Comparative Evaluation and Refinement of Algorithm for Water Depth Estimation using Medium Resolution Remote Sensing Data.

Abstract

Information concerning water depth of near shore water region is one of the most basic requirements for coastal zone management. Depth is especially important for near coastal lines, in harbors, and near shoals andbanks, where changes can occur rapidly as sedimentation, erosion and scouring of channels alters underwater topography. Although, the algorithms used to estimate depth from passive Remote Sensing (RS) satellite data were practicing last two decades, it is necessary to test the algorithms with recently available better quality RS data archives. In addition to that this paper proposes two methods to estimate water depth from single spectral band and from multispectral band. The single band algorithm called Radiance Based Estimation (RBE) is a modification of Stoffle and Halmo, (1991). Here onwards the single band method proposed by Stoffle and Halmo, (1991) is referred to as SLR (Single-band Linear Regression). A constant attenuation coefficient of a particular band is estimated using SLR method with limited in-situ depth; further the RBE is used to estimate depth from same band of multi-temporal images even when there is no in-situ depth available. RBE only depends upon the radiance value of the log-transformed band. The single spectral band from each satellite imagery is captured at different time is used to estimate multi-temporal depth from multi-source optical RS data such as Landsat 7, Landsat 8 and ASTER. The RBE applied to Landsat 8 data which has high radiometric resolution provided better accuracy compared to other satellite data in terms of correlation coefficient (R=0.89 and 0.80), coefficient of determination (R2=0.81 and 0.66) and Root Mean Square Error (RMSE=1.39 m and 2.09 m) from images collected on 31 January, 2014 and 4 March, 2014 respectively.The multispectral algorithm originally proposed by Clark et al., (1987, 1988) used only visible bands to estimate depth. Since the Clark et al., (1987and1988) wereusing a linear regression between multispectral bands and in-situ depth, here onwards depth estimation algorithm using multi-spectral bands from visible region from electromagnetic spectrum is called as Multiple-band Linear Regression (MLR). Multispectral bands are account for the varying degree of attenuation coefficient by addressing the heterogeneity of bottom types. This study proposes a new band combination to include Near InfraRed (NIR) band for depth estimation. The method to estimate depth using new band combination with visible and NIR band is hereafter called as MLR-NIR. The MLR and MLR-NIR were tested with the same data sets and observed that the results significantly improved by the proposed MLR-NIR. The depth estimation results derived from Landsat 8 data collected on 12 November, 2013, 31 January, 2014 and 4 March, 2014 by applying MLR-NIR shown better correlation coefficient (R=0.95, 0.95 and 0.90), coefficient of determination (R2=0.91, 0.91 and 0.83) and RMSE (1.53 m, 0.83 m and 1.23 m) as compared with MLR. The depth estimates derived from Landsat 8 data were used to investigate the reduction in accuracy due to turbidity has been carried out. The results indicate that NDWI (Normalized Difference Water Index) impacts the accuracy of depth estimates but not very significantly.