Development of long short-term memory models using rainfall and soil moisture to predict soil moisture dynamics.

Long short-term memory (LSTM) models were developed using rainfall and soil moisture data to predict soil moisture at various depths in an Asian pear orchard. Two types of rainfall inputs were tested: hourly rainfall data treated as individual values and event-based rainfall data, where cumulative rainfall was calculated by applying the minimum inter-event time threshold (12 h). Soil moisture was measured at depths of 20, 40, and 60 cm from the soil surface during 2023 and 2024 using frequency domain reflectometry. The models were trained to predict soil moisture at time horizons of 't + 1', 't + 3, 't + 6', and 't + 12' when 't' is present time. Short-term predictions more accurately followed the observed soil moisture trends than long-term predictions. During the collection period, rainfall varied from 0.5 mm to 48.5 mm per hour. Soil moisture contents increased immediately following the onset of rainfall. These results are consistent with existing knowledge. As soil depth increased, soil moisture contents tended to increase and respond more gradually to rainfall. During rainfall, in the topsoil (20 cm depth) moisture content fluctuated significantly, while the subsoil (60 cm depth) remained relatively stable for several hours after rainfall ended. The model performance was evaluated using mean absolute error, root mean square error (RMSE), and normalized RMSE values. In both models using hourly and event-based rainfall inputs, errors in soil moisture prediction tended to increase with longer forecast time horizons. However, these errors were significantly lower when using event-based rainfall data. These findings indicate that event-based rainfall is a more effective input for LSTM models in predicting soil moisture in an Asian pear orchard, and such models can support precision irrigation systems that optimize water use by delivering the right amount of water at the right time.