Abstract: Cotton is one of the most important cash crops for regional food security in Xinjiang of western China. However, cotton is also susceptible to various diseases. Among them, Verticillium Wilt (VW) has caused severe economic losses to the cotton industry. Conventional manual monitoring cannot fully meet the demands of rapid and accurate detection, due to being time-consuming, labor-intensive, and prone to bias. Hyperspectral Imaging (HSI) can be used for rapid detection of plant diseases to acquire spectral information from plant leaves and canopies. It is still lacking in the refined analytical and discriminative criteria for the severity grading of cotton Verticillium wilt. In this study, a Cotton Verticillium Wilt Index (CVWI) was proposed to accurately assess the disease severity using hyperspectral technology and deep learning. A deep convolutional neural network, designated as ResNet-NL, was first constructed to integrate a non-local module into a residual network (ResNet) architecture. Disease features were then characterized at different severity levels. Subsequently, the significance score of each wavelength was calculated to identify the feature wavelengths most sensitive to the disease response, according to the network’s backpropagation. Finally, the CVWI was formulated to combine the optimal spectral features. A comparative analysis was conducted on the classification performance of various machine learning and deep learning models (ResNet and ResNet-NL). Four types of feature selection were used: successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), ReliefF, and random forest (RF). The results demonstrated that the CARS with extreme learning machine (ELM) achieved a superior classification accuracy of 84.6% in the feature selection. The ResNet and ResNet-NL models attained classification accuracies of 87.1% and 89.1%, respectively, where the ResNet-NL model exhibited optimal performance. The non-local module was effectively incorporated to capture the spectra sensitive to Verticillium wilt. Feature wavelengths were selected by the ResNet-NL-Saliency, indicating favorable intra-class aggregation and inter-class separability. Jeffries-Matusita (JM) distances exceeded 1.95, with Within-class scatter matrix (S_w) values below 0.2. Compared with classical spectral indices, the CVWInl shared a stronger correlation with the severity of cotton Verticillium wilt. A more refined assessment of disease severity was realized with the correlation coefficient of -0.888, indicating the superior anti-noise performance. The findings can provide a solution to overcome the existing grading standards for plant diseases. Significant theoretical value and application prospects were offered for the precise monitoring and early prevention and control of cotton Verticillium wilt.