Remote Sensing Image Scene Classification Based on Mutual Learning With Complementary Multi-Features

A novel neural network model based on mutual learning, with complementary multi-features (MLCMFNet), is proposed for scene classification, addressing common issues with insufficient extraction to more effectively learn target features from remote sensing images. First, a DenseNet-67 framework is adopted as the backbone network for local feature extraction, to identify multi-level image features. Second, a multi-feature fusion module (MFFM) is designed to extract and fuse feature information from different images by effectively locating, normalizing, fusing, and integrating significant regional features. Third, a multi-attention fusion module (MAFM) is introduced to extract and fuse feature information from different depths by focusing on specific features used to extract more representative local image details. In addition, a Swin transformer module (STM) is employed to extract global image features. Finally, unlike conventional knowledge distillation in which a teacher network guides a student network, the two networks constructed in this study establish a complementary relationship between local and global features. As such, these networks learn from each other and share experiences to more effectively aggregate extracted details and identify target features. The proposed MLCMFNet method was evaluated using four popular open-source datasets (RSSCN7, AID, UC-Merced, and NWPU-RESISC45). Experimental results demonstrated that the proposed MLCMF-Net framework effectively improved the accuracy of scene classification compared to existing models.