Abstract:

 This study utilized microclimate data from high−standard farmlands during wheat growing season (October to May) from 2021 to 2023. By investigating the lagged response of soil relative humidity (SRH) to microclimate factors, this study developed three machine learning models, Random Forest (RF), Backpropagation Neural Network (BPNN) and Support vector regression (SVR), using the Optuna framework for hyperparameter optimization. The models predicted SRH at three forecasting horizons (3−, 5− and 10−days) across five soil depths (10cm, 20cm, 30cm, 40cm and 50cm) to establish a predictive reference system for high−standard farmland. The results indicated that: (1) SRH exhibited a fluctuating decrease throughout winter wheat growth stages, with maximum values (90.4%) during sowing to emergence and minimum values (73.9%) at anthesis to maturity stage. (2) The response characteristics of SRH to microclimate factors varied significantly. SRH demonstrated the strongest yet slowest response to ground temperatures (r=0.32–0.57; 5–10d lag), and the weakest yet fastest response to air relative humidity (r<0.20; 1–3d lag). As soil depth increased, the correlation between SRH and precipitation, daily mean air temperature and daily maximum temperatures decreased, whereas correlations with maximum daily wind speed and soil temperatures (10cm, 20cm and 50cm depths) increased gradually. (3) Among the three simulation models, the RF model achieved superior performance across all prediction horizons (R²=0.87−0.98, RMSE=0.02−0.05, MAE=0.01−0.03), significantly outperforming SVR (R²=0.77−0.97, RMSE=0.03−0.07, MAE=0.02−0.04) and BPNN (R²=0.60−0.97, RMSE=0.04−0.07, MAE=0.01−0.06). A comprehensive evaluation showed that the RF model was better suited for short−term predictions of soil moisture in high−standard farmland, providing valuable technical support for precise water management in Henan. 

Key words: High?standard farmland, Microclimate factor, Machine learning, Soil relative humidity