Self-enhanced SVM Extraction of Building Objects from High Resolution Satellite Images

Abstract

A self-enhanced SVM (Support Vector Machines) building detection scheme is discussed. The scheme was designed for 1-metre resolution satellite imagery analysis. The scheme is a learning based segmentation without any prior prepared training data set. In the initial stage, an adaptive two-dimension Otsu algorithm is adopted to segment the image primarily into buildings and non-buildings. Then the segmented regions are modeled as second order GMRF (Gaussian Markov Random Fields), and a six element characteristic vectors are extracted. In the final stage, a SVM classifier is trained on the characteristic vector and region label, then the trained SVM classifier re-segment the image on pixel to get an enhanced result. Experiment shows that the system is efficient and robust.