving as the dependent variable for this study. Accurately mapping the SES of the Saddang watershed required identifying both eroded and non-eroded areas. To facilitate this, the coordinates of 1992 locations—993 with soil erosion and 999 without—were collected through field surveys and analyzed using SAS Planet and Google Earth. These data points were then used to create a binary SES model, categorizing locations by soil erosion occurrence or non-occurrence. For model development, 1195 samples (60% of the total) were randomly selected, with the remaining 797 samples (40%) set aside for validation (Fig. 1). The types of soil erosion identified in the study included sheet, rill, gully, and mass movements.
2.3.3 Selection of the SES Factors
The selection of factors for this study was carefully guided by several criteria: the availability of reliable data, insights from existing literature and prior research, the connectivity and variability of the data, and the specific geo-environmental characteristics of the study area. Based on these considerations, a comprehensive set of 11 key SES factors was identified and compiled. This set includes various hydrological factors such as rainfall erosivity, which measures the potential of rainfall to cause soil erosion, and the Topographical Wetness Index (TWI), which indicates areas of water accumulation. Additional hydrological factors include the distance to the river, the Stream Power Index (SPI), which assesses the energy of water flow, and drainage density, which reflects the network of water channels in the area. Topographic factors include the slope-length factor, which quantifies the impact of slope length on soil erosion, and the Topographic Roughness Index (TRI), which measures the variability of the terrain. Environmental factors encompass bulk density, which affects soil cohesion, clay ratio, which influences soil texture, soil organic carbon, which contributes to soil health, and the Normalized Difference Vegetation Index (NDVI), which indicates vegetation cover. The spatial distribution of SES factors can be seen in Fig. 3.
2.3.3.1 Rainfall Erosivity
Rainfall erosivity measures the potential of rainfall to cause soil erosion based on the intensity and kinetic energy of rainfall events. High rainfall erosivity indicates a greater potential for soil detachment and transport. The most common equation used to calculate rainfall erosivity is the R-factor in the Universal Soil Loss Equation (USLE):
$$R=\sum\limits_{{i=1}}^{{12}} {1.735 \times {{10}^{\left( {1.5{{\log }_{10}}\left( {\frac{{{P_m}^{2}}}{{{P_a}}}} \right) - 0.018188} \right)}}}$$
1
where R is the rainfall erosivity factor (MJ mm ha− 1 h− 1 year− 1), Pm is the monthly rainfall (mm), and Pa is the annual rainfall (mm).
2.3.3.2 Topographical Wetness Index (TWI)
The Topographical Wetness Index indicates an area's susceptibility to soil saturation and water accumulation, which can influence soil erosion by increasing soil moisture and reducing its stability. High TWI values are often associated with greater soil erosion potential due to waterlogged conditions. It is calculated using the formula:
$${\text{TWI}}=\ln \left( {\frac{{{A_s}}}{{\tan \beta }}} \right)$$
2
where As is the upstream contributing area and β is the slope gradient (in radians).
2.3.3.3 Stream Power Index (SPI)
The Stream Power Index represents the erosive power of flowing water and its capacity to transport sediment. Higher SPI values suggest a higher potential for soil erosion due to the increased force exerted by flowing water. The value can be computed with the following formula:
$${\text{SPI}}=\ln \left( {{A_s}\tan \beta } \right)$$
3
where As is the upstream contributing area and β is the slope gradient (in radians).
2.3.3.4 Distance to River
The distance to a river significantly influences soil erosion by determining the likelihood of sediment transport into water bodies. Areas closer to rivers are more susceptible to soil erosion due to the increased potential for sediment movement and higher water flow, which can lead to both surface soil erosion and riverbank instability. As proximity to the river increases, the SES and subsequent sedimentation in the water body also rise, exacerbating soil erosion processes in these vulnerable zones.
2.3.3.5 Drainage Density
Drainage density reflects the total length of streams and rivers per unit area in a watershed. Higher drainage density indicates a more dissected landscape, which can enhance surface runoff and increase soil erosion potential. It is calculated using the following expression:
$${D_d}=\frac{{{l_s}}}{{{W_A}}}$$
4
where lS is the total length of the river (km) and WA is a watershed area (km2).
2.3.3.6 Slope Length Factor
The Slope Length Factor represents the effect of slope length and steepness on soil erosion. Longer and steeper slopes generally contribute to greater soil erosion due to increased velocity and volume of surface runoff. It is commonly calculated using the LS factor in the USLE:
\({\text{LS}}={\left( {\frac{\lambda }{{22.13}}} \right)^m}10.8\sin \beta +0.03\) if tan β < 0.09 (5)
\({\text{LS}}={\left( {\frac{\lambda }{{22.13}}} \right)^m}16.8\sin \beta - 0.5\) if tan β ≥ 0.09 (6)
$${\text{m}}=\frac{{\text{F}}}{{1+{\text{F}}}}$$
7
$$F=\frac{{{{\sin \beta } \mathord{\left/ {\vphantom {{\sin \beta } {0.0896}}} \right. \kern-0pt} {0.0896}}}}{{3{{\left( {\sin \beta } \right)}^{0.8}}+0.56}}$$
8
where λ is slope length (m), β is the slope gradient (in radians), m is the slope length exponent, and F is the ratio between rill soil erosion and interrill soil erosion.
2.3.3.7 Topographic Roughness Index (TRI)
The Topographic Roughness Index measures the variability of terrain elevation, with higher values indicating a more rugged landscape. Rugged terrains often have more intense soil erosion due to increased surface runoff and reduced vegetation cover. It is calculated by the standard deviation of elevation within a specified window or grid cell using the equation:
$${\text{TRI}}=Y{\left( {\sum {{{\left( {{x_{ij}} - {x_{00}}} \right)}^2}} } \right)^{0.5}}$$
9
where xij is the elevation of the neighbor grid (0,0).
