Abstract: Leaf diseases have threatened the yield and quality of apples, leading to economic losses. There is an ever-increasing demand to rapidly and accurately identify leaf disease. Among them, the commonly used RGB images are limited in capturing the full spectral information of the plant diseases, thus hindering accurate recognition. In contrast, hyperspectral imaging (HSI) can be expected to capture a broader range of wavelengths beyond visible light in the detection of plant diseases. The spectral information can also be detailed across the multiple bands. However, this high-dimensional spectral data can present information redundancy and computational complexity, which makes it difficult to efficiently process the hyperspectral images. In this study, a multi-head self-attention dynamic resnet (MHA-DResNet) model was proposed to recognize the apple leaf diseases. The multi-dimensional data of the hyperspectral images was fully utilized to mitigate the high dimensionality. A global semantic-guided feature selection mechanism (a multi-head attention mechanism) and dynamic convolution were integrated with the residual connections, in order to enhance the accuracy and efficiency of the disease detection. Firstly, a hyperspectral dataset was specifically constructed for the apple leaf diseases. A variety of hyperspectral images also corresponded to the different disease conditions on apple leaves. The dataset was then trained to evaluate the improved model. Secondly, the multi-head self-attention was introduced to select the most relevant spectral features from the hyperspectral data. Different regions of the image were focused on to learn the relationships among the spectral bands. The most relevant features were better distinguished between healthy and diseased leaves. The overall performance was improved after recognition. Moreover, the dynamic convolution and residual connections were integrated into the MHA-DResNet model. Dynamic convolution was used to adaptively capture the spatial features of the image. While the residual connections were more effectively trained to alleviate the vanishing gradient. As such, the local and global spatial features were extracted to accurately classify the diseases. The dynamic convolution was integrated with the self-attention mechanism in order to extract the spectral information from the spatial patterns. The more robust model was obtained for the disease detection in real-world applications. A series of experiments were performed on the hyperspectral dataset of the apple leaf diseases. The results showed that the MHA-DResNet achieved a significant accuracy of 96.18%. The recognition accuracy of the apple leaf disease was improved by 16.88% to 20.18%, compared with the traditional RGB-based method and the rest. The better performance of the improved model highlighted the promising potential of the hyperspectral data in disease identification. The self-attention mechanisms were also combined with the dynamic convolution. MHA-DResNet effectively extracted the global and local features for the accurate recognition of the disease-affected areas. The complex types of the diseases were also treated to remove the nonlinear relationships among the spectral bands. The improved model provided a promising solution to the accurate recognition of the apple leaf diseases. Some insights were offered for the detection of agricultural diseases. The MHA-DResNet was also combined with the hyperspectral data, self-attention mechanisms, and dynamic convolution. The finding can greatly contribute to the more efficient and precise recognition of plant diseases in modern agriculture, particularly in early disease warning and rapid diagnosis.