To determine the optimal classification model for unmanned aerial hyperspectral remote sensing for detection of leaf blast in rice canopies, the research is based on rice field trials, hyperspectral images of unmanned aerial vehicles (UAVs) in the 400~1 000 nm band were acquired, referring to the national standard GBT 15790-2009 specification for rice blast detection and survey, leaf blast is categorized into five classes according to the disease index, a total of 227 hyperspectral data sets were extracted for levels 0 to 4. The data were preprocessed using Savitzky- Golay smoothing (SG smoothing), first-order differential spectroscopy (1-Der) and second-order differential spectroscopy (2-Der) methods, and SVM models are constructed and compared to arrive at a better preprocessing method. Principal component analysis (PCA) was used to select the cumulative contribution of the principal components, continuous projection (SPA) and random frog (RF) methods for screening spectral signature bands, and using the results of the screening as inputs to the model, constructing Particle Swarm Optimization for Extreme Learning Machines (PSO-ELM), Extreme Learning Machines (ELM), Support Vector Machines (SVM) and Decision Tree (DT) classification models, respectively. The results show that compared with 1-Der and 2-Der, the SG smoothing method has better denoising effect, higher classification accuracy, and is a better preprocessing method, the classification accuracy and Kappa coefficient were 93. 47% and 91. 85%, respectively. The cumulative contribution of the first 2 PCs of the PCA was 93. 13%, and for the effective construction of the model, the first 6 PCs were finally selected with a cumulative contribution of 99. 02%,SPA used RMSE as the criterion for the selection of the best spectral signature bands, showing a total of seven best spectral signature bands, The visible wavelength bands are 400. 8, 416. 7 and 426. 2 nm, the green wavelength band is 566nm, the red wavelength band is 683. 9nm, and the near-infrared wavelength bands are 830. 2 and 916. 4 nm, RF selected the bands with a screening probability greater than 0. 2 as the best spectral signature bands, and finally screened eight spectral signature bands, including 663. 4 and 694. 2 nm for red light, and 784. 4, 787. 9, 791. 4, 905. 5, 927. 2, and 930. 9 nm for the near-infrared band, this method effectively reduces the inter-band correlation and redundancy, while the three screening results are constructed into classification models separately, and the results show that the overall classification accuracies of all models are all greater than 92. 61%, the modeling results were better, in which the PSO-ELM model was used to classify PCA with an accuracy of 97. 77% and a Kappa coefficient of 97. 22%, the highest classification accuracy among all models, 1. 42% and 1. 56% higher compared to the highest classification accuracy and Kappa coefficient of the ELM model, the highest classification accuracy and Kappa coefficient are 2. 12% and 2. 66% higher compared to SVM model and 4. 44% and 5. 58% higher compared to DT model. The comprehensive evaluation of PSO-ELM model modeling is better than ELM model, SVM model and DT model, which is the optimal classification model. Therefore, it is feasible to use UAV hyperspectral remote sensing to detect rice leaf blast, providing scientific basis and technical support for rice production and leaf blast detection.