LSTM time series prediction of soil moisture content in kiwifruit root zone based on meteorological data fusion

• Fusion of meteorological data improves prediction accuracy of soil moisture content • Time-series networks are more effective in capturing key features in time series • Novel ATT-LSTM model performs best in predicting soil moisture content • Time-series networks are a feasible solution for predicting soil moisture content Accurate prediction of soil moisture content (SMC) is important for water resource management, irrigation optimization, and drought monitoring. Current research focuses on inverse monitoring of SMC, but management decisions based on inverse results often suffer from lags. In contrast, predictive studies of SMC can provide information on future dynamics, which can help to respond proactively to environmental changes, make long-term plans, and have more substantial prospective and practical value. In this study, we constructed an integrated network model (ATT-LSTM) based on the attention mechanism, which uses vegetation index (VI) and meteorological data (MD) as input features to realize short-term (next two days) prediction of SMC in the root zone of kiwifruit plants. On this basis, combined with the Shapley additive explanations (SHAP) interpretable analysis method, this study systematically quantified the key influences on SMC prediction accuracy. It clarified the features in the VIs and MDs that contributed most to SMC prediction. To fully evaluate the model performance, we compared ATT-LSTM with traditional artificial neural network models, including non-temporal feedforward neural network (FFNN) and temporal Long Short-Term Memory (LSTM). The experimental results show that FFNN's prediction on the test set is poor due to its inability to capture time series information effectively. In the SMC prediction for the next day, the ATT-LSTM model performs the best, with a coefficient of determination (R²) of 0.749 and a root mean square error (RMSE) of only 2.04% on the test set. In the prediction for the next two days, despite the increased prediction difficulty, ATT-LSTM maintains its lead with R² and RMSE of 0.713 and 2.31% on the test set, respectively. The results show that ATT-LSTM has significant advantages in SMC prediction in agriculture and provides adequate technical support for irrigation scheduling and water resource management.