Soil moisture is one of the most important parameters in agriculture, environmental science, and hydrology. It is commonly expressed as volumetric water content, defined as the ratio of the volume of water in the soil to the total soil volume. With the rapid development of machine learning, particularly the widespread application of convolutional neural networks in soil moisture prediction, SD-CNN (shallow-deep coupled convolutional neural network) was proposed. By processing remote sensing imagery or historical soil moisture data, the model automatically learns both shallow and deep features embedded in the data and utilizes these features to accurately predict soil moisture in unobserved regions. Compared with traditional observation methods and sensor-based measurements, SD-CNN can obtain wide-coverage, continuous soil moisture information within a shorter time and effectively capture its spatial variability, thereby significantly improving prediction accuracy and reliability. By applying SD-CNN to complex mountainous terrain and performing regional adaptation through iterative optimization and parameter tuning, the model achieves high-precision predictions of soil moisture for the upcoming year and demonstrates strong correlation and stability when validated with field measurements. Experimental results show that this soil moisture prediction capability not only improves precise irrigation and agricultural production efficiency, but also strengthens drought monitoring, flood forecasting, and ecological regulation. These advancements provide essential technical support for stable agricultural output, optimized water resource allocation, and ecological security, ultimately contributing to sustainable regional and national economic development.