Accurate prediction of crop water diagnostic indicators is crucial for improving crop water productivity; however, the applicability of water stress indices for diagnosing crop water status requires study. We analyzed the accuracy of machine learning (ML) algorithms in predicting the Crop Water Stress Index (CWSI) and transpiration-based Plant Water Deficit Index (PWDI). Their applicability across the growth cycle and threshold values was determined. Using field experiment data (n = 2 years) from cotton under mulch drip irrigation in Xinjiang, we evaluated the predictive performance of four ML algorithms for the cotton CWSI and PWDI. Using XGBoost's variable importance analysis and the random forest (RF) two-dimensional partial dependence analysis, we determined responses of indices to key factors and their applicability. Tree-based models, specifically RF and XGBoost, performed better than other models. The best CWSI predictors were midday air temperature (T_a-13:00), leaf temperature (T_c), and vapor pressure deficit (VPD). Key variables PWDI factors included stem sap flow rate (V), transpiration rate (T_r), and solar radiation (SR). CWSI was most accurately predicted when T_a-13:00 < 27°C and T_c < 27°C, whereas for PWDI, optimal T_a-13:00 > 27°C, root-zone soil water content (SWC-RZ) < 0.175 cm³ cm⁻³ , and SR < 400 W m⁻². Threshold analyses suggested CWSI values for the squaring and boll stages of 0.79 and 0.80, respectively, but 0.65 for PWDI across stages. Using the indices, critical insights were determined for developing and implementing irrigation strategies to optimize crop water-use efficiency and precise water management in agricultural systems.