Accurate and effective soil moisture prediction has gradually attracted attention due to the management of agricultural activities and the practical usage of water resources. Therefore, this research presents an integrated deep learning-based framework for soil moisture prediction, where long short-term memory (LSTM) layers, an attention mechanism, and fully connected layers are combined. In this framework, LSTM layers are applied to extract the complicated, long- and short-term dependencies from the time series soil moisture data. Besides, the attention mechanism is employed to learn the discriminative information from the extracted features. The genetic algorithm (GA), which aims to enhance the prediction accuracy, is applied to simultaneously determine the hyperparameters of the proposed network, called the GA-LSTM-ATT, consisting of the LSTM layer count, the hidden units in layers, the rate of dropout, and the learning rate. The experimental consequences have indicated that the GA-LSTM-ATT method outperforms the benchmark approaches in different evaluation indices. [ABSTRACT FROM AUTHOR]