In order to study the prediction of crop water stress index (CWSI) of tomato canopy in greenhouse, through the deployment of multi parameter sensors, the environmental parameter inside and outside the greenhouse can be obtained in real time. Using gray correlation analysis, the correlation degree between environmental parameters and tomato canopy CWSI and the sub factor correlation coefficient between environmental parameters was calculated, the environmental parameters were sorted according to the correlation degree, and the impact on the accuracy of the model was considered. Finally, a total of seven parameters from nine environmental parameters were selected as the model input, and a prediction model of greenhouse tomato canopy crop water stress index (CWSI) based on LightGBM was established. Combined with Bayesian algorithm to optimize the key parameters, the correlation between the prediction results of the model and the CWSI value calculated by Jones empirical formula was analyzed. Under the same computing environment, it was compared with GBRT and SVR models respectively. The experimental results showed that the coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE) and operation time of the Bayesian optimized LightGBM model were 0.9601, 0.0218, 0.0314 and 0.0518s, respectively. Compared with GBRT and SVR models, R2 was increased by 2.14% and 14.05% respectively, MAE was reduced by 0.0093 and 0.0612 respectively, RMSE was reduced by 0.0097 and 0.0591 respectively, and the time was shortened by 0.0459s and 0.0612s respectively. It was showed that the LightGBM model proposed had better performance, which could effectively improve the prediction accuracy of greenhouse tomato canopy CWSI, and provide a strategy for realizing greenhouse tomato on-demand irrigation and a reference for water requirement research.