Decision level fusion of multi-frequency polarimetric SAR and optical data with Dempster-Shafer evidence theory

Abstract

This paper investigated and tested the multi-source data fusion of multi-frequency SAR data with optical sensor, utilizing the mathematical theory of evidence. In this type of data fusion, the correct definition of uncertainty and mass function is very important. Firstly each mono-source data is independently classified using the Bayesian maximum likelihood classification method and the fusion is processed during the decision making stage. For the representation of support and plausibility, mass of evidence is assigned to the candidate labels associated with the frequencies of neighborhood pixels' decision. Then the mass is weighted again by both the distance effect and the global source-specific reliability. In the combination and fusion stage, unions of classes get same mass values to emboss its uncertainty of decision making. Finally Dempster's orthogonal sum method was used to combine the evidence information and the associated level of ignorance, and then the maximum support rule decides the final land-cover types. Classification accuracies before and after each step of fusion processing are compared with reference training data set. We applied this method to the