FEATURE SELECTION IN LIDAR HEIGHT METRICS USING DECISION TREE FOR SVM CLASSIFICATION: APPLICATION IN AGRICULTURAL RESOURCES MAPPING

Hits: 115
Research areas:
Year:
2016
Type of Publication:
In Proceedings
Keywords:
Support Vector Machine, Decision Tree, LiDAR, Feature Selection, Object-based Image Analysis
Authors:
David, Lawrence Charlemagne G.; Sarte, Shydel M.; Laguerta, Joanne Rizza V.; Ballado, Alejandro H.; Jr.,
Abstract:
The fusion and combined use of LiDAR and other remotely sensed data have been widely explored in land cover and land use mapping. With the scarcity of available imagery data, this study focuses on classifying agricultural resources from LiDAR data only. Specifically, using object-based image analysis on non-ground objects, classes with CHM > 0.5m were extracted. It has been found that even without spectral data, LiDAR data alone can be used to map agricultural resources. Initially, confusion among classes was observed in doing Support Vector Machine (SVM) classification because of the presence of too many inputs, e.g. popular forestry metrics and other height information from LiDAR data, which contain irrelevant and redundant data. In addition, processing takes longer time because SVM needs to solve a quadratic programming problem on all the input variables. Thus, feature selection on the input layers and the variables derived from them is essential. In this study, a Decision Tree (DT) was constructed to determine the importance of 56 variables derived from 14 LiDAR height metrics. The input layers were filtered for SVM classification based on the results of the DT. The effectiveness of the method was assessed in three different locations. Through this method, the relevant features derived from the LiDAR height metrics were accurately identified. At least 71% reduction in the number of variables was achieved, as well as 56% reduction in SVM’s training and classification time. With the DT-based feature selection, the overall accuracy and kappa index of agreement were effectively increased, saving time and minimizing inputs for SVM.
Full text: Ab 0492.pdf [Bibtex]
You are here: Home ACRS ACRS Overview Proceedings FEATURE SELECTION IN LIDAR HEIGHT METRICS USING DECISION TREE FOR SVM CLASSIFICATION: APPLICATION IN AGRICULTURAL RESOURCES MAPPING