This study explores the application of deep learning in predicting soil moisture, a crucial factor in the hydrological cycle. The authors evaluate 10 different network structures, including three basic feature extractors and seven hybrid structures, to uncover their data utilization mechanisms and maximize the potential of deep learning for soil moisture prediction. The models are systematically compared across different soil textures and depths, and their interpretability is analyzed using Shapley (SHAP) additive explanations. The results highlight that long short-term memory (LSTM) is well-suited for temporal modeling, and the improved accuracy achieved by feature attention LSTM (FA-LSTM) and generative-adversarial-network-based LSTM (GAN-LSTM) demonstrates the effectiveness of attention mechanisms and adversarial training in feature extraction. The study also reveals varying data leveraging approaches among different models, as indicated by Shapley values, and visualizes encoded features using t-distributed stochastic neighbor embedding (t-SNE). The findings provide insights into effective network design principles for soil moisture prediction and serve as a reference for deep learning applications in other hydrology problems. The codes for 3 machine learning and 10 deep learning models are open-source.This study explores the application of deep learning in predicting soil moisture, a crucial factor in the hydrological cycle. The authors evaluate 10 different network structures, including three basic feature extractors and seven hybrid structures, to uncover their data utilization mechanisms and maximize the potential of deep learning for soil moisture prediction. The models are systematically compared across different soil textures and depths, and their interpretability is analyzed using Shapley (SHAP) additive explanations. The results highlight that long short-term memory (LSTM) is well-suited for temporal modeling, and the improved accuracy achieved by feature attention LSTM (FA-LSTM) and generative-adversarial-network-based LSTM (GAN-LSTM) demonstrates the effectiveness of attention mechanisms and adversarial training in feature extraction. The study also reveals varying data leveraging approaches among different models, as indicated by Shapley values, and visualizes encoded features using t-distributed stochastic neighbor embedding (t-SNE). The findings provide insights into effective network design principles for soil moisture prediction and serve as a reference for deep learning applications in other hydrology problems. The codes for 3 machine learning and 10 deep learning models are open-source.