摘要：In empirical approach, the satellite-derived bathymetry (SDB) is usually derived from a linear regression.However, the depth variable in surface reflectance has a more complex relation.In this paper, a methodology was introduced using a nonlinear regression of Random Forest (RF) algorithm for SDB in shallow coral reef water.Worldview-2 satellite images and water depth measurement samples using single beam echo sounder were utilized.Furthermore, the surface reflectance of six visible bands and their logarithms were used as an input in RF and then compared with conventional methods of Multiple Linear Regression (MLR) at ten times cross validation.Moreover, the performance of each possible pair from six visible bands was also tested.Then, the estimated depth from two methods and each possible pairs were evaluated in two sites in Indonesia: Gili Mantra Island and Panggang Island, using the measured bathymetry data.As a result, for the case of all bands used the RF in compared with MLR showed better fitting ensemble, -0.14 and -1.27m of RMSE and 0.16 and 0.47 of R2 improvement for Gili Mantra Islands and Panggang Island, respectively.Therefore, the RF algorithm demonstrated better performance and accuracy compared with the conventional method.While for best pair identification, all bands pair wound did not give the best result.Surprisingly, the usage of green, yellow, and red bands showed good water depth estimation accuracy.
其他摘要：In empirical approach, the satellite-derived bathymetry (SDB) is usually derived from a linear regression. However, the depth variable in surface reflectance has a more complex relation. In this paper, a methodology was introduced using a nonlinear regression of Random Forest (RF) algorithm for SDB in shallow coral reef water. Worldview-2 satellite images and water depth measurement samples using single beam echo sounder were utilized. Furthermore, the surface reflectance of six visible bands and their logarithms were used as an input in RF and then compared with conventional methods of Multiple Linear Regression (MLR) at ten times cross validation. Moreover, the performance of each possible pair from six visible bands was also tested. Then, the estimated depth from two methods and each possible pairs were evaluated in two sites in Indonesia: Gili Mantra Island and Panggang Island, using the measured bathymetry data. As a result, for the case of all bands used the RF in compared with MLR showed better fitting ensemble, -0.14 and -1.27m of RMSE and 0.16 and 0.47 of R 2 improvement for Gili Mantra Islands and Panggang Island, respectively. Therefore, the RF algorithm demonstrated better performance and accuracy compared with the conventional method. While for best pair identification, all bands pair wound did not give the best result. Surprisingly, the usage of green, yellow, and red bands showed good water depth estimation accuracy.
关键词：Satellite-derived bathymetry; Worldview-2; random forest; multiple linear regression
其他关键词：Satellite-derived bathymetry;Worldview-2;Random Forest;Multiple Linear Regression
