Objective  Brassica chinensis var. parachinensis was used as the experimental object to study the change rule of water stress during growth process based on canopy temperature, and the machine learning method was used for predicting the photosynthetic rate based on crop water stress index (CWSI) and photosynthetically active radiation.

Method  During the test, the experiment adopted different field capacities for B. chinensis var. parachinensis at the vegetative growth stage (V stage) and reproductive growth stage (R stage), collected the canopy temperature, air temperature and humidity data, established non-transpiration-baseline (upper limit equation), non-water-stress-baseline (lower limit equation), and calculated CWSI by empirical formulation. Cluster method of density-based spatial clustering of application with noise and air temperature were used to study the upper limit distribution of canopy temperature minus air temperature of B. chinensis var. parachinensis, and the fixed values were selected as the upper limit. Based on the CWSI empirical formulation, CWSI was calculated using the non-transpiration-baselines with different temperature fixed values to verify the clustering effect. In order to obtain the photosynthetic rate more easily, four machine learning methods of k-nearest neighbor (KNN), support vector regression (SVR), extreme gradient boosting (XGBoost) and random forest (RF) were used for prediction, and the prediction effects were compared.

Result  Under different field capacities, CWSI could better monitor the water stress status of B. chinensis var. parachinensis. Through cluster analysis, the upper limit of canopy temperature minus air temperature at V stage and R stage was classified into two clusters, and the cluster centers were 3.4 and 4.2 ℃, respectively, which were significantly correlated with the values calculated by the empirical formula of CWSI, indicating that it was feasible to use a fixed value as the upper limit of canopy temperature minus air temperature in B. chinensis var. parachinensis. The prediction models of KNN, SVM, XGBoost and RF all achieved good results, and the correlation coefficients were 0.873, 0.877, 0.887 and 0.863, respectively.

Conclusion  Machine learning can be used for predicting the photosynthetic rate of B. chinensis var. parachinensis, avoid the use of large and cumbersome instruments, reduce the damage to the leaves of B. chinensis var. parachinensis, and reduce the measurement time.