Advanced attention-driven deep learning architectures for multi-depth soil temperature prediction

Le résumé fourni par la source

• Soil temperature (SDT) at four different depths was monitored and analyzed in the eastern Mediterranean. • Attention-based deep learning (DL) architecture was used for SDT prediction. • SDT was modeled using four DL models, LSTM, GRU, CNN, and Transformer. • LSTM outperformed the other models, followed by GRU and CNN. Soil temperature plays a crucial role in ecological stability by supporting balanced processes such as nutrient cycling, and the flow of water and energy. This research aimed to analyze and predict the dynamic relationship of multi depth soil temperature (SDT) at (5 cm, 10 cm, 20 cm, and 50 cm) with meteorological variables using Bi-wavelet coherence and deep learning models. Four well-structured, attention-based deep learning (DL) architectures, namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Transformer were employed for predicting multi depth soil temperature. The wavelet analysis revealed a high coherence of 0.8 to 1.0 between maximum and minimum temperature (Tmax & Tmin) and multi depth soil temperature from 2003 to 2007. Among deep learning models, proposed LSTM with attention layer outperformed GRU, CNN, and Transformer with highest prediction accuracy of R² = 0.951, RMSE = 1.809, and MSE = 3.273 during testing stage. Similarly, it gained highest accuracy of R² = 0.947, RMSE = 1.925, and MSE = 3.706 during validation at a soil depth of 10 cm. Notably, perturbation-based sensitivity analysis also confirmed LSTM as the superior model, with RMSE ranging from 0.180 at the lowest noise level (0.01) to 2.34 at the highest noise level (0.3). The SHAP kernel explanation of best performed LSTM architecture showed the highest positive contribution of Tmax and Tmin in predicting multi depth soil temperature. The output will support sustainability plans in Syria.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.