AI Summary & Key Takeaways
AI InsightThis study provides a scientific and practical framework for soil management, vegetation conservation, and dust risk mitigation.
📌 3 Key Takeaways
Abstract
Atmospheric mineral dust critically influences climate systems, ecological frameworks, and human health, particularly in arid and semi-arid regions. While various modeling approaches have been applied to map dust susceptibility, conventional methods often fail to capture the complex spatio-temporal interactions governing dust emission dynamics. This study addresses this gap by evaluating four hybrid deep learning-support vector machine (DL-SVM) frameworks—CNN-SVM, DBN-SVM, LSTM-SVM, and AE-SVM—for predicting dust-prone areas in the Lake Urmia Basin, Iran. Quantitative analyses demonstrated that all models possess significant predictive capability. The LSTM-SVM model achieved the highest discriminative power during training (AUC = 0.981), while the AE-SVM model delivered the most robust generalization during testing, with an accuracy of 87.63%. SHAP-based feature importance analysis revealed that the Normalized Difference Vegetation Index (NDVI) was the uniformly dominant factor controlling dust susceptibility across all models. Complementary soil properties—texture, pH, and salinity—played a decisive secondary role. These findings confirm that hybrid deep learning architectures, which simultaneously leverage spatial, temporal, and non-linear feature representations, offer superior capability in identifying dust-source areas compared to traditional models. This study provides a scientific and practical framework for soil management, vegetation conservation, and dust risk mitigation.