Object-based land use and land cover mapping from LiDAR data and orthophoto application of decision tree-based data selection for SVM classification

Abstract

The major disadvantage of Support Vector Machine (SVM) happens in its training phase as it requires to solve a quadratic programming problem, making computation very costly. With the integration of LiDAR data and high spatial resolution orthophoto, more input data layers are available for object-based Support Vector Machine classification. Initially, confusion among classes arises because of the presence of irrelevant and redundant data layers. Hence, this study applies Decision Tree (DT), a popular data mining technique, as a pre-classification process in SVM to select the relevant features from the input variables. We assessed the effectiveness of seven vegetation indices, two vegetation index combinations and 14 LiDAR height metrics for mapping agricultural resources in Calatagan, Batangas. We were able to filter the input variables and subsequently achieve at least 73% reduction of training features. With the DT-based feature selection, we were able to reduce the number of input features as well as make the SVM training and classification time shorter by more than 90%. Importantly, the overall accuracy and kappa index of agreement both increased when DT-based SVM was used in contrast with using all the variables for SVM classification.