Domain Adaptation using NDWI index for Water-land Semantic Segmentation

Recently, deep-learning based method has achieved great success in remote sensing image segmentation. Deep neural networks for semantic segmentation always require a large number of samples, which becomes a tough problem especially in remote sensing area. Unsupervised domain adaptation utilizes the labeled source dataset to train the unlabeled target dataset, which can reduce the labeling cost greatly. The key challenge of this task is how to alleviate the data distribution discrepancy of feature between the two datasets meanwhile keep the detail information in target dataset. This paper proposes a novel framework based on AdaptSegNet model that combined with remote sensing derived indices. We concatenate the NDWI index to the output layer of segmentation module, which can reflect waterbody at pixel-level with value from -1 to 1 and improve the segmentation accuracy greatly. we adopt 3 × 3 convolution to learn a weight that update the class probability after concatenation. Experiments are conducted from the GID dataset to the Landsat8 datasets. Our method achieves 90.0% mean intersection-over-union (IoU) and outperforms all of the baselines