Towards robust classification of multi-view remote sensing images with partial data availability

Utilizing remote sensing to monitor and obtain the land use information is crucial for sustainable development goals (SDGs), including sustainable agriculture, urbanization processes, land reclamation, etc. The development of remote sensing big data and deep learning has greatly promoted the use of multi-source images to understand land use. However, in practical applications, missing data often occurs due to high cost and environmental limitation. Therefore, we propose a framework which can towards robust classification of multi-view remote sensing images with partial data availability. First, we construct a student model and teacher model mutual learning framework. In particular, we promote the consistency of student model and teacher model which enhances robustness under missing view and further improves performance under complete views. Second, we propose a parameter-free channel & spatial attention (PFCSA) module embedded in the image encoders, which allows the architecture with three encoders to balance performance and lightweight. Further, the cross-view attention fusion (CAF) module is designed to enhance the fusion of multi-view images. The experiments based on global data show that the proposed method can utilize multi-views more effectively than common fusion strategies. Our method also ensures robustness against missing view compared to other methods. And we reveal the effectiveness of each proposed strategies. In addition, the contribution of different views and the mechanism of the model under missing view are analyzed. The proposed method in this study can be used to generate land use and its derived geographic information products on a global scale (including different natural and development regions), and further serve the realization of global SDGs. The code will be publicly available at https://github.com/mfzhao1998/multi_view_incomplete_learning.