2.3.3.8 Bulk Density
Bulk density reflects the soil's compactness and porosity, influencing water infiltration and root penetration. High bulk density indicates compacted soil with lower infiltration rates, leading to increased surface runoff and potential soil erosion. Bulk density is typically measured directly using soil samples.
2.3.3.9 Clay Ratio
The clay ratio indicates the proportion of clay particles in the soil, which affects soil structure and its susceptibility to soil erosion. Soils with higher clay content can be more cohesive, reducing soil erosion, but under certain conditions, they may also be susceptibility to crusting and soil erosion. The clay ratio is calculated as the percentage of clay particles relative to other soil particles. It is measured by applying:
10
where %clay is the percentage of clay content, %sand is the percentage of sand content, and %silt is the percentage of silt content.
2.3.3.10 Carbon Organic
Soil organic carbon content influences soil structure, stability, and resistance to soil erosion. Higher organic carbon levels improve soil aggregation, reducing soil erosion by increasing infiltration and decreasing runoff. It is measured in units of decigrams per kilogram (dg/kg) of soil.
2.3.3.11 Normalized Difference Vegetation Index (NDVI)
NDVI is a measure of vegetation cover and health, with higher values indicating more dense and vigorous vegetation. Dense vegetation protects soil from soil erosion by reducing the impact of raindrops and slowing surface runoff. NDVI is calculated using satellite imagery as:
$${\text{NDVI}}=\frac{{{\text{NI}}{{\text{R}}_{{\text{band}}}} - {\text{Re}}{{\text{d}}_{{\text{band}}}}}}{{{\text{NI}}{{\text{R}}_{{\text{band}}}}+{\text{Re}}{{\text{d}}_{{\text{band}}}}}}$$
11
Where NIR is light reflected in the near-infrared spectrum and REDband is light reflected in the red range of the spectrum.
2.3.4 Soil Erosion Modeling Using Machine Learning Models
Soil erosion modeling using ML involves predicting the likelihood and extent of soil erosion based on various environmental factors. ML models can capture complex relationships between these factors and soil erosion processes, offering a flexible and data-driven approach to SES assessment.
2.3.4.1 Support Vector Machines (SVM)
Support Vector Machines (SVM) are applied in soil erosion modeling to classify and predict areas at SES based on various hydrological, environmental, and topographical factors. SVM works by identifying the optimal hyperplane that separates different classes of SES with the maximum margin, effectively distinguishing between high-susceptibility and low- low-susceptibility zones. The model handles complex, non-linear relationships between factors by using kernel functions to map the input data into a higher-dimensional space where the classes become linearly separable. A common equation used in SVM classification is:
$$f\left( x \right)=sign\left( {w\phi \left( x \right)+b} \right)$$
12
where w is the weight vector, ϕ(x) represents the transformation of the input data into a higher-dimensional space, and b is the bias term. The function f(x) determines the class label based on the sign of the output. By training on known soil erosion location and no-soil erosion location data, SVM can effectively predict SES in new, untested areas, making it a valuable tool in soil conservation and land management planning.
2.3.4.2 Generalized Linear Models (GLM)
Generalized Linear Models (GLM) are used in soil erosion studies to analyze the relationship between soil erosion factors and soil erosion location or no-soil erosion location data by extending linear regression to handle non-normal distributions of the response variable, such as binary or count data. GLMs link predictors to soil erosion outcomes through a specified link function, allowing for the modeling of complex and non-linear relationships. The flexibility of GLM makes it suitable for predicting SES under varying hydrological, environmental, and topographical conditions. The general equation for a GLM is:
$$g\left( \mu \right)={\beta _0}+{\beta _1}{X_1}+{\beta _2}{X_2}+ \cdots +{\beta _p}{X_p}$$
13
where g(µ) is the link function that relates the mean of the response variable µ to the linear predictors β0, β1, …, βp and X1, X2, …, Xp are the independent variables (SES factors). GLMs are valuable in understanding the probabilistic nature of soil erosion and predicting its occurrence under various scenarios.
2.3.5 Evaluating The Models’ Performance
Several metrics and methods are employed to evaluate the performance of SVM and GLM in predicting SES, assessing the models' accuracy, reliability, and ability to generalize to new data, each serving to assess different aspects of model performance. These evaluations are crucial for determining how well the models perform, ensuring they provide reliable predictions across various scenarios and conditions. These evaluations are essential for determining the accuracy of the models in correctly identifying areas at SES, their reliability in providing consistent results across different datasets, and their generalization ability to make accurate predictions in new, unseen data. Metrics such as accuracy, precision, recall, and the F Mesure provide insights into the model’s ability to balance true positive (TP) and false positive rates (FPR), while the ROC curve and AUC measure the overall discriminative power of the model. Cross-validation techniques further ensure that the models are not overfitting to the training data, enhancing their robustness and reliability when applied to different environmental conditions. Together, these methods create a comprehensive framework for evaluating and refining the performance of SVM and GLM in SES modeling.
2.3.6 Normalization of Results
Normalization of soil erosion susceptibility (SES) results involves scaling the output predictions to a common range to ensure consistent interpretation and comparison across different datasets or scenarios. This process adjusts the SES values to a standardized scale, typically [0, 1], to make the results more comparable and to avoid biases introduced by varying magnitudes in the raw data. It can be measured by applying the equation:
$${X_{norm}}=\frac{{X - {X_{\hbox{min} }}}}{{{X_{\hbox{max} }} - {X_{\hbox{min} }}}}$$
14
where X is the original value, Xmin is the minimum value in the dataset, and Xmax is the maximum value in the dataset.
