Feature extraction combined with data generation powers the grading of cotton Verticillium wilt under small-sample condition

Rapid and accurate grading of cotton Verticillium wilt (CVW) is critical for the cotton industry. However, it is difficult to collect sufficient diseased samples in practical scenarios, resulting in the problem of small sample. This study focus on exploring the potential of feature extraction and data generation to improve the performance of CVW grading under the small-sample condition based on hyperspectral imaging. 207 leaves with different degrees of infection were utilized and divided into five categories. Four types of features, including full wavelengths, feature wavelengths, color features, and texture features, were extracted respectively. Then, a DCNN integrating an attention mechanism was designed, and the effectiveness of different types of features was evaluated. Obtaining the highest accuracy of 94.29%, the DCNN based on the feature wavelengths outperformed DCNNs based on other features and conventional SVC models. Feature fusion was further investigated for improving model performance, which was proven to be effective to a certain extent. Considering the small-sample condition, CGAN was introduced for data generation. The generated spectra were evaluated for their similarity to the real data. Then, new grading models were constructed through gradually adding the generated spectra to the original training set. The CGAN-SPA-SVC achieved superior performance, with an accuracy of 97.14%. The overall results indicated the effectiveness of the combination of feature extraction with data generation in the grading of CVW under small-sample conditions, which would be conducive to the timely and accurate monitoring and assessment of CVW in actual applications.