Modeling land susceptibility to wind erosion hazards using LASSO regression and graph convolutional networks

Predicting land susceptibility to wind erosion is necessary mitigate the negative impacts of on soil fertility, ecosystems, and human health. This study first attempt model hazards through application a novel approach, graph convolutional networks (GCNs), as deep learning models with Monte Carlo dropout. approach applied Semnan Province in arid central Iran, an area vulnerable dust storms climate change. We mapped 15 potential factors controlling erosion, including climatic variables, characteristics, lithology, vegetation cover, use, digital elevation (DEM), then least absolute shrinkage selection operator (LASSO) regression discriminate most important factors. constructed predictive by randomly selecting 70% 30% pixels, training validation datasets, respectively, focusing locations severe inventory map. The current LASSO identified eight out features (four property categories, speed, evaporation) Province. These were adopted into GCN model, which estimated that 15.5%, 19.8%, 33.2%, 31.4% total characterized low, moderate, high, very high respectively. under curve (AUC) SHapley Additive exPlanations (SHAP) game theory assess performance interpretability output, AUC values for datasets at 97.2% 97.25%, indicating excellent prediction. SHAP ranged between −0.3 0.4, while analyses revealed coarse clastic component, use effective output. Our results suggest this suite methods highly recommended future spatial prediction other environments around globe.