Accurate water extraction using remote sensing imagery based on normalized difference water index and unsupervised deep learning

Abstract

• Our method has the potential to extract water with high accuracy on a large scale. • Our method realizes the conversion from unlabeled learning to noisy label learning. • The model performance is significantly higher than traditional methods. • Our method provides a new idea for unsupervised learning. Large-scale monitoring of surface water bodies is of great significance to the sustainable development of regional ecosystems. Remote sensing is currently the main method of global earth observation. On the one hand, the traditional water index is simple and efficient, but it relies on a fixed global threshold, which leads to low accuracy for water extraction. On the other hand, deep learning has achieved state-of-the-art results in classification of spectral data, but it consumes a substantial amount of manpower and time to label sufficient high-quality samples. Furthermore, spectral characteristics of water bodies in different areas vary greatly due to changes in atmospheric conditions and viewing geometry. In this paper, we propose a new accurate water extraction framework based on unsupervised deep learning and NDWI of multispectral images. Binarized NDWI images are used to identify potential water bodies, and deep learning training is performed using these pseudo samples and labels. This process realizes the conversion from unlabeled learning to noisy label learning. First, we proposed a simple and fast binarization algorithm to segment as many real water bodies as possible from NDWI images. Then a set of water confidence assessment rules was constructed from the four aspects of the spectrum, shape, agglomeration, and range. The water segments were scored and sorted to make the model start learning from easy samples and gradually advance to complex samples. Finally, the adjusted co-teaching learning strategy is adopted to filter errors introduced by noisy labels through peer networks with different learning capabilities during training. We tested the accuracy of our method using Gaofen Image Dataset (GID), Sentinel-2 and Landsat images of several water bodies in China. Compared with other methods, our method improved the F1 score by 18.1–40.3% and 6.8–22.2% for GID and Sentinel-2 images, respectively. In addition, our method is more stable in long-term water monitoring. The proposed method has the potential to be used for extracting water bodies with high accuracy on a large scale, especially in areas with complex environments and a lack of deep learning samples. And it provides a new idea for unsupervised learning in the current remote sensing field by fully combining remote sensing expertise and spectral information of ground objects.