Two Specific Multiple-Level-Set Models for High-Resolution Remote-Sensing Image Classification

Abstract

This letter adopts level-set methods in order to seek a novel classification strategy in which classification is free of segmentation. A region-driven multiple-level-set (MLS) framework is used to perform very high resolution image classification. Two specific unsupervised classification models are presented. First, from the point of view of feature fusion, an MLS model is suggested by fusing texture features and spectral information (TSMLS model). The model combines spectral information, texture features extracted from the image, and geometrical characteristics of closed curves to achieve effective classification for high-resolution imagery. Second, an alternative MLS model with quadratic image energy (GMMLS model) is presented, which can efficiently integrate the level-set method with Bayesian theory. The model benefits from both the level-set method and Bayesian theory and performs satisfactory classifications. The experiments have demonstrated that our methods can obtain better or similar classification results as compared to support vector machine and Mansouri's method.