Abstract: Abstract: An accurate and rapid detection has been one of the most important steps to monitor the drought stress status of cotton at flowering and boll-forming stage in the cotton yield and quality. This study aims to realize the rapid detection of key drought resistance indexes of cotton. 253 cotton resources were taken as the raw materials in this experiment. Two treatments were set for the drought stress and normal irrigation. The images were also obtained by the DJI Jingling 4 multispectral UAV at the flowering and boll-forming stage. A systematic analysis was made to determine the spectral reflectance of drought stress and normal treatment cotton at the flowering and boll-forming stage. The predicted model was established to combined with the relative value of chlorophyll (SPAD) and leaf water content (LWC) data of ground survey using the radial basis function in the neural network algorithm. The results showed that the red-light reflection value of cotton increased slightly before and after drought stress, whereas, the other four spectra were less than normal treatment. The best estimation models were cubic function to predict the SPAD by radial basis function. The best estimation models of LWC were quadratic function. The ideal model of quadratic function SPAD was y=0.010 6x2-0.673x+61.045, where determination coefficient R2 was 0.848 8, root mean square error RMSE was 2.005 and relative error RE was 0.004. The ideal model of LWC was y=-0.017 6x2+0.709 7x+3.249 3, where R2 was 0.936 6, RMSE was 0.930, and RE was 0.011. The improved model was also applied to discriminate the drought image in the experimental area. The SPAD and LWC prediction values were clustered and analyzed during this time. The drought classification based on SPAD clustering results was that extreme drought, severe drought, moderate drought, light drought and no drought, and the drought classification based on LWC clustering results was that extreme drought, severe drought, moderate drought, drought, light drought and no drought. The obtained models and classification effects were better conform to the characteristics of cotton drought stress. The finding can provide a strong reference for the rapid and accurate acquisition during cotton drought monitoring.