Adzuki bean rust is a common fungal disease that significantly reduces crop yield by infecting leaves and impairing photosynthesis. This paper proposes a novel diagnostic method for adzuki bean rust based on an improved Empirical Mode Decomposition-Wavelet Transform (E-DWT) algorithm and a deep learning model. Using “Baoqinghong” adzuki beans as the experimental material, 960 leaf spectral datasets were collected over 10 days using a handheld visible/near-infrared spectrometer, obtaining reflectance data in the wavelength range of 326~1 075 nm. First, the improved E-DWT algorithm was applied for spectral denoising. This algorithm combines Empirical Mode Decomposition (EMD) and wavelet threshold denoising technology to retain effective signal information while removing noise. The optimal wavelet basis function (sym5) and the number of decomposition layers (4 layers) were determined by comparing the RMSE and SNR indicators. To further reduce redundancy in high-dimensional data, the Successive Projections Algorithm (SPA) was employed to select 12 representative wavelengths from the initial 750 features, resulting in a 98.4% reduction in feature wavelength count. Next, the Gramian Angular Field (GAF) method was employed to convert the one-dimensional wavelength sequence into a two-dimensional spectral image, thereby enhancing correlations between different bands for subsequent model training. The designed deep learning model combines a Convolutional Neural Network (CNN) with a Convolutional Block Attention Module (CBAM). The CBAM module effectively discriminates the weights of different feature wavelengths and time nodes through channel and spatial attention mechanisms, enabling the model to focus on key features influencing adzuki bean rust identification. Experimental results show that the CBAM-CNN model achieved 99.31% accuracy in the training set, 98.33% accuracy in the test set, and 98.89% recall, significantly outperforming traditional CNN models. Compared to existing methods, the proposed model exhibits superior performance in terms of recognition accuracy, stability, and training convergence speed. Additionally, the model structure is more concise, which optimizes parameter adjustment and improves operability in practical applications. In conclusion, the proposed diagnostic method based on the improved E-DWT algorithm and CBAM-CNN model not only achieves efficient and precise disease detection but also provides a theoretical foundation and technical support for constructing data-driven crop disease diagnosis systems in the future.