Improving maize water stress diagnosis accuracy by integrating multimodal UAVs data and leaf area index inversion model

Unmanned aerial vehicles (UAVs) allow field monitoring of crop water status using multimodal UAV data. However, the accuracy and evaluation of predictive models in large-scale applications remain challenging issues. We conducted a two-year field experiment on spring maize in the Hetao Irrigation District, Inner Mongolia, using plant water content (PMC) and normalized stomatal conductance (NGS) as indicators of crop water status. UAV-data-derived multi-spectral indices (MIs) were used to construct a leaf area index (LAI) inversion-model. Crop water-stress diagnostic models were developed by integrating UAV multi-spectral data, thermal infrared data, and LAI inversion results using machine-learning algorithms. The best-performing models were applied in a large-scale experiment to evaluate their transferability. MIs and LAI showed higher correlations with PMC and NGS during early and mid-growth stages (V9 and VT), whereas temperature indices (TIs) strongly correlated with NGS during the peak growth stages (VT and R1). Although models built using the random forest regression (RFR) algorithm and the combination of MIs+TIs+LAI performed best (R2 ≥ 0.575, RMSE ≤ 0.073, and RRMSE ≤ 0.18) across growth stages, their predictive advantages for PMC and NGS varied with the growth stage: PMC predictions were more accurate during stages V9 and R3, whereas NGS predictions were more accurate during stages VT and R1. In the AE scene, the prediction accuracy for both PMC and NGS decreased, although NGS predictions outperformed PMC predictions, with R2, RMSE, and RRMSE values of 0.5836, 0.2301, and 0.3437, respectively. Nonetheless, the models underestimated NGS results. Our study provides practical recommendations and insights for diagnosing plant water stress based on multimodal UAV data.