Multi-Level Relation Embedding with Attention Mechanism for Few-Shot Remote Sensing Image Scene Classification

Given the limited availability of annotated remote sensing image samples, which renders traditional deep learning algorithms ineffective, and considering the complex spatial semantic features with high intra-class variability and inter-class similarity, we propose the RNC_SCAM model based on multi-level relation embedding with an attention mechanism. This model is designed for few-shot remote sensing image scene classification in defense and military applications, offering significant benefits for military reconnaissance, target identification, and localization. A hybrid attention mechanism is integrated into the feature extraction network to enhance the network's ability to extract and represent fundamental image features more effectively. The basic feature maps are input into the Self-Association Module (SAM) and Cross-Association Module (CAM) in sequence to achieve multi-level relationship embedding to fully learn the feature relationships within and between image classes. The experimental results show that when the experimental settings are 5way-1shot and 5way-5shot, the average classification accuracy of the RNC_SCAM model on the NWPU-RESISC45 and AID datasets are 74.57±0.39%, 90.01±0.19% and 66.70±0.43%, 80.73±0.26% respectively, the classification effect has been significantly improved.