Author(s): Ankita Misra, Zoran Vojinovic, Balaji Ramakrishnan, Arjen Luijendijk, Roshanka Ranasinghe
Linked Author(s): Ankita Misra
Keywords: Near-shore bathymetry, Landsat 8 OLI, Landsat 7ETM, SONAR, Support Vector Machine.
Abstract: This paper demonstrates the capability of the Support Vector Machine (SVM) technique to obtain Satellite Derived Bathymetry (SDB) maps of near-shore regions using freely available Landsat 7 ETM and Landsat 8 OLI imagery of 30 m resolution (medium). The Support Vector Machine is a non- linear machine learning approach that utilizes the Radial Basis kernel function; and the other training factors such as the smoothing parameter, penalty parameter C and insensitivity zone ? are selected and tuned based on the learning (i. e. training) process. The approach is applied here with available bathymetric SONAR data for two regions, first, one of the Dutch Wadden sea inlets, the Ameland inlet, which has a significant mean tidal range of 2m and the second, the Dutch territory of Sint Maarten Island which has a rather insignificant mean tidal range of less than 10 cm. A comparison of the SVM retrieved depths with the available in-situ data is presented for the regions and the results highlight the ability of this approach to estimate near-shore bathymetry accurately. For Sint Maarten Island, 20. 38% of training data is used as support vectors and the overall error obtained is 8. 26%. The inter-comparisons between the sonar and SDB depth values shows r2 = 0. 97 with RMSE of 0. 69m. In the case of Ameland inlet, the number of support vectors obtained in the training phase is 21. 56% of the training data and the overall errors during training and test phases are 6. 88% and 6. 48%, respectively. The comparison of the predicted and actual depths shows r2 = 0. 98 and a RMSE of 0. 64m. These results indicate that the SVM technique performs well for shallow depths and can be used effectively for deriving accurate and updated bathymetric maps (at the temporal resolution of satellite return periods of ~2 weeks)
Year: 2017