River Water Mapping by fusing Landsat-8, Sentinel-1 and Sentinel-2 based on Spatio-temporal Weighted Dempster-Shafer Evidence Theory

Abstract:

River water extent information is the basic input of river hydrodynamics and the core parameter for calculating river discharge. Studying the spatial and temporal dynamics of river water extent is of great significance for grasping the hydrological situation of the river and understanding the hydrological characteristics of the basin. Although numerous methods have been proposed for mapping river water from either optical or Synthetic aperture radar (SAR) remotely sensed images, uncertainties are broadly existing in most of these methods, due to all kinds of reasons, such as cloud cover, image quality, and mixed pixel issues. Image fusion is an effective approach to take advantage of multiple data sources and achieve high quality and high intensity earth observations. Therefore, how to quantify the uncertainties of river water mapping, and reduce the uncertainties through multi-source image fusion, in order to achieve higher accuracy and higher frequency of river water mapping is of great significance for monitoring river hydrological conditions.

In this study, we propose a spatio-temporal weighted Dempster-Shafer evidence theory method (STWDS) for river water mapping. It uses multisource remote sensing imagery including Sentinel-1, Sentinel-2 and Landsat-8 as the input. A posterior probability Support Vector Machine (PPSVM) method was first applied to all these different data sources in order to generate a consistent quantification of river water mapping uncertainties. A spatio-temporal weighted Dempster-Shafer evidence theory was then employed to fuse the results of PPSVM, aiming to reduce river water mapping uncertainties by combing multisource observations . A distance decay model that incorporates both spatial and temporal information was developed to assign weights to neighboring pixels from different data, which were then employed to calculate the similarities of multiple evidences (probabilities of neighboring pixels). The similarities were then normalized and considered as the reliability of different evidences. Different weights were assigned to different evidences according to their reliability, which were then combined with the initial probabilities to generate a weighted Mass function. A rule of judgement was established based on the D-S evidence theory, which helped identify final river water pixels based on the Mass function.

We used three river sections with different characteristics in northwest inland region of China as the case study areas. High resolution aerial images were acquired in 2019 summer using DJI Phantom 4. They were used as the major reference to validate and evaluate the results of our proposed STWDS method. It was found that the river water maps obtained by the proposed method had higher accuracy than those derived from a single data source.

Keywords: Dempster-Shafer Evidence Theory; Support Vector Machine; Optical images; Synthetic aperture radar images