Self-Explaining Attention-Based Deep Learning Models for Soil Property Regression Using Multi-Temporal Sentinel-2 Data

This study explores the use of deep learning models with attention mechanisms for predicting multiple soil properties from multi-temporal Sentinel-2 satellite images. We developed and tested two convolutional neural network architectures: one with a shared attention layer for all target properties, and another with separate attention layers for each property. The models were trained and evaluated on data from 45.000 locations across the United States. The main goal was to analyze which satellite bands and months are most relevant for soil property regression, as identified by the learned attention weights. In additional experiments, we excluded the most important bands to test their impact on prediction accuracy. The results show that attention mechanisms help highlight relevant input features and that removing key bands generally reduces model performance. To the best of our knowledge, such a direct integration of attention mechanisms with regression models for multi-temporal Sentinel-2-based soil property mapping has not been previously reported. This approach provides insight into model behavior and may support future applications in remote sensing-based soil analysis.