SoftFormer: SAR-optical fusion transformer for urban land use and land cover classification

Abstract

Classification of urban land use and land cover is vital to many applications, and naturally becomes a popular topic in remote sensing. The finite information carried by unimodal data, the compound land use types, and the poor signal-noise ratio caused by restricted weather conditions would inevitably lead to relatively poor classification performance. Recently in remote sensing society, multimodal data fusion with deep learning technology has gained a great deal of attention. Existing research exhibit integration of multimodal data at a single level, while simultaneously lacking exploration of the immense potential provided by popular transformer and CNN structures for effectively leveraging multimodal data, which may fall into the trap that makes the information fusion inadequate. We introduce SoftFormer, a novel network that synergistically merges the strengths of CNNs with transformers, as well as achieving multi-level fusion. To extract local features from images, we propose an innovative mechanism called Interior Self-Attention, which is seamlessly integrated into the backbone network. To fully exploit the global semantic information from both modalities, in the feature-level fusion, we introduce a joint key–value learning fusion approach to integrate multimodal data within a unified semantic space. The decision and feature level information are simultaneously integrated, resulting in a multi-level fusion transformer network. Results on four remote sensing datasets show that SoftFormer is able to achieve at least 1.32%, 0.7%, and 0.99% performance improvement in overall accuracy, kappa index, and mIoU, compared to other state-of-the-art methods, the ablation studies show that multimodal fusion outperforms the unimodal data on urban land cover and land use classification, the highest overall accuracy, kappa index as well as mIoU improvement can be up to 5.71%, 10.32% and 7.91%, and the proposed modules are able to boost performance to some extent, even with cloud cover. Code will be publicly available at https://github.com/rl1024/SoftFormer.