Answers are grounded in this paper and cite their sources.
Few-Shot Remote Sensing Scene Classification (FSRSSC) aims at recognizing novel categories with only a few labeled samples. Local representations can retain richer details than image-level feature vectors and adapt well to the characteristics of spatially scattered objects in remote sensing images. However, the existing local representation-based few-shot learning works generally overlook the role of class-semantic information. To address this issue, in this paper, a Class-aware Local Representation Learning (CLRL) method is proposed for FSRSSC. It generates class-aware support local representations by calibrating the local features with class-level global features to alleviate the problem of ignoring the class-semantic information. The Image-to-Class (I2C) metric is used to measure the similarities between query samples and support samples. Moreover, a novel Local Representation-based Self-supervised Learning (LRSL) strategy is proposed. The main idea of LRSL is to maximize the relationship consistency between the query local representations and different views of support local representations. A novel consistency loss is further proposed as the objective function of LRSL to supervise the model training. Different from the existing image-level feature vector based self-supervised learning tasks, the proposed LRSL forces the model to capture the structure information of local representations. A robust model can be trained by jointly optimizing LRSL tasks and FSRSSC tasks. The experimental results on the NWPU-RESISC45, WHU-RS19 and UC Merced datasets show the superiority of our proposed method.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: