MDHANet: Rethinking HOG as a Prior in Dense Networks with Self-Attention for Hyperspectral Image Classification

Convolutional neural network (CNN)-based methods have been widely adopted in hyperspectral image (HSI) classification tasks and have demonstrated superior performance due to their potent nonlinear fitting capabilities. However, as network depth increases, existing methods suffer from significant issues such as gradient vanishing and information loss, which limit further improvements in model performance. Meanwhile, current approaches primarily design entirely new network architectures to extract abstract features but fail to fully exploit the explicit gradient prior information inherent in the data, which affects model performance in complex scenarios. To address these issues, this paper proposes a Multiscale Dense HOG-driven Self-Attention Network (MDHANet) that incorporates the Histogram of Oriented Gradients (HOG) as prior knowledge. First, the model employs three densely connected branches at different scales to obtain diverse multi-level feature representations through multi-scale feature reuse. Furthermore, since HOG can explicitly encode the distribution of gradient orientations and magnitudes in local neighborhoods and accurately characterize the feature distribution, a dynamic HOG-driven self-attention module is designed. Within this module, a HOG prior-based feature rearrangement strategy together with a dual-branch self-attention mechanism enables the network to more effectively learn the spatial information of the data, enhancing the extraction of discriminative spatial-structure features and further improving classification performance. Extensive experimental results on multiple datasets demonstrate that the proposed method outperforms existing state-of-the-art approaches. Extensive experimental results on four HSI datasets (IP, LK, HH, and HC) demonstrate that the proposed method outperforms existing state-of-the-art approaches, achieving OA of 99.13%, 97.32%, 99.02%, and 99.24%, respectively, with improvements of 0.83%, 0.41%, 2.69%, and 2.47% over the best competing methods.