Accurate and nondestructive detection of cotton aphids is crucial for effective pest control and enhancing cotton yield and quality. Aiming to propose a multi-feature fusion method for cotton aphid damage level (CADL) monitoring, spectral feature wavelengths, vegetation indices, and cotton canopy texture characteristics were integrated to enhance the accuracy of cotton aphid damage level determination. A UAV-mounted hyperspectral imaging system was employed to collect hyperspectral image data of cotton canopy. Pre-processing of the extracted spectral data involved Savitzky-Golay smoothing (SG smoothing) and multiple scattering correction (MSC). Support vector machine (SVM) modeling was applied to the pre-processed spectral data, results revealed that MSC performed better than SG smoothing in pre-processing. Thus the spectral data pre-processed by MSC was used for characteristic wavelengths extraction. Characteristic wavelengths extraction was conducted by using the competitive adaptive reweighting algorithm (CARS) and the shuffled frog leaping algorithm (SFLA), totally 31 and 37 characteristic wavelengths were extracted by CARS and SFLA, respectively. Subsequently, the successive projections algorithm (SPA) was utilized for secondary characteristic wavelengths extraction. Ultimately，six sensitive wavelengths at wavelengths of 650nm, 786nm, 931nm, 938nm, 945nm and 961nm were extracted. Based on six secondarily extracted characteristic wavelengths, nine vegetation indices and eight texture features were calculated, followed by correlation analysis between these vegetation indices/texture features and CADL. Four machine learning models (LightGBM, XGBoost, SVM, RF) were developed to evaluate the classification performance by using characteristic wavelengths alone, vegetation indices alone, texture features alone, combined characteristic wavelengths and vegetation indices, and integrated characteristic wavelengths, vegetation indices, and texture features. Results indicated that vegetation indices (RDVI, SAVI, MSAVI, OSAVI) and texture features (MEA, VAR, DIS, HOM) exhibited strong correlations with CADL. The XGBoost model incorporating the tri-feature combination (characteristic wavelengths, vegetation indices, texture features) achieved optimal CADL classification performance, yielding an overall accuracy (OA) of 86.99% and a Kappa coefficient of 0.8371 on the test set. Compared with models by using characteristic wavelengths alone, vegetation indices alone, texture features alone, or the dual-feature combination (characteristic wavelengths, vegetation indices), this integrated approach improved OA by 4.88, 27.64, 21.95, and 2.44 percentage points, respectively.