Sheath blight is one of the main diseases of rice, whose control is of great significance to ensure rice yield and quality. Hyperspectral detection of rice diseases has been widely adopted in recent years, and hyperspectral dimensionality reduction is an important part of spectral analysis. In this study, the hyperspectral data of low altitude remote sensing canopy and rice ground canopy were obtained in Shenyang Agricultural University rice proving ground in 2019, and were smoothed by Savitzky-Golay with a window width of 15 and order of 3, as well as spectral transformations (original reflection spectrum, first-order differential reflection spectrum and inverse-log reflection spectrum), were carried out. To reduce the dimension of hyperspectral data in these 3 spectra, the split-window Gram-Schmidt transform method was used to find the projection space and map the main substrate, in which the main base with significant probability was drawn, and its maximum and minimum value was the characteristic band. The principal component analysis and successive projections algorithm were also used for dimensionality reduction of three spectra. Dimension-reduced data and rice sheath blight disease index were modeled by support vector machine regression, which was used for particle swarm optimization and radial basis function as the kernel function. The effect of three-dimensionality reduction methods was compared and analyzed. The results showed that the modeling effect of the rice ground canopy scale was better than that of the low-altitude remote sensing scale; in the aspect of hyperspectral data processing, the inverse logarithm transformation effect of low-altitude canopy hyperspectral data was better, and the first-order differential transformation effect of ground canopy hyperspectral data was better; the split-window Gram-Schmidt transformation algorithm was better than principal component analysis and successive projections algorithm; particle swarm optimization could optimize the penalty coefficient and kernel function parameters in SVR, and improve the inversion accuracy; in the low-altitude remote sensing canopy scale, the hyperspectral spectrum was processed by using the inverse logarithm processing and the split-window Gram-Schmidt transform, whose sensitive bands were 427.3, 539.6, 749.5 and 825.4 nm respectively. The determination coefficient R² was 0.731 and RMSE was 0.151 by using the PSO-SVR model; in the ground canopy scale, the hyperspectral spectrum was processed by using the first order differential processing and the split-window Gram-Schmidt transform, whose sensitive bands were 552, 607, 702 and 730 nm respectively. The determination coefficient R² was 0.778 and RMSE was 0.147 by using the PSO-SVR model. In conclusion, rice sheath blight can be effectively detected by hyperspectral technology, and its disease index can be retrieved by canopy hyperspectral analysis. The split-window Gram-Schmidt transform has a good effect on the dimensionality reduction of hyperspectral data. PSO-SVR modeling can significantly improve the inversion of rice sheath blight disease index. The results can provide a theoretical basis and technical support for the detection of rice sheath blight and disease occurrence on the canopy scale.