2.3.6 Weighting and Scoring
Modeling SES using weighted forms of SVM and GLM involves assigning weights to various factors to reflect their relative significance in predicting SES. In this approach, each factor is assigned a score based on its influence on SES (Table 2). These weights are then multiplied by factor-specific scores for different factor classes in Table 2, ensuring that the contributions of each factor are accurately represented in the model's predictions.
| No. | SES Factors | Categories | Weight | Classes of Factors | Area (km2) | Area (%) | Scores | |
| SVM | GLM | |||||||
| 1 | Rainfall erosivity (MJ mm ha− 1 h− 1 year− 1) | Hydrological Data | 0.076 | 0.085 | < 1,750 | - | - | 1 |
| 1,750-2,000 | 1,430 | 29.1 | 2 | |||||
| 2,000–2,250 | 1,951 | 39.7 | 3 | |||||
| 2,250-2,500 | 1,528 | 31.1 | 4 | |||||
| > 2,500 | - | - | 5 | |||||
| 2 | Topographical Wetness Index (TWI) | 0.093 | 0.063 | < 5 | 1,960 | 39.9 | 1 | |
| 5–10 | 2,692 | 54.8 | 2 | |||||
| 10–15 | 227 | 4.6 | 3 | |||||
| 15–20 | 28 | 0.6 | 4 | |||||
| > 20 | 2 | 0.0 | 5 | |||||
| 3 | Stream Power Index (SPI) | 0.038 | 0.032 | < 0 | 19 | 0.4 | 1 | |
| 0–5 | 4,014 | 81.8 | 2 | |||||
| 5–10 | 830 | 16.9 | 3 | |||||
| 10–15 | 41 | 0.8 | 4 | |||||
| > 15 | 5 | 0.1 | 5 | |||||
| 4 | Distance to River (m) | 0.107 | 0.376 | > 1,600 | 2,777 | 56.6 | 1 | |
| 1,200-1,600 | 454 | 9.3 | 2 | |||||
| 800-1,200 | 492 | 10.0 | 3 | |||||
| 400–800 | 540 | 11.0 | 4 | |||||
| < 400 | 645 | 13.1 | 5 | |||||
| 5 | Drainage Density (km/km2) | 0.134 | 0.282 | 0.0-0.2 | 2,011 | 41.0 | 1 | |
| 0.2–0.4 | 1,815 | 37.0 | 2 | |||||
| 0.4–0.6 | 925 | 18.8 | 3 | |||||
| 0.6–0.8 | 150 | 3.1 | 4 | |||||
| 0.8-1.0 | 8 | 0.2 | 5 | |||||
| 6 | Slope Length Factor | Topographic Data | 0.121 | 0.052 | < 0.4 | 949 | 19.3 | 1 |
| 0.4–1.4 | 95 | 1.9 | 2 | |||||
| 1.4–3.1 | 210 | 4.3 | 3 | |||||
| 3.1–6.8 | 614 | 12.5 | 4 | |||||
| > 6.8 | 3,041 | 61.9 | 5 | |||||
| 7 | Topographic Roughness Index (TRI) | 0.064 | 0.052 | 0.0-0.2 | 949 | 19.3 | 1 | |
| 0.2–0.4 | 95 | 1.9 | 2 | |||||
| 0.4–0.6 | 210 | 4.3 | 3 | |||||
| 0.6–0.8 | 614 | 12.5 | 4 | |||||
| 0.8-1.0 | 3,041 | 61.9 | 5 | |||||
| 8 | Bulk Density (cg/cm3) | Environmental Data | 0.022 | 0.001 | < 50 | - | - | 1 |
| 50–75 | 18 | 0.4 | 2 | |||||
| 75–100 | 2,898 | 59.0 | 3 | |||||
| 100–125 | 1,993 | 40.6 | 4 | |||||
| > 125 | - | - | 5 | |||||
| 9 | Clay Ratio | 0.099 | 0.143 | 0.0-0.2 | - | - | 1 | |
| 0.2–0.4 | 129 | 2.6 | 2 | |||||
| 0.5–0.6 | 2,837 | 57.8 | 3 | |||||
| 0.7–0.8 | 1,915 | 39.0 | 4 | |||||
| 0.8-1.0 | 28 | 0.6 | 5 | |||||
| 10 | Carbon Organic (dg/kg) | 0.088 | 0.098 | > 125 | 62 | 1.3 | 1 | |
| 100–125 | 1,865 | 38.0 | 2 | |||||
| 75–100 | 2,175 | 44.3 | 3 | |||||
| 50–75 | 788 | 16.0 | 4 | |||||
| < 50 | 19 | 0.4 | 5 | |||||
| 11 | Normalized Difference Vegetation Index (NDVI) | 0.108 | 0.283 | > 0.7 | - | - | 1 | |
| 0.5–0.7 | 313 | 6.4 | 2 | |||||
| 0.3–0.5 | 3,889 | 79.2 | 3 | |||||
| 0.2–0.3 | 305 | 6.2 | 4 | |||||
| < 0.2 | 402 | 8.2 | 5 | |||||
3.1 Model Performance Evaluation
The comparative analysis of SVM and GLM performance in predicting SES reveals nuanced differences that can guide model selection based on specific research goals. Table 3, SVM shows a slight advantage in accuracy (87.4%) over GLM (87.2%), a crucial metric for overall model performance in correctly predicting both positive and negative outcomes. Accuracy remains a widely used measure in SES studies, indicating the proportion of correctly classified instances across the total cases. On the other hand, GLM excels in several critical performance areas. Figure 4 shows the higher Area AUC for GLM (0.939 vs. SVM's 0.916) suggesting that GLM is better at distinguishing between SES areas and those that are not, which is vital in developing reliable SES maps. Furthermore, GLM's higher recall and sensitivity (both at 89.4%) compared to SVM (both at 86.5%) indicate that GLM is more effective in identifying true positives, reducing the susceptibility of underestimating areas susceptibility to soil erosion—a key factor in environmental conservation and land management (Rahmati et al., 2017). Additionally, GLM's higher specificity (88.3% vs. SVM's 85%) underscores its effectiveness in correctly identifying non-susceptible areas, thereby minimizing false positives, which can lead to more targeted and cost-effective soil conservation strategies (Bui et al., 2020). Although SVM shows better precision (87.6% vs. GLM's 86.1%), indicating fewer false positives, the comprehensive performance of GLM across multiple metrics suggests that it may offer a more balanced approach for SES modeling, particularly when the goal is to minimize both false negatives and false positives.
