Soil water content prediction across seasons using random forest based on precipitation-related data

• Hourly Soil Water Content (SWC) is reasonably predicted by random forest based on past hourly precipitation. • Consistent cumulative rainfall is identified as the input variable. • Prediction errors are minor except for the extreme SWC observations in the rainy season. Predicting the soil water content (SWC) is crucial to prepare for and mitigate risks during dry periods, particularly before droughts. It also ensures effective water management and precise irrigation of agricultural land. Few studies have focused solely on the use of precipitation data to predict SWC. This study aimed to predict the hourly SWC for one or two days in advance at depths of 10 and 20 cm below the surface of agricultural land in Taichung, Taiwan, using Random Forest (RF), which has demonstrated promising results in previous studies. The model used hourly precipitation data from January 19, 2022, to April 18, 2023. Seven sets of strategically selected cumulative rainfall days were incorporated to balance the computational load and prediction accuracy. Based on the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) curves, a cumulative rainfall of 6–8 days was optimal for RF prediction at 10 and 20 cm SWC. The RF model demonstrated reasonable performance, with MAE of 0.6 % and 1.0 %, R 2 values of 0.5 and 0.9, MAPE of 25.2 % and 5.1 %, and RMSE of 2.4 % and 2.0 % for the 10 and 20 cm SWC predictions, respectively. The RF model performed well during dry periods but showed less accuracy during the rainy season, as indicated by the MAPE for the entire period, compared with the rainy season alone. This discrepancy may be due to the unusually high SWC in response to storm events. In conclusion, this study provides insights for improving SWC prediction accuracy across different seasons and practical guidelines for employing RF models in SWC prediction.