EUVIMEDEUVIMED entdecken
Crossref

Soil erosion susceptibility assessment based on the hybrid deep learning models and SHapley analyses

Thuy Ha Thi Nguyen, Tuyen Thi Tran, Hieu Ngoc Nguyen, Giang Hương Pham, Pal Souvik, Binh Thai Pham

Vietnam Journal of Earth Sciences · 17.07.2026

DOI: 10.15625/2615-9783/24948 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Vietnam Journal of Earth Sciences, online erschienen am 17.07.2026. Verfasst von Thuy Ha Thi Nguyen und 5 weiteren Autor:innen. Die Arbeit nennt 88 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study assesses spatial soil erosion susceptibility in Nghe An province, Vietnam, using four novel deep learning models: Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM architecture. A key innovation of this study is the integration of temporal sequence modeling with spatial feature learning via a hybrid CNN-LSTM architecture, an approach not commonly used in soil erosion studies. In addition, the SHapley Additive exPlanations (SHAP) technique was used to explain how the model worked and how it performed. CNN-LSTM demonstrated the highest validation accuracy (85.7%), sensitivity (0.964), Kappa (0.714), and area under the curve (AUC) of 0.927, along with the lowest error (0.378), outperforming all other models in prediction capability and robustness. While the LSTM model achieved the highest training AUC (0.947) and accuracy (87.7%), it showed signs of overfitting, with a validation accuracy of 83.9%. CNN offered stable performance (validation AUC: 0.892), while DNN lagged, limited by its lack of temporal context (validation AUC: 0.884). Using SHAP, elevation, geology, and aspect were identified as the most critical drivers of erosion susceptibility. The findings underscore the superiority of hybrid deep learning models in capturing the spatiotemporal complexity of erosion processes. The proposed CNN-LSTM model demonstrates strong potential for real-world implementation environments, where both spatial and temporal patterns influence erosion dynamics. This research presents a novel framework for erosion-susceptibility mapping and highlights the importance of integrating interpretable machine learning (via SHAP) with advanced modeling strategies.

Zitieren

APA

Nguyen, T. H. T., Tran, T. T., Nguyen, H. N., Pham, G. H., Souvik, P., & Pham, B. T. (2026). Soil erosion susceptibility assessment based on the hybrid deep learning models and SHapley analyses. Vietnam Journal of Earth Sciences. https://doi.org/10.15625/2615-9783/24948

Vancouver

Nguyen THT, Tran TT, Nguyen HN, Pham GH, Souvik P, Pham BT. Soil erosion susceptibility assessment based on the hybrid deep learning models and SHapley analyses. Vietnam Journal of Earth Sciences. 2026. doi:10.15625/2615-9783/24948

BibTeX
@article{Nguyen2026,
  author = {Nguyen, Thuy Ha Thi and Tran, Tuyen Thi and Nguyen, Hieu Ngoc and Pham, Giang Hương and Souvik, Pal and Pham, Binh Thai},
  title = {Soil erosion susceptibility assessment based on the hybrid deep learning models and SHapley analyses},
  journal = {Vietnam Journal of Earth Sciences},
  year = {2026},
  doi = {10.15625/2615-9783/24948}
}
Kostenlos starten →