Abstract:
Threshold-based index approaches are highly successful in differentiating between water and non-water components in optical satellite images. More importantly, water index offers a high level of precision in monitoring surface water. Index-based methods use different spectral bands that lead to a generation of different results. Different indices can generate good results in different challenging situations, such as environmental noise like shadows, forests, built-up areas, snow, and clouds. Taking that into consideration, this study evaluates the performance of the most widely used water indices: Normalized Difference Water Index (NDWI), Modified NDWI (MNDWI), Normalized Difference Moisture Index, Water Ratio Index, automated water index, new water index (NWI), and Sentinel water index (SWI) in a Sentinel-2 scene of Ukhiya and Teknaf region of Bangladesh. Two major problems of applying these index-based methods except for NDWI and SWI for Sentinel-2 are inconsistent band resolutions and the lack of a panchromatic band. This study solves the later problem of panchromatic band by doing a linear regression analysis among different spectral bands. After the selection of a panchromatic band, the latter problem is solved by comparing different established Pan sharpening methods like Intensity Hue Saturation, Gram-Schmidt, High Pass Filter, Wavelet Resolution Merge, and Sen2Res tool. From the Linear Regression analysis during 2021-23, it appears that Short Wave Infrared-1 (SWIR-1), SWIR-2, and Vegetation Red Edge (VRE-1) correlate most with the average of 4 spectral bands, Red band, respectively, with the highest R2 values ranging from 0.42 to 0.74. Therefore, an average of 4 spectral bands and a Red band are selected as panchromatic bands. From the field validation and the analysis of Spatial Correlation Coefficient, Universal Image Quality Index, Root Mean Squared Error, and Spatial Correlation Coefficient, it appears that the High Pass Filter Pan sharpening method is the best in identifying water bodies for all years (2021-23) and bands, specifically SWIR-1, SWIR-2, and VRE-1. Subsequently, an appropriate threshold is identified using Isodata, Li, Triangular, Mean, Yen, Otsu, Huang, and Minimum methods. Here, out of 350 validation points, 96 field survey locations in 2023 and 254 Google Earth Pro data for 2021-22 are used in deriving a confusion matrix. Results show that NDWI commonly performs better for the studied region with the highest kappa range varying between 0.81 and 0.94 for Yen, Isodata, and Triangle thresholding techniques in 2021, 2022, and 2023, respectively. However, historical image analyses from 2016 to 2023 reveal that the Isodata is consistent in producing high false positive values. In contrast, Yen consistently outperforms other methods, including manual thresholding. Therefore, the NDWI method with Yen thresholding is ultimately selected to identify changes in the historical waterbody area (WBA) within the examined region. There is undeniable evidence of a substantial decline in the WBA from 2016 to 2023, with an annual decreasing rate of 3.17 Km2 WBA each year. Therefore, it is of the utmost importance to preserve the watershed and prevent its deterioration before it becomes irreversible.