Automatically Designed CNN for HSI Classification Based on Circular Kernel Convolution

In order to address the challenges of geometric feature distortion in traditional square convolutional kernels and the inefficiency of manual architecture design for hyperspectral image classification (HSIC), this paper proposes a Neural Architecture Search framework with Circular Kernel Convolution and Transformer (CT-NAS). Firstly, a hybrid search space is constructed by introducing 3×3/5×5 circular convolutions and separable circular convolutions, which enhances local feature capture through isotropic receptive fields while reducing boundary effects via periodic padding. Secondly, a dual-layer search strategy dynamically selects spatial/spectral-dominant cells in the outer layer and optimizes multi-scale circular convolution topologies in the inner layer, achieving adaptive fusion of spatial-spectral features. The framework further integrates a lightweight Transformer module to supplement global context awareness. Experimental results demonstrate that CT-NAS achieves state-of-the-art overall accuracy (OA) of 99.62% and 99.44% on Pavia and PaviaU datasets respectively, with a 14.53% accuracy improvement for gravel-type features compared to 3D-Auto-CNN. Visualization results confirm its superior boundary consistency in complex scenarios, proving that circular convolution's rotational invariance and NAS-driven architecture co-design effectively address geometric distortion problems in HSIC.