Flood water body identification and change monitoring in semi-arid areas using multi-source heterogeneous remote sensing images

Abstract

[Objective] In order to solve the problem that the low temporal resolution and high spectral heterogeneity of remote sensing images in flood monitoring in arid areas of northwest China lead to low recognition accuracy of flood areas and inability to extract more detailed flood change information, a flood area recognition and change monitoring method based on spatial-spectral feature fusion and multi-source heterogeneous remote sensing images correlation analysis is proposed. [Methods] Taking the flood event in Sarhusong Township, Altay Prefecture, Xinjiang as an example, seven temporal multispectral images of Landsat8, HJ-1A, Sentinel-2A and GF-1 before and after the flood event are obtained. Then multi-dimensional feature vectors, including water index(NDWI, NDMBWI, WI2021) of image spectral reflectance, entropy, homogeneity texture features and as well as are extracted from them. PCA technology is used to reduce feature dimension; Finally, the random forest(RF) classifier is used to fuse multi-dimensional spatial-spectral features so as to identify water bodies and to recognize the flooding areas from remote sensing images acquired in every periods. After comparing the water body recognition result of adjacent temporal images, the dynamic change information of flood submerged areas is obtain. [Results] Through experimental verification, the result indicate that the false alarm rates for water extraction from Landsat images are 0. 21%, 0. 28%, and 0. 32%, with corresponding the miss rates of 2. 17%, 3. 37%, and 0. 110%. The maximum flooded area is 355. 1 km², with submerged farmland and grassland covering areas of 134. 3 km² and 229. 2 km², respectively. [Conclusion] The following conclusion can be obtained that the RF recognition algorithm with PCA for multi-feature fusion significantly improves the low recognition accuracy of scattered water bodies in single temporal Landsat8 images, and the overall accuracy of water body recognition is 13. 7%, 10. 8% and 2. 03%, higher than that of NDWI water body index method; The use of multi-source remote sensing image data makes the monitoring cycle as short as 1 week, which makes the extracted flood evolution process information more detailed and makes up for the shortage of satellite transit time; In addition, the remote sensing monitoring result of flood dynamic changes are basically consistent with the development trend of meteorological and hydrological observation data. Through the identification and change monitoring of flood water bodies in Sarhusong Township, it is fully demonstrated that in semi-arid areas, multi-source optical remote sensing images can effectively identify flooded areas, providing important data support for emergency disaster relief.

Keywords

multi-source remote sensing images / water body identification / spatial-spectrum features fusion / random forest / time-spatial change analysis / flood disaster monitoring / flood / climate change