DB-FusionNet: A dual-branch CNN framework with spatial pyramid-based multi-scale spatial-spectral fusion for HSI classification

Le résumé fourni par la source

Hyperspectral image (HSI) classification is critical for accurate land cover mapping by utilizing rich spectral and spatial information.However, existing methods often struggle with inadequate fusion of spatial and spectral features, excessive computational burden, and limited generalization capability. To overcome these challenges, this study proposes a novel dual-branch convolutional neural network framework, termed DB-FusionNet. The architecture consists of two complementary branches: the first integrates three-dimensional (3D) and two-dimensional (2D) convolutional neural networks along with a Spatial Pyramid Pooling Fusion (SPPF) module, facilitating efficient multi-scale spatial-spectral feature extraction while reducing computational complexity. The second branch adopts a one-dimensional (1D) CNN to effectively capture fine-grained spectral variations, thereby enhancing the model’s sensitivity to subtle inter-band differences. DB-FusionNet achieves overall accuracies (OA) of 98.93% on Indian Pines, 99.95% on Salinas, 99.96% on WHU-Hi-LongKou, 99.75% on WHU-Hi-HanChuan, and 99.87% on WHU-Hi-HongHu, consistently outperforming traditional and state-of-the-art models including SVM, 3D CNN, and HybridSN. Besides accuracy gains, DB-FusionNet shows strong generalization across different scenes and improved convergence efficiency.These results confirm the effectiveness and robustness of the proposed architecture for hyperspectral image classification.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.