Weakly supervised high spatial resolution land cover mapping based on self-training with weighted pseudo-labels

Despite its success, deep learning in land cover mapping requires a massive amount of pixel-wise labeled images. It typically assumes that the training and test scenes are similar in data distribution. The performance of models trained on any particular dataset could degrade significantly on a new dataset due to the domain shift or domain gap across datasets, resulting in new training data requiring labor-intensive manual pixel-wise labeling. This paper proposes a land cover mapping framework combining Feature Pyramid Network (FPN) and self-training. In the FPN, we integrate ConvNeXt with a Pyramid Pooling Module (PPM). Combining the FPN and the PPM improves the segmentation performance, which benefits from the multiscale aggregation of pyramid features. To fully exploit pseudo-labels, we design an Unsupervised Domain Adaptation (UDA) land cover mapping scheme with self-training using weighted pseudo-labels of the target samples. The proposed land cover mapping framework could benefit from multiscale aggregation of pyramid features and the full use of the pseudo-labels. Comparison results on the LoveDA dataset, the latest large-scale unsupervised domain adaptation dataset for land cover mapping, empirically demonstrated that our land cover mapping approach significantly outperforms the baselines in both UDA scenarios, i.e., Urban-* Rural and Rural-* Urban. The models of this paper are now publicly available on GitHub.1

Automated summary

This paper proposes a land cover mapping framework combining FPN and self-training, which improves segmentation performance through multiscale aggregation and the use of pseudo-labels. The method significantly outperforms baselines on the latest unsupervised domain adaptation dataset.