An index and approach for water extraction using Landsat–OLI data

Abstract

ABSTRACTThe extraction of water distribution is extremely useful in research and planning activities, including those associated with water resources, environments, disasters, local climates, and other factors. Remote-sensing images with moderate resolution have been the main data source due to the vast distribution of water and the high cost, access difficulty, and massive size of high-resolution images. Although some water indices and methods for water extraction have been proposed, there is still a lack of these resources to easily, accurately, efficiently, and automatically extract water. This paper focused on some improvements that mainly used the most traditional but also the newest Operational Land Imager (OLI) images in Landsat 8. This study first analysed the variation features of previous water indices. Secondly, taking the city of Beijing and its surrounding area as the experimental site, a spectral curve analysis was performed and a new water index was proposed. This index was compared to three typical indices. Thirdly, a new approach was proposed to accurately and easily extract water. It included four major steps: background partitioning, thresholding and preliminary segmentation, noise removal by patch size, and local region growth. Next, the stricter and more effective stratified random sampling method was used to test the accuracy. Then, we tested the generality of the proposed water index and extraction method using nine typical test sites from around the world and tried to simplify the workflow. Finally, this paper discusses threshold optimization issues, such as automatic selection and reduction of the number of thresholds. The results show that the normalized water index (NDWI), modified normalized water index (MNDWI), and normalized difference built-up index (NDBI) may fail in some situations due to the complex spectrum of the impervious surface class. Some shadow pixels were impossible to remove using only spectral analysis because both the digital number (DN) trends and values were similar to those of water. The proposed water index was easy and simple, but it corresponded better to water bodies. Additionally, it was more accurate and universal and showed greater potential for extracting water. This method relatively accurately and completely extracted various water bodies from plain city, plain country, and natural mountainous regions in many typical climate zones, eliminating interference caused by dark impervious surfaces, plants, sand, suspended sediments, snow, ice, bedrock, reservoir drawdown areas, shadows from mountains and buildings, mixed pixels, etc. The mean kappa coefficients were 0.988, 0.982, and 0.984 in plain city, plain country, and natural mountainous regions, respectively. This paper suggests that thresholds can be automatically determined by comparing the accuracy changes of different thresholds according to preselected sample and test points. Furthermore, the combined use of the maximum class square error method (also known as the Ostu algorithm) and the adaptive thresholding method exhibits great potential for automatic determination of thresholds in regions without many noises with higher water index values. In addition, water bodies could also be accurately extracted by setting these thresholds to fixed values based on the results at more test sites.