OPTIMIZING DEEP LEARNING COMPONENTS FOR ENHANCEMENT OF SOIL MOISTURE PREDICTION MODELS FROM REMOTE SENSING DATA

Designing deep learning models for remote sensing based soil moisture estimation is a complex task that requires careful selection of architectural components to achieve high accuracy and generalization. This study aims to evaluate the impact of different deep learning configurations on soil moisture prediction performance. Four model variants were developed, each exploring distinct architectural features, including the presence or absence of feature fusion, residual connections, and squeeze and excitation attention blocks. The study collected dataset size of 2200 splited at the ratio of 70% used for training while 15% used for validation and the remaining 15% was used for testing. All models were trained on the same dataset using identical hyperparameters to ensure a fair comparison. Their performance was assessed using three standard regression metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²). Among the tested models, the best-performing variant achieved an RMSE of 0.0117 (1.17%), MAE of 0.0148 (1.48%), and an R² of 0.8639 consistently outperforming several baseline models reported in existing literature. These findings demonstrate competitive performance of targeted architectural enhancements in improving deep learning based soil moisture estimation and offer a valuable benchmark for future research in this domain.