Sparseness of the LS-SVM algorithm based on the leave one out cross validation method with Multibeam data

Abstract

The LS-SVM algorithm is non-sparse, so the seabed trend surface constructed by it is impacted by the abnormal sounding training samples. Filtering the sounding training samples before the model construction is necessary, which is to address the non-sparseness of the algorithm. The residual sequence obtained by the leave one out cross validation (LOOCV) method can show the deviation of the water depth estimated by the model from the real depth, the Lagrange multipliers reconstructed based on which can evaluate the contribution of the sounding data to the model construction. The sample data after pruning bring sparseness to the algorithm, and make the model reasonable. The validation experiments using the multibeam sounding measurements show that the LOOCV method can select the sounding training samples that contribute to the model construction a lot to generate reasonable models.