In order to reduce the influence of dust retention on the extraction of effective spectral information of tea leaf and to establish a more robust water content estimation model of tea leaf by spectrum. We took "Shu Chazao" as the research object and collected samples of tea leaves by random sampling. Then the hyperspectral information, leaf water content and dust retention rate of leaves were measured. The correlation coefficient method was used to extract feature information. Newly-built vegetation indexes were constructed by the normalization calculation method and ratio calculation method, The relative variability analysis was used to screen the candidate indexes that reduce the impact of dust retention on the accuracy of the leaf water content estimation model. By comparing the response relationship between newly-built vegetation indexes and existing water indexes under the different conditions of dust retention, the optimal vegetation index estimation model of tea leaf water content which less affected by dust retention, was selected. Finally, the high-precision estimation models of the tea leaf water content with the optimal vegetation index were established and verified. The results show that, dust leaves' spectral reflectance is higher than clean leaves in 711-1 378 nm bands. The correlation between the water content of the tea leaves and vegetation index is affected by dust retention, but its correlation direction is not. Dust retention also makes the accuracy value of tea leaf water content estimation model decreased. The newly-built ratio index (RVI(1 298, 1 325)) with 1 298 and 1 325 nm as the center band is least affected by leaf dust retention under complex environmental conditions. Therefore, it is the optimal vegetation index, and the hyperspectral estimation model of tea leaf water content constructed by RVI(1 298, 1 325) has higher estimation accuracy, better sensitivity and stability (y= 0.245x - 0.241, R-2 = 0.854, RMSE= 0.001). In conclusion, this study provides a basis for the refined water management of tea trees and provides new ideas that high-precision models of water content estimation is constructed by hyperspectral information under complex environmental conditions.
