Evaluating the impact of residual learning and feature fusion on soil moisture prediction accuracy

The architectural design of deep learning models significantly influences their predictive capabilities in environmental monitoring tasks. This paper investigates the individual and collective effects of residual learning and feature fusion mechanism to improve the performance of soil moisture estimation on the designed architecture of the deep learning model. In this study, the data fusion mechanism was used to integrate Normalized Difference Water Index (NDWI), Synthetic Aperture Radar (SAR), and satellite imagery datasets containing Red, Green, and Blue (RGB) color channels, which consist of images or data collected by a radar system that uses microwaves to produce images of the Earth's surface. Three model variants were developed, each selectively omitting one or more of these architectural elements, and their performance was evaluated using three standard metrics, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2). The results of the final proposed model architecture showed that while each component contributes to accuracy improvements, the combination of residual learning and feature fusion yields the most significant gains. Improved results of RMSE = 0.0117, R²=0.814 and Mean Absolute Error=0.0148 were obtained. These performance indicators were superior to the results of most of the baseline models after comparative analysis. Thus, this study provides insights into model component selection for deep learning soil moisture prediction applications.