109 / 2024-08-15 23:51:05

A Few-Shot Remote Sensing Scene Classification Method Based on Transfer Learning and a Multi-layer Attention Structure.

Keywords: Remote Sensing classification ,Transfer learning ,Multilayer attention.

Abstract. Remote sensing scene classification is a critical application in the field of remote sensing image interpretation. However, due to factors such as extreme environments and limited detection time, acquiring remote sensing scene image samples is challenging, resulting in insufficient data. Additionally, in practical applications, remote sensing images are often combined with similar information from different platforms to provide more comprehensive and multi-layered geographic information support. This leads to limitations in deep learning, particularly due to the scarcity of remote sensing data and the adaptability of models. In response to these challenges, this paper proposes a few-shot remote sensing scene classification method based on transfer learning and a multi-layer attention structure. The method leverages pre-trained feature extraction networks for few-shot remote sensing scene classification tasks through transfer learning, combined with multi-layer attention modules. This approach allows the model to learn general features in advance, enhancing its adaptability, effectively filtering out irrelevant information, and focusing on key scene areas across different levels. As a result, the model achieves strong performance with limited labeled data, improving its robustness. The proposed method demonstrated advanced performance on the NWPU-RESISC45 and MEDIC datasets. The experimental results indicate that this approach effectively enhances the accuracy of few-shot remote sensing scene classification, offering valuable insights for addressing practical application issues in scenarios like disaster detection with limited sample sizes.