GEOGRAPHICALLY WEIGHTED REGRESSION APPROACH FOR SHALLOW WATER DEPTH ESTIMATION USING MULTISPECTRAL SATELLITE IMAGERIES

Abstract

Shallow water depth is essential for coastal planning, monitoring, and research. Bathymetry data is mostly produced from hydrographic survey using echosounder. The generic result from those measurements is discrete values while the desired output is a continuous depth model. To fill the gaps in the sounding data, we use Satellite Derived Bathymetry (SDB) approach with Geographically Weighted Regression (GWR). This study aims to investigate the feasibility of GWR to model bathymetry of shallow water in the eastern part of Indonesia. We explore the correlation between the number of training data and the predicted result. Two different satellites images are used, namely: Sentinel-2A and Landsat 8 OLI/TIRS with 10 and 30 m resolutions respectively. For the experiment, in-situ data are set into training and validation in three different ratios. The model is developed using adaptive GWR approach in which the parameter of regression would adapt the local data set within different kernel sizes. Finally, we compute RMSE (Root Mean Square Error), R2, and TVU (Total Vertical Uncertainty) to assess the quality of our model. In general, Sentinel-2A produces more detailed information due to higher resolution than Landsat 8 OLI/TIRS. Sentinel-2A also obtains more accurate results based on RMSE values. The percentage number of the estimated depth that fulfils TVU requirements is up to 83%. These assessment quality results give us an insight that the SDB approach using GWR is promising. Thus, the GWR method may be able to provide an estimate of bathymetry for many coastal areas in Indonesia.