Surface Water body Extraction using a Progressive Enhancement Model from remote sensing images

Abstract

The most commonly used water extraction methods, the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Water Index (NDWI) and the Modified Normalized Difference Water Index (MNDWI), use spectral bands to differentiate between water and other features in images. Each method has its drawbacks: NDVI cannot effectively distinguish between vegetation and water; NDWI confuses building shadows and water; and MNDWI misinterprets mountain shadows as water. Meanwhile a single water index method is not always the best choice when people extract water bodies from images because a threshold value to be selected for extracting water bodies is empirical. This paper presents a model to progressively enhance surface water information extraction. This model first calculates three indexes (NDVI, NDWI, and MNDWI), extracts water information by setting thresholds, and adds up the three isolated results to progressively enhance accuracy. This model can be used to quickly extract water information with high accuracy and addresses problems with previous indexes. In order to test accuracy, the results of the new water extraction model were compared to reality in selected test areas. The results show that compared with those from the traditional methods, the overall accuracy and Kappa value of the new method increased by 13% and 0.26 to extract lakes and fishponds, and 1% and 0.02 to extract rivers.