Gully erosion prediction using weight of evidence and advanced machine learning models

Xiaohui Chen, Alireza Arabameri, M. Santosh, Hasan Raja Naqvi, Mohd Ramiz,
Gully erosion prediction using weight of evidence and advanced machine learning models,
Advances in Space Research,
2025,
,
ISSN 0273-1177,
https://doi.org/10.1016/j.asr.2025.10.047.
(https://www.sciencedirect.com/science/article/pii/S0273117725011780)
Abstract: Spatial modelling and susceptibility mapping of soil erosion are crucial for planning effective control strategies and land-use changes to mitigate future degradation and aid soil conservation. In this study we employed an ensemble statistical data-driven based prediction of gully erosion (GUE) susceptibility in Central Iran using the Weight of Evidence (WOE) method combined with four models: Stochastic Gradient Descent (SGD), Entropy (ENT), Multi-layer Perceptron (MLP), and Rotation Forest (RF). We used, 424 gully erosion locations, with 292 samples used to train the models and the rest for validation. Sixteen conditioning factors were considered in predicting gully erosion susceptibility zones. For accuracy assessment, the ROC curve was used, and the results show that the RF-WOE model surpassed all other ensemble models with an accuracy of 93 %. This was followed by the SGD-WOE model at 92.6 %. Conversely, the WOE model alone had the lowest accuracy at 89 %. Meanwhile, the MLP-WOE and ENT-WOE models showed accuracies of 92.3 % and 92.0 %, respectively. According to the RF-WOE model, about 40 % and 47.17 % land was classified under high and very high susceptible zone respectively that are dominantly located in the north, northeast, and southwest of the study region. Furthermore, analysis using the entropy model indicated that geomorphology, with a value of 0.684, was the highest contributing factor to gully erosion susceptibility, followed by distance to streams at 0.449. Conversely, soil depth was the least contributing factor. In conclusion, study demonstrates that machine learning-based ensemble models, particularly RF-WOE, effectively identify gully erosion susceptibility zones with high accuracy, providing a valuable tool. This finding is significant for land managers and policymakers aiming to implement effective soil conservation measures in vulnerable regions.
Keywords: Gully erosion; Data-driven model; Remote sensing; Machine learning