In conclusion, while SVM's slightly higher accuracy and precision may be advantageous in certain contexts, GLM's superior performance in AUC, recall, sensitivity, and specificity makes it a more robust choice for SES modeling, especially in scenarios where accurate identification of both high-susceptibility and low- susceptibility areas is crucial for sustainable land management and environmental protection.
| Accuracy Metrics | Unit | ML Algorithm | |
|---|---|---|---|
| SVM | GLM | ||
| Accuracy | % | 87.4 | 87.2 |
| Classification Error | % | 12.6 | 12.8 |
| AUC | 0.916 | 0.939 | |
| Precision | % | 87.6 | 86.1 |
| Recall | % | 86.5 | 89.4 |
| F Measure | % | 87.0 | 87.7 |
| Sensitivity | % | 86.5 | 89.4 |
| Specificity | % | 85.0 | 88.3 |
3.2 Spatial Distribution of Soil Erosion Susceptibility (SES)
Table 4 compares SES across different classes using the SVM and GLM presents interesting contrasts in their predictions. GLM categorizes a notably larger area under "Very Low" susceptibility (24.55%) compared to SVM (4.37%), indicating that GLM tends to classify a larger portion of the landscape as having minimal SES. This could be attributed to GLM's generalization tendencies, possibly smoothing over finer variations in the data, resulting in broader classifications. On the other hand, SVM's higher sensitivity to data nuances leads it to identify smaller, more concentrated areas of "Very Low" susceptibility, suggesting a more conservative approach. Additionally, the "Moderate" susceptibility category shows a significant difference, with SVM assigning 40.59% of the area to this class, while GLM assigns only 21.84%. This discrepancy further underscores SVM's tendency to distribute SES across a wider range of moderate susceptibility areas, possibly indicating its precision in capturing gradual variations in SES.
In contrast, both models show remarkable alignment in the "Low" susceptibility category, with almost identical percentages (38.50% for SVM and 38.59% for GLM). This similarity suggests a consensus between the models in identifying regions with low SES, which might be attributed to clearer patterns or more robust data. However, the differences reemerge in the "High" and "Very High" classes, where SVM again shows a higher area in the "High" susceptibility category (15.56% compared to GLM's 13.32%), while GLM slightly exceeds SVM in the "Very High" class (1.23% versus 0.98%). This pattern could indicate that SVM is more sensitive to identifying areas that transition from moderate to high susceptibility, whereas GLM might be better at distinguishing extreme cases of susceptibility. Figure 5 of each model depicts that these highly vulnerable zones are primarily found along rivers and river branches. This pattern emerges because water movement in these areas frequently causes greater soil separation and transport. Because of their proximity to water bodies, soil particles are more likely to be swept away by surface runoff, resulting in increased soil erosion. Furthermore, the characteristics of river flow, such as meandering and the power of water during high discharge events, make these places more SES.
The overall distribution suggests that SVM provides a more nuanced and detailed classification of SES, which could be useful for targeted management interventions, while GLM offers a broader, more generalized assessment that might be advantageous for broader, strategic planning and prioritization of soil erosion control measures. These differences have important implications for land management, where the choice of model might depend on the specific goals of the soil erosion assessment—whether it is for detailed site-specific management or general planning and resource allocation.
| No. | Normalization Range | Classes of Soil Erosion Susceptibility | Support Vector Machines (SVM) | Generalized Linear Models (GLM) | ||||
|---|---|---|---|---|---|---|---|---|
| Grid Total | Area (km2) | % | Grid Total | Area (km2) | % | |||
| 1 | 0.0–0.2 | Very Low | 238,436 | 215 | 4.4 | 1,339,038 | 1,205 | 24.6 |
| 2 | 0.2–0.4 | Low | 2,099,817 | 1,890 | 38.5 | 2,104,967 | 1,894 | 38.6 |
| 3 | 0.4–0.6 | Moderate | 2,213,906 | 1,993 | 40.6 | 1,191,423 | 1,072 | 21.8 |
| 4 | 0.6–0.8 | High | 848,573 | 764 | 15.6 | 726,729 | 654 | 13.3 |
| 5 | 0.8–1.0 | Very High | 53,583 | 48 | 1.0 | 67,158 | 60 | 1.2 |
| Total | 5,454,315 | 4,909 | 100 | 5,429,315 | 4,886 | 100 | ||
3.3 Feature Importance and Interpretation
Table 2 provided highlights the weights assigned to various factors by each model, revealing insights into their relative significance. For SVM, the highest weights are assigned to distance to the river (0.107), drainage density (0.134), and slope length factor (0.121). These weights suggest that SVM places substantial emphasis on topographic and hydrological features. Recent studies support this prioritization. For example, Band et al. (2020) emphasized the importance of slope length and drainage density in predicting soil erosion due to their direct influence on runoff and soil erosion processes. Slope length affects the velocity of water flow, which is crucial in soil erosion, while drainage density impacts how water is distributed across the landscape (Vu Dinh et al., 2021). Additionally, the distance to rivers, which influences sediment transport and water flow patterns, is critical for SVM models, aligning with findings by Pourghasemi et al. (2020) that proximity to water bodies significantly affects soil erosion dynamics.
