Abstract: This study aimed to improve the accuracy, reliability, and risk control of irrigation decision-making for greenhouse tomatoes by integrating crop water status sensing with a prover-verifier (P-V) constrained framework. A multispectral leaf water content (LWC) inversion model was developed to provide nondestructive crop water information, and the estimated LWC was further integrated with soil moisture, meteorological conditions, and irrigation-management information to support irrigation decision-making. Multispectral imagery was acquired using an unmanned aerial vehicle (UAV), and canopy spectral reflectance over the whole growth period was extracted. Twenty-four spectral indices were screened using the Pearson correlation coefficient (PCC), and nine machine-learning algorithms were compared to establish an optimal LWC inversion model. A multilayer perceptron (MLP) served as the prover to generate irrigation decisions, while logistic regression served as the verifier to evaluate decision rationality and constrain unreasonable irrigation outputs. The results showed that partial least squares regression (PLSR) achieved the best performance among the nine LWC inversion models, with a validation coefficient of determination of 0.828. The selected multispectral features therefore provided reliable information for characterizing tomato water status and supporting subsequent irrigation decisions. At the irrigation decision level, the P-V framework improved the balance between prediction accuracy and risk control compared with the traditional threshold-based model and the prover-only model. When the adversarial weight was set to 0.12, the comprehensive decision risk decreased from 12.12% in the baseline model to 6.15%. The over-irrigation and under-irrigation error rates were controlled at 5.26% and 7.04%, respectively, indicating that the verifier constraint effectively reduced unreasonable irrigation decisions. The P-V framework therefore provided an additional rationality-checking mechanism beyond direct prediction and improved the robustness of irrigation outputs. Multiple random-seed experiments further showed that the model maintained relatively stable performance under different initialization conditions, with an average classification accuracy of 91.78% and a mean absolute error (MAE) of 1.95 mm for irrigation amount prediction. Practical validation during the second growing season further confirmed the applicability of the proposed framework. Compared with traditional scheduled fixed-amount irrigation, the P-V model reduced total irrigation by 9.7% and increased irrigation water use efficiency (IWUE) by 11.1%. These results indicated that the improvement in model-level risk control could be translated into measurable water-saving benefits under actual greenhouse irrigation conditions. The results demonstrated that the combination of UAV multispectral sensing, LWC inversion, multidimensional water-status information, and verifier-based constraints effectively improved the robustness and reliability of irrigation decision-making for greenhouse tomatoes. The proposed P-V framework reduced irrigation decision risk while maintaining stable prediction performance and practical water-saving effects, providing a feasible decision-support approach for precision irrigation management in greenhouse agriculture.