A Multi-scale Densely Connected and Feature Aggregation Network for Hyperspectral Image Classification

Jian Zhu, Yi Liu, Jiajie Feng, Caihong Mu

Abstract: Convolutional neural networks have been widely used in the field of hyperspectral image (HSI) classification due to their excellent ability to model local regions, and have achieved good classification performance. However, HSI classification still faces challenges such as insufficient representation of spectral-spatial features and inadequate fusion of multi-level features. To address these issues, we propose a Multi-scale Densely Connected and Feature Aggregation Network (MSDC-FAN) for HSI classification. The network mainly consists of a Spectral-Spatial Feature Extraction (SSFE) module, three Multi-scale Feature Extraction (MSFE) modules, and a Multilevel Feature Aggregation Module (MFAM). Firstly, the SSFE module is carried out to extract more comprehensive spectral-spatial features. Secondly, three MSFE modules are used in sequence to extract multi-scale features and highlight significant features, thus further improving the model's performance. Finally, the MFAM is designed to aggregate features at different levels, enhancing the model's feature representation ability. Experimental results on two commonly used hyperspectral datasets demonstrate the superiority of the proposed method.