In contrast, GLM assigns the highest weights to distance to river (0.376), drainage density (0.282), and NDVI (0.283). This prioritization indicates GLM's greater sensitivity to proximity to water bodies and land cover changes. The importance of distance to rivers and drainage density in GLM reflects their significant roles in soil erosion processes, as supported by Band et al. (2020), who found that proximity to rivers affects sediment transport and soil erosion, while drainage density influences runoff patterns. The Normalized Difference weight in GLM underscores the role of vegetation cover in SES. Recent research by Bai et al. (2020) highlights that NDVI is a critical factor in soil erosion modeling because it reflects vegetation cover, which stabilizes soil and reduces soil erosion rates.
4.1 Strengths and Weaknesses of Each Model
For SES modelling, the GLM offers several advantages due to its flexibility and interpretability. GLM allows for the inclusion of various types of predictor variables and can model different types of SES responses, such as binary or continuous outcomes. This interpretability is crucial in understanding the relationship between environmental factors and SES, which aids in making informed land management decisions. For instance, the GLM's higher emphasis on factors like distance to river and NDVI (with parameter weights of 0.376 and 0.283, respectively) aligns with their significant roles in soil erosion processes. These factors are influence sediment transport and land cover, as supported by studies, which emphasize their importance in soil erosion dynamics (Arabameri et al., 2019; Gayen et al., 2019; Igwe et al., 2020; Rahmati et al., 2017). However, GLM’s assumption of a linear relationship between predictors and response can be a limitation, especially in the context of the complex and non-linear processes governing soil erosion (Jiang et al., 2021). Moreover, GLM may struggle with high-dimensional data or significant interactions between variables, as seen in the moderate performance it shows in the accuracy metrics with an AUC of 0.939 compared to the SVM's 0.916.
On the other hand, SVM are particularly well-suited for SES modelling due to their ability to handle non-linear relationships between predictors and response variables. SVM can effectively manage complex interactions and higher-dimensional data, making it advantageous in modelling the intricate processes involved in soil erosion. The SVM model’s emphasis on drainage density and slope length factor (with parameter weights of 0.134 and 0.121, respectively) reflects its strength in capturing topographic influences on soil erosion, which is consistent with the findings of (Olii et al., 2023). SVM's flexibility with different kernel functions allows it to adapt to various types of data distributions, enhancing its predictive accuracy in SES applications (Mustafa et al., 2018). However, SVM also has drawbacks, including being computationally intensive and less interpretable (Devos et al., 2009), which can be challenging for practitioners who need to understand the factors influencing SES. Despite these challenges, SVM's ability to provide accurate predictions is reflected in its performance metrics, such as a precision of 87.6% and recall of 86.5%, indicating its robustness in SES classification.
4.2 Implications for Spatial Susceptibility Assessment and Environmental Management
The findings of SES assessments have significant implications for land management practices, particularly when integrating ML models like SVM and GLM with geospatial data. The data from the SES models, illustrated in the provided Table 4, shows that both SVM and GLM effectively identify areas with varying degrees of SES, with GLM highlighting a larger area of very low susceptibility (24.55%) compared to SVM (4.37%). This precision in identifying high-susceptibility areas allows for more targeted interventions, such as reforestation, terracing, and the construction of check dams, which are crucial in preventing soil erosion and enhancing land productivity (Mosavi et al., 2020). Moreover, the data suggests that different models may emphasize different areas, as seen with the distribution of high and very high susceptibility areas, where SVM shows a higher percentage of moderate susceptibility (40.59%) compared to GLM (21.84%). These insights support findings that targeted soil conservation measures can lead to improved agricultural yields and reduced environmental degradation, demonstrating the practical benefits of using ML models in SES assessments (Folharini et al., 2023; Ghorbanzadeh et al., 2020; Nguyen et al., 2023; Wang et al., 2023). The application of such data-driven insights in land management optimizes resource allocation and aligns with sustainable development goals by mitigating the impact of soil erosion on ecosystems and human livelihoods (Arabameri et al., 2018; Rahmati et al., 2017).
The broader environmental implications of spatial SES assessments extend beyond direct land management applications, particularly when considering the different weights and distributions identified by SVM and GLM models. For instance, SVM’s higher sensitivity in identifying moderately susceptible areas (40.59%) suggests that it may be more effective in predicting and managing soil erosion in regions where SES is not immediately apparent but still significant. In contrast, GLM’s emphasis on larger very low susceptibility areas (24.55%) can help prioritize areas for preventive measures. Accurate SES mapping is crucial for watershed management (Farhan et al., 2013), as it helps predict sediment loads in rivers and reservoirs, which is essential for maintaining water quality and preventing downstream flooding (Olii, et al., 2024). Additionally, the adaptability of these models to various geographic and environmental contexts, as indicated by the correlation between different model outputs and environmental factors, highlights their potential in global environmental management efforts. This adaptability is vital for integrating SES models into climate change adaptation strategies, predicting shifts in soil erosion patterns due to changing precipitation regimes and land use dynamics (Eekhout & de Vente, 2022). Ultimately, SES models, with their nuanced understanding of different susceptibility levels, provide valuable tools for combating desertification, preserving biodiversity, and supporting sustainable development on a global scale (Allafta & Opp, 2022; Das et al., 2020; François et al., 2024).
