Timely monitoring of cotton canopy physiological status and the identification of growth stages sensitive to irrigation are crucial for optimizing irrigation management. This study aimed to utilize UAV multispectral data to identify growth stages associated with cotton yield variation and to evaluate the response of SPAD values to different lower-limit irrigation thresholds. UAV multispectral imagery and ground-truth data were collected during key growth stages. A Random Forest (RF) yield model combined with SHAP analysis was employed to investigate the relative contributions of traits from different growth stages to yield variation. Simultaneously, XGBoost, Random Forest Regression, 1D-CNN, and CNN-BiLSTM-Attention models were constructed and compared for SPAD estimation based on UAV data. Furthermore, crop water consumption was calculated using the soil water balance method, and water use efficiency (WUE) and irrigation water productivity (IWP) were determined to further evaluate the effectiveness of water management. The RF yield model demonstrated moderate predictive performance, with a coefficient of determination (R²) of 0.40, a root mean square error (RMSE) of 771.96 kg ha⁻¹, and a mean absolute error (MAE) of 671.31 kg ha⁻¹. SHAP analysis revealed that traits measured during the flowering and boll-setting stage contributed more significantly to yield variation than those from other stages. Regarding SPAD estimation, the CNN-BiLSTM-Attention model performed best, achieving an R² of 0.89, an RMSE of 0.03, and a residual prediction deviation (RPD) of 3.08. Raising the lower-limit irrigation threshold during the flowering and boll-setting stage from 50% field capacity (FC) to 60% FC significantly improved root-zone soil moisture status, SPAD values, yield, and IWP. Further increasing the threshold to 70% FC was more beneficial for enhancing yield and WUE but did not lead to further improvements in IWP. These results indicate that UAV multispectral data can support the monitoring of cotton SPAD and irrigation threshold responses. Under the present field conditions, a flowering-boll lower irrigation threshold range of 60–70% FC may provide a practical reference for balancing canopy physiological status, yield formation, and water-management performance, with 60% FC being more favorable for IWP and 70% FC being more favorable for yield and WUE.