【Objective】 To solve the problems of difficult selections of high-dimensional variables and low accuracy of random forest algorithm in tree water prediction model, this paper studies the fusion algorithm based on recursive feature elimination (RFE) and random forest (RF), constructs the estimation model of visible light images and leaf water content of young A. sinensis, explores the water conditions suitable for the growth of young Aquilaria sinensis, and provides a feasible reference method for realizing the nondestructive monitoring of water deficit degree of A. sinensis seedlings. 【Method】 Taking two-years-old precious tree species of A. sinensis as the research object, the visible light images of young A. sinensis under four different water gradients were obtained by camera, 15 image features were extracted, the recursive feature elimination method was used to screen the optimal image feature subset of agallochum leaves, and then the prediction model of water content of A. sinensis leaves was constructed combined with RF algorithm. Finally, the RFE_RF model is compared with the conventional RF and least square support vector machine (LSSVM) to test the feasibility of the model. Three features of standard red light value (I_NR), saturation (S) and rectangularity (E_R) of young A. sinensis leaf images were screened out by recursive feature elimination and random forest fusion (RFE_RF) algorithm, and used as the model independent variables. 【Result】 The results showed that compared with severe flooding stress, young A. sinensis was more sensitive to long-term severe drought stress, and the drought time exceeded two weeks, the leaves of the seedlings were severely damaged, threatening the growth of A. sinensis; the optimum leaf water growth range of A. sinensis was 50%-65%, moderately increase moisture, which is conducive to the growth of A. sinensis. The false alarm rate and accuracy of the prediction model constructed based on the RFE_RF fusion algorithm reached 88.64%, 85.31%, 14.39% and 91.62%, respectively, which were better than the LSSVM model; the sensitivity is increased by 3.34%, the specificity is increased by 10.87%, the false alarm rate is decreased by 36.83%, and the precision is increased by 13.39%. 【Conclusion】 The model of Aquilaria sinensis leaf color, shape characteristics and water content established by RFE_RF solves the problem of high-dimensional variable selections in the random forest process, improves the accuracy of RF in the forest moisture prediction model, and realizes the analysis of leaf water content of Aquilaria sinensis seedlings. The non-destructive estimation and diagnosis provide new ideas for accurate water management and management of precious tree species.