4.3 Limitations and Future Research Directions
Despite advancements in modelling techniques, the study faces several limitations that could impact the accuracy and generalizability of the results. One significant limitation is the dependency on the quality and resolution of input data, which can directly affect the precision of model outputs (Olii et al., 2021). For instance, the models' ability to accurately classify areas into different SES classes, as shown in the provided Tables 3, relies heavily on the quality of geospatial data. The SVM model, which identified 40.59% of the area as moderately susceptible, and the GLM model, which identified 21.84% in the same category, demonstrate how variations in data quality can influence model outcomes. The generalizability of these findings may also be constrained by the specific environmental and geographic characteristics of the study area, making it challenging to apply the results to regions with different climates, soil types, and land use patterns (Mosavi et al., 2020). Additionally, while SVM and GLM are robust models, their application may introduce biases if not properly calibrated, or if the classification and weighting systems, such as those seen in the different parameter weights for factors like distance to river and drainage density, are overly simplistic (Senanayake et al., 2020). These limitations highlight the importance of carefully interpreting the results and validating the models with independent datasets to ensure their reliability (Conoscenti et al., 2014).
Future research should focus on addressing these limitations by exploring new avenues to enhance the accuracy and applicability of SES models. For instance, the use of higher-resolution spatial data could improve model performance by providing more detailed inputs, which would be particularly beneficial in refining the classification of areas with very high or very low SES, as shown by the differences in areas classified by SVM and GLM. Research could also investigate the effectiveness of more advanced ML models, such as Deep Learning or Neural Networks, which may better capture the complex interactions between environmental variables and soil erosion processes (Khosravi et al., 2023). Moreover, future studies should explore integrating real-time data from remote sensing technologies to enable dynamic monitoring of SES (Musasa et al., 2024). This approach could significantly enhance the responsiveness of land management strategies to ongoing environmental changes, aligning with the need for more adaptive and resilient frameworks for SES assessment and environmental management (Ejegu & Yegizaw, 2021). By tackling these areas, future research can contribute to developing more robust and comprehensive models that not only improve predictive accuracy but also broaden the applicability of SES assessments across diverse environmental contexts.
The integration of Machine Learning (ML) techniques such as Support Vector Machines (SVM) and Generalized Linear Models (GLM) in the geospatial modelling of Soil Erosion Susceptibility (SES) provides a significant enhancement in predictive accuracy and reliability over traditional methods. The analysis shows that the SVM model predominantly classifies areas as moderate (40.59%) and low (38.50%) susceptibility, while the GLM model identifies a larger portion of the area as very low (24.55%) and low (38.59%) susceptibility. Both models exhibit high accuracy metrics, with SVM achieving an accuracy of 87.4% and GLM close behind at 87.2%. Furthermore, the GLM model demonstrates a slightly better Area Under the Curve (AUC) score of 0.939, compared to 0.916 for the SVM model, indicating superior model performance in distinguishing between different classes of SES. The analysis of contributing factors highlights distinct parameter weights, with the GLM model placing greater emphasis on hydrological factors such as distance to rivers and drainage density, while the SVM model gives more balanced attention across multiple environmental variables. These differences suggest that GLM might offer a more nuanced understanding of hydrological influences on soil erosion. The robust performance of these ML models, validated through cross-validation techniques, underscores their potential for dynamic and accurate SES assessments. This is crucial for guiding effective land management and conservation strategies, particularly in regions vulnerable to soil erosion. These findings emphasize the value of adopting innovative, data-driven approaches in environmental modelling to address the complex challenges of land degradation and soil conservation.
Acknowledgements
We sincerely thank the Faculty of Engineering at Universitas Gorontalo for their invaluable support and resources, which were crucial to the success of this research. We also appreciate the contributions of our colleagues, research assistants, and collaborators, whose expertise and dedication significantly enhanced the quality of this study.
Author Contributions
Olii, M.R. led the conceptualization, methodology design, and overall supervision of the research project, and played a significant role in drafting and revising the manuscript. Nento, S. was responsible for data collection, statistical analysis, and interpretation of results, contributing to the writing of the results and discussion sections. Doda, N. conducted the literature review and developed the theoretical framework, assisting in writing the introduction and literature review sections. Djafar, H. handled fieldwork, data acquisition, and processing, and contributed to the development of environmental variables for the SES model. Olii, R.Z.N. supported software development, model simulation, and validation, and was involved in visualizing the results. Pakaya, P. contributed to writing, formatting the manuscript, and preparing tables and figures for publication. All authors reviewed and approved the final manuscript.
Data availability
The data supporting this study's findings are available from the corresponding author upon request, for academic and research purposes, subject to confidentiality agreements and data use policies.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper
- Al-Bawi, A. J., Al-Abadi, A. M., Pradhan, B., & Alamri, A. M. (2021). Assessing gully erosion susceptibility using topographic derived attributes, multi-criteria decision-making, and machine learning classifiers. Geomatics, Natural Hazards and Risk, 12(1), 3035–3062. https://doi.org/10.1080/19475705.2021.1994024
- Allafta, H., & Opp, C. (2022). Soil Erosion Assessment Using the RUSLE Model, Remote Sensing, and GIS in the Shatt Al-Arab Basin (Iraq-Iran). Applied Sciences (Switzerland), 12(15), 1–17. https://doi.org/10.3390/app12157776
- Almouctar, M. A. S., Wu, Y., Zhao, F., & Dossou, J. F. (2021). Soil erosion assessment using the rusle model and geospatial techniques (Remote sensing and gis) in south-central niger (maradi region). Water (Switzerland), 13(24). https://doi.org/10.3390/w13243511
- Arabameri, A., Cerda, A., & Tiefenbacher, J. P. (2019). Spatial pattern analysis and prediction of gully erosion using novel hybrid model of entropy-weight of evidence. Water (Switzerland), 11(6), 1–23. https://doi.org/10.3390/w11061129
- Arabameri, A., Pradhan, B., Pourghasemi, H. R., & Rezaei, K. (2018). Identification of erosion-prone areas using different multi-criteria decision-making techniques and gis. Geomatics, Natural Hazards and Risk, 9(1), 1129–1155. https://doi.org/10.1080/19475705.2018.1513084
- Band, S. S., Janizadeh, S., Pal, S. C., Saha, A., Chakrabortty, R., Shokri, M., & Mosavi, A. (2020). Novel ensemble approach of deep learning neural network (Dlnn) model and particle swarm optimization (pso) algorithm for prediction of gully erosion susceptibility. Sensors (Switzerland), 20(19), 1–28. https://doi.org/10.3390/s20195609
- Bui, D. T., Tsangaratos, P., Nguyen, V. T., Liem, N. Van, & Trinh, P. T. (2020). Comparing the prediction performance of a Deep Learning Neural Network model with conventional machine learning models in landslide susceptibility assessment. Catena, 188(December 2019), 104426. https://doi.org/10.1016/j.catena.2019.104426
- Conoscenti, C., Angileri, S., Cappadonia, C., Rotigliano, E., Agnesi, V., & Märker, M. (2014). Gully erosion susceptibility assessment by means of GIS-based logistic regression: A case of Sicily (Italy). Geomorphology, 204, 399–411. https://doi.org/10.1016/j.geomorph.2013.08.021
- Das, B., Bordoloi, R., Thungon, L. T., Paul, A., Pandey, P. K., Mishra, M., & Tripathi, O. P. (2020). An integrated approach of GIS, RUSLE and AHP to model soil erosion in West Kameng watershed, Arunachal Pradesh. Journal of Earth System Science, 129(1), 1–18. https://doi.org/10.1007/s12040-020-1356-6
- Devos, O., Ruckebusch, C., Durand, A., Duponchel, L., & Huvenne, J. P. (2009). Support vector machines (SVM) in near infrared (NIR) spectroscopy: Focus on parameters optimization and model interpretation. Chemometrics and Intelligent Laboratory Systems, 96(1), 27–33. https://doi.org/10.1016/j.chemolab.2008.11.005
- Eekhout, J. P. C., & de Vente, J. (2022). Global impact of climate change on soil erosion and potential for adaptation through soil conservation. Earth-Science Reviews, 226, 103921. https://doi.org/10.1016/j.earscirev.2022.103921
- Ejegu, M. A., & Yegizaw, E. S. (2021). Modeling soil erosion susceptibility and LULC dynamics for land degradation management using geoinformation technology in Debre Tabor district, Northwestern highlands of Ethiopia. J. Degrade. Min. Land Manage, 8(2), 2623–2633. https://doi.org/10.15243/jdmlm.2021.082.2623
- Farhan, Y., Zregat, D., & Farhan, I. (2013). Spatial Estimation of Soil Erosion Risk Using RUSLE Approach, RS, and GIS Techniques: A Case Study of Kufranja Watershed, Northern Jordan. In Journal of Water Resource and Protection (Vol. 05, Issue 12, pp. 1247–1261). https://doi.org/10.4236/jwarp.2013.512134
- Folharini, S., Vieira, A., Bento-Gonçalves, A., Silva, S., Marques, T., & Novais, J. (2023). Soil Erosion Quantification using Machine Learning in Sub-Watersheds of Northern Portugal. Hydrology, 10(1). https://doi.org/10.3390/hydrology10010007
- François, M., Gonçalves Pontes, M. C., de Vasconcelos, R. N., de Oliveira, U. C., Peixoto da Silva, H., Faria, D., & Mariano-Neto, E. (2024). Assessing soil erosion and its drivers in agricultural landscapes: a case study in southern Bahia, Brazil. Journal of Water and Climate Change, 00(0), 1–16. https://doi.org/10.2166/wcc.2024.147
- Gayen, A., Pourghasemi, H. R., Saha, S., Keesstra, S., & Bai, S. (2019). Gully erosion susceptibility assessment and management of hazard-prone areas in India using different machine learning algorithms. Science of the Total Environment, 668, 124–138. https://doi.org/10.1016/j.scitotenv.2019.02.436
- Ghorbanzadeh, O., Shahabi, H., Mirchooli, F., Valizadeh Kamran, K., Lim, S., Aryal, J., Jarihani, B., & Blaschke, T. (2020). Gully erosion susceptibility mapping (GESM) using machine learning methods optimized by the multi‑collinearity analysis and K-fold cross-validation. Geomatics, Natural Hazards and Risk, 11(1), 1653–1678. https://doi.org/10.1080/19475705.2020.1810138
- Golijanin, J., Nikolić, G., Valjarević, A., Ivanović, R., Tunguz, V., Bojić, S., Grmuša, M., Lukić Tanović, M., Perić, M., Hrelja, E., & Stankov, S. (2022). Estimation of potential soil erosion reduction using GIS-based RUSLE under different land cover management models: A case study of Pale Municipality, B&H. Frontiers in Environmental Science, 10(August), 1–13. https://doi.org/10.3389/fenvs.2022.945789
- Golkarian, A., Khosravi, K., Panahi, M., & Clague, J. J. (2023). Spatial variability of soil water erosion: Comparing empirical and intelligent techniques. Geoscience Frontiers, 14(1), 101456. https://doi.org/10.1016/j.gsf.2022.101456
- Huang, D., Su, L., Zhou, L., Tian, Y., & Fan, H. (2023). Assessment of gully erosion susceptibility using different DEM-derived topographic factors in the black soil region of Northeast China. International Soil and Water Conservation Research, 11(1), 97–111. https://doi.org/10.1016/j.iswcr.2022.04.001
- Igwe, O., John, U. I., Solomon, O., & Obinna, O. (2020). GIS-based gully erosion susceptibility modeling, adapting bivariate statistical method and AHP approach in Gombe town and environs Northeast Nigeria. Geoenvironmental Disasters, 7(1). https://doi.org/10.1186/s40677-020-00166-8
- Jiang, C., Yang, Z., Wang, X., Dong, X., Li, Z., & Li, C. (2021). Examining the reversal of soil erosion decline in the hotspots of sandstorms: A non-linear ecosystem dynamic perspective. Journal of Arid Environments, 186(October 2020), 104421. https://doi.org/10.1016/j.jaridenv.2020.104421
- Khosravi, K., Rezaie, F., Cooper, J. R., Kalantari, Z., Abolfathi, S., & Hatamiafkoueieh, J. (2023). Soil water erosion susceptibility assessment using deep learning algorithms. Journal of Hydrology, 618. https://doi.org/10.1016/j.jhydrol.2023.129229
- Kucuker, D. M., & Cedano Giraldo, D. (2022). Assessment of soil erosion risk using an integrated approach of GIS and Analytic Hierarchy Process (AHP) in Erzurum, Turkiye. Ecological Informatics, 71, 101788. https://doi.org/https://doi.org/10.1016/j.ecoinf.2022.101788
- Mosavi, A., Sajedi-Hosseini, F., Choubin, B., Taromideh, F., Rahi, G., & Dineva, A. A. (2020). Susceptibility mapping of soil water erosion using machine learning models. Water (Switzerland), 12(7). https://doi.org/10.3390/w12071995
- Musasa, T., Dube, T., & Marambanyika, T. (2024). Landsat satellite programme potential for soil erosion assessment and monitoring in arid environments: A review of applications and challenges. International Soil and Water Conservation Research, 12(2), 267–278. https://doi.org/10.1016/j.iswcr.2023.10.003
- Mustafa, M. R. U., Sholagberu, A. T., Yusof, K. W., Hashim, A. M., Khan, M. W. A., & Shahbaz, M. (2018). SVM-Based Geospatial Prediction of Soil Erosion under Static and Dynamic Conditioning Factors. MATEC Web of Conferences, 203. https://doi.org/10.1051/matecconf/201820304004
- Nguyen, C. Q., Tran, T. T., Nguyen, T. T. T., Nguyen, T. T., Astarkhanova, T. S., Vu, L. V., Dau, K. T., Nguyen, H. N., Pham, G. H., Nguyen, D. D., Prakash, I., & Pham, B. (2023). Mapping of soil erosion susceptibility using advanced machine learning models at Nghe An, Vietnam. Journal of Hydroinformatics, 26(1). https://doi.org/10.2166/hydro.2023.327
- Olii, M. R., Kironoto, B. A., Olii, A., Pakaya, R., & Olii, A. K. Z. (2024). Advancing Soil Erosion Assessment: Application of Remote Sensing and Geospatial Techniques in Bulango Ulu Reservoir Basin. E3S Web of Conferences, 476, 1–15. https://doi.org/10.1051/e3sconf/202447601041
- Olii, M. R., Olii, A. K. Z., Olii, A., Pakaya, R., & Kironoto, B. A. (2024). Spatial modeling of soil erosion risk a multi‑criteria decision‑making (MCDM) approach in the paguyaman watershed, gorontalo, Indo.pdf. Arabian Journal of Geosciences, 17(226), 1–13. https://doi.org/https://doi.org/10.1007/s12517-024-12032-0
- Olii, M. R., Olii, A., & Pakaya, R. (2021). The Integrated Spatial Assessment of The Flood Hazard Using AHP-GIS: The Case Study of Gorontalo Regency. Indonesian Journal of Geography, 53(1), 126–135. https://doi.org/10.22146/ijg.59999
- Olii, M. R., Olii, A., Pakaya, R., & Olii, M. Y. U. P. (2023). GIS-based analytic hierarchy process (AHP) for soil erosion-prone areas mapping in the Bone Watershed, Gorontalo, Indonesia. Environmental Earth Sciences, 82(9), 1–14. https://doi.org/10.1007/s12665-023-10913-3
- Phinzi, K., & Szabó, S. (2024). Predictive machine learning for gully susceptibility modeling with geo-environmental covariates: main drivers, model performance, and computational efficiency. In Natural Hazards (Vol. 120, Issue 8). Springer Netherlands. https://doi.org/10.1007/s11069-024-06481-9
- Pourghasemi, H. R., Sadhasivam, N., Kariminejad, N., & Collins, A. L. (2020). Gully erosion spatial modelling: Role of machine learning algorithms in selection of the best controlling factors and modelling process. Geoscience Frontiers, 11(6), 2207–2219. https://doi.org/10.1016/j.gsf.2020.03.005
- Rahmati, O., Tahmasebipour, N., Haghizadeh, A., Pourghasemi, H. R., & Feizizadeh, B. (2017). Evaluating the influence of geo-environmental factors on gully erosion in a semi-arid region of Iran: An integrated framework. Science of the Total Environment, 579, 913–927. https://doi.org/10.1016/j.scitotenv.2016.10.176
- Saini, S. S., Jangra, R., & Kaushik, S. P. (2015). Vulnerability Assessment of Soil Erosion Using Geospatial Techniques-a Pilot Study of Upper Catchment of Markanda River. Research Article International Journal of Advancement in Remote Sensing, 3(1), 9–21.
- Senanayake, S., Pradhan, B., Huete, A., & Brennan, J. (2020). A review on assessing and mapping soil erosion hazard using geo-informatics technology for farming system management. Remote Sensing, 12(24), 1–25. https://doi.org/10.3390/rs12244063
- Vu Dinh, T., Hoang, N. D., & Tran, X. L. (2021). Evaluation of Different Machine Learning Models for Predicting Soil Erosion in Tropical Sloping Lands of Northeast Vietnam. Applied and Environmental Soil Science, 2021. https://doi.org/10.1155/2021/6665485
- Wang, Y., Zhang, Y., & Chen, H. (2023). Gully erosion susceptibility prediction in Mollisols using machine learning models. Journal of Soil and Water Conservation, 78(5), 385–396. https://doi.org/10.2489/jswc.2023.00019
No competing interests reported.