In this study, an approach is proposed for determining the depths of shallow waters from multispectral satellite images. This approach is a hybrid system that uses a Random Forest Ensemble (RFE) technique for Neural Network (NN) and Support Vector Regression (SVR) algorithms to improve accuracy of bathymetry estimations by reducing both errors and prediction variance. The study was performed over two different coastal areas: the first is the port of Alexandria, Egypt, which has a low turbidity silt–sand bottom area with a depth range of 10.5 m; the second area is Shiraho, Japan, which is a heterogeneous bottom composition coral reef area with a varied depths range up to 12.5 m. Landsat-8 and Quickbird satellite images were used for these two study areas, respectively, as examples of low- and high-resolution images. These images were corrected for both atmospheric and sun-glint errors. The results were validated using echosounder field points. Four reflectance values of green, red, blue divided by red, and green divided by red bands were assigned to reference point locations. These values were used for algorithm training and testing processes. The SVR and NN output values were combined using RFE. In order to evaluate the accuracy of the BE approach, we compared its outputs to that of the NN and SVR approaches. All of the results were evaluated using RMSE and R2 values. The RFE approach was found to yield a RMSE of 0.64 and 0.90 m for the two areas, resulting in an almost 20 and 10 cm improvement. These results suggest that the proposed RFE approach outperforms both the NN and SVR algorithms, when they are used individually. |
