Improving water status prediction of winter wheat using multi-source data with machine learning

Water and nitrogen (N) are the most important factors limiting crop productivity. Effective monitoring of the water status of winter wheat under different N treatments is imperative for precision irrigation in improved crop management. Hyperspectral remote sensing is widely used for monitoring the crop water status. However, changes in canopy architecture during ontogeny lead to poor inversion of crop properties and limit the use of spectral indices. This study aimed to improve the water status prediction of winter wheat using multi-source data with machine learning (ML). Two multi-irrigation levels (0, 120, 240, 360 mm) and N rates (75, 225 kg N ha⁻¹) experiments were conducted during the 2019–2021 wheat growing seasons under field conditions using a rainout shelter. Hyperspectral, soil, plant, and climate data were evaluated with two feature selection methods. Selected results were chosen as input variables for prediction models by using three ML algorithms. By constructing the normalized difference spectral index (NDSI), ratio spectral index (RSI), and difference spectral index (DSI), the DSI₍₂₀₁₅, ₂₃₇₅₎, NDSI₍₂₁₇₅, ₂₂₄₅₎, and RSI₍₇₂₀, ₁₂₀₀₎ showed the strongest correlation with canopy water content (CWC), plant water content (PWC), and canopy equivalent water thickness (CEWT), respectively. The best feature selection method and data were delivered by the Pearson correlation coefficient together with the soil, plant, and climate data. The best performing ML algorithm for CWC and PWC prediction was RF, while SVM was the best ML algorithm for CEWT prediction. The R² of the optimal models ranged from 0.86 to 0.97. These models with multi-source data significantly improved the prediction accuracy of the water status and can thus assist in precision irrigation management of winter wheat.