Ensembles of multiple spectral water indices for improving surface water classification.

  • SCI-E
作者: Zhaofei Wen;Ce Zhang;Guofan Shao;Shengjun Wu;Peter M. Atkinson
作者机构: Department of Forestry and Natural Resources, Purdue University, West Lafayette 47906, USA
Key Laboratory of Reservoir Aquatic Environment, Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing 400714, China
Lancaster Environment Centre, Lancaster University, Lancaster LA1 4YQ, UK
UK Centre for Ecology & Hydrology, Library Avenue, Bailrigg, Lancaster LA1 4AP, UK
语种: 英文
关键词: Water index;Threshold;Integrated decision making;Mixed pixels;MNDWI
ISSN: 1569-8432
年: 2021
卷: 96
页码: 102278
基金类别: Open Fund of the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University [18R07]; National Natural Science Foundation of ChinaNational Natural Science Foundation of China (NSFC) [41501096, 51779241]
摘要: Mapping surface water distribution and its dynamics over various environments with robust methods is essential for managing water resources and supporting water-related policy design. Thresholding Single Water Index image (TSWI) with threshold is a common way of using water index (WI) for mapping water for it is easy to use and could obtain acceptable accuracies in many applications. As more and more WIs are available and each has its distinct merits, the real-world application of TSWI, however, often face two practical concerns: (1) selection of an appropriate WI and (2) determination of an appropriate threshold for a given WI. These two issues are problematic for many users who rely either on trial-and-error procedures that are time-consuming or on their personal preferences that are somewhat subjective. To better deal with these two practical concerns, an alternative way of using WIs is suggested here by transforming the current paradigm into a simple but robust ensemble approach called Collaborative Decision-making with Water Indices (CDWI). A total of 145 subsite images (900 × 900 m) from 22 Landsat-8 OLI scenes that covering various water-land environments around the world were used to assess the performance of TSWI and the CDWI. Five benchmark WIs were adopted in five TSWI methods and CDWI method: Normalized Difference Water Index (NDWI), the Modified NDWI (MNDWI), the Automated Water Extraction Indices without considering (AWEI0) and with considering (AWEI1) shadows, and the state-of-the-art 2015 water index (WI2015). Two aspects of performance were analyzed: comparing their accuracies (indicated by both F1-scores and Youden’s Index) over various environments and comparing their accuracy sensitivities to threshold. The results demonstrate that CDWI produced higher accuracies than the other five TSWI methods for most application cases. Particularly, more cases (indicated by percentage) produced higher F1-scores by CDWI than the other five TSWI methods, i.e. 67% (CDWI) vs. 15% (TSWINDWI), 54% (CDWI) vs. 22% (TSWIMNDWI), 42% (CDWI) vs. 12% (TSWIAWEI0), 57% (CDWI) vs. 17% (TSWIAWEI1), and 34% (CDWI) vs. 12% (TSWIWI2015). Moreover, the F1-score of the CDWI is less sensitive to the change of thresholds compared with that of the five TSWI methods. These important benefits of CDWI make it a robust approach for mapping water. The uncertainty of CDWI method was thoroughly discussed and a general guidance (or look-up-table) for determining parameters of CDWI method was also suggested. The underlying framework of CDWI could be readily generalizable and applicable to other satellite images, such as Landsat TM/ETM+, MODIS, and Sentinel-2 images.

Ensembles of multiple spectral water indices for improving surface water classification.