Scholarlink

<p>In recent years, remote sensing scene classification has witnessed substantial advancements in knowledge and information systems, particularly through the integration of deep learning technologies. Conventional methods frequently depend on voluminous datasets with extensive labels and complex deep learning models, which present challenges in settings with limited resources. This study proposes a novel classification method that aims to enhance performance in a few-shot learning framework. The proposed method integrates multi-source data augmentation and knowledge distillation strategies to achieve this objective. In light of the complexity and variations inherent in remote sensing imagery, a multi-source data augmentation technique was initially implemented. This technique encompassed a range of processes, including multi-scale scaling, multi-angle rotations, and mixed sample enhancements. The primary objective of this approach was to enhance the model’s generalizability capabilities and to facilitate the recognition of complex scene features. Furthermore, a lightweight deep learning model was developed that integrates a feature pyramid network with attention mechanisms to expedite the processing of multi-scale information and accentuate critical features. In order to optimize the performance of the model in environments characterized by limited computational resources, a knowledge distillation technique was applied. This technique involves the transfer of insights from complex models to our streamlined model, with the objective of refining the learning process through feature distillation of mixed samples and scale-aware features. Our method achieves accuracies of 86.60%, 94.59%, and 83.57% on the three datasets, respectively, under the 5-way 5-shot setup. The experimental results, derived from four public remote sensing datasets (UCM, WHU-RS19, and NWPU-RESISC45), demonstrate the efficacy of the proposed method in enhancing the accuracy of scene classification. This research constructs a lightweight deep learning model that can autonomously parse complex remote sensing scenes by simulating human visual cognitive mechanisms (e.g., multi-scale feature processing and attentional focusing), which embodies the ability of cognitive computing systems to intelligently understand environmental information.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-source data augmentation and knowledge distillation for few-shot remote sensing classification

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Chaunjun Zhao,
  • Meiqi Liu,
  • Xuzhuang Sun

摘要

In recent years, remote sensing scene classification has witnessed substantial advancements in knowledge and information systems, particularly through the integration of deep learning technologies. Conventional methods frequently depend on voluminous datasets with extensive labels and complex deep learning models, which present challenges in settings with limited resources. This study proposes a novel classification method that aims to enhance performance in a few-shot learning framework. The proposed method integrates multi-source data augmentation and knowledge distillation strategies to achieve this objective. In light of the complexity and variations inherent in remote sensing imagery, a multi-source data augmentation technique was initially implemented. This technique encompassed a range of processes, including multi-scale scaling, multi-angle rotations, and mixed sample enhancements. The primary objective of this approach was to enhance the model’s generalizability capabilities and to facilitate the recognition of complex scene features. Furthermore, a lightweight deep learning model was developed that integrates a feature pyramid network with attention mechanisms to expedite the processing of multi-scale information and accentuate critical features. In order to optimize the performance of the model in environments characterized by limited computational resources, a knowledge distillation technique was applied. This technique involves the transfer of insights from complex models to our streamlined model, with the objective of refining the learning process through feature distillation of mixed samples and scale-aware features. Our method achieves accuracies of 86.60%, 94.59%, and 83.57% on the three datasets, respectively, under the 5-way 5-shot setup. The experimental results, derived from four public remote sensing datasets (UCM, WHU-RS19, and NWPU-RESISC45), demonstrate the efficacy of the proposed method in enhancing the accuracy of scene classification. This research constructs a lightweight deep learning model that can autonomously parse complex remote sensing scenes by simulating human visual cognitive mechanisms (e.g., multi-scale feature processing and attentional focusing), which embodies the ability of cognitive computing systems to intelligently understand environmental information.