Identification of key growth stages for cotton yield formation and SPAD response to irrigation thresholds using UAV multispectral data

This study aimed to identify the critical growth stages governing cotton yield formation and to characterize SPAD responses to irrigation thresholds under precision irrigation conditions. Field irrigation experiments, unmanned aerial vehicle (UAV) multispectral remote sensing, and machine learning methods were integrated to analyze cotton yield, canopy SPAD, and irrigation responses. Multispectral images acquired across six growth stages, together with synchronous ground observations, were used to develop a plot-level yield prediction model and a canopy SPAD estimation model. SHAP analysis was further employed to identify key growth stages and influential indicators associated with yield formation. The results showed that the Random Forest model achieved moderate predictive performance for plot-scale cotton yield, with out-of-sample values of R² = 0.40, RMSE = 771.96 kg ha⁻¹, and MAE = 671.31 kg ha⁻¹. SHAP analysis indicated that variables with high explanatory importance were mainly concentrated during the flowering and boll-forming stage, suggesting that this stage plays a key role in cotton yield formation. Among the evaluated SPAD estimation models, the CNN-BiLSTM-Attention model performed best, with validation metrics of R² = 0.89, RMSE = 0.03, and RPD = 3.08. Irrigation response analysis further showed that SPAD during the flowering and boll-forming stage was most sensitive to changes in the lower irrigation threshold. Increasing the threshold from 50% to 60% of field capacity (FC) significantly increased predicted SPAD values, whereas a further increase to 70% FC resulted in only limited gains. Under the conditions of this study, 60% FC appeared to be an effective threshold for maintaining relatively high SPAD levels during the flowering and boll-forming stage. These findings suggest that SPAD can serve as a useful process-based indicator linking irrigation regulation with canopy physiological status and yield formation, thereby providing a scientific basis for optimizing precision irrigation thresholds in cotton production.