LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation

• DOI: 10.48550/arxiv.2110.08733
• Publication Date: 2021-01-01
• Authors: Junjue Wang; Zhuo Zheng; Ailong Ma; Xiaoyan Lu; Yanfei Zhong
• Keywords: Land Cover; Domain Adaptation; Representation

Deep learning approaches have shown promising results in remote sensing high spatial resolution (HSR) land-cover mapping. However, urban and rural scenes can show completely different geographical landscapes, the inadequate generalizability of these algorithms hinders city-level or national-level Most existing HSR datasets mainly promote research semantic representation, thereby ignoring model transferability. In this paper, we introduce Land-cOVEr Domain Adaptive segmentation (LoveDA) dataset to advance transferable learning. The LoveDA contains 5987 images with 166768 annotated objects from three cities. Compared datasets, encompasses two domains (urban rural), which brings considerable challenges due the: 1) multi-scale objects; 2) complex background samples; 3) inconsistent class distributions. is suitable for both unsupervised domain adaptation (UDA) tasks. Accordingly, benchmarked on eleven methods eight UDA methods. Some exploratory studies including architectures strategies, additional supervision, pseudo-label analysis were also carried out address challenges. code data are available at https://github.com/Junjue-Wang/LoveDA.