Abstract: In order to improve the accuracy of hyperspectral prediction model for soil texture, the soil texture spectral prediction model was conducted based on the 177 soil samples taken from Chaohu Lake Basin using the methods of the ant colony optimization to select the characteristic wavelengths, and combined with BP neural network. The results showed that, compared with the spectral prediction model constructed by full spectrum, the accuracy of the spectral prediction model constructed in this study was better. The determination coefficient R² of the prediction set of the soil slit content prediction model was 0.76 and RPIQ was 2.23, that of the soil sand content prediction model was 0.72 and RPIQ was 1.94. The prediction set R² and RPIQ of the full spectra soil slit content prediction model were 0.57 and 1.75, respectively, and that of the full spectra soil sand content prediction model were 0.48 and 1.82, respectively. Compared with the full spectrum prediction model of soil texture, the ant colony optimization selected the spectral characteristic wavelengths for modeling could reduce the redundancy of spectral information and improve the accuracy of the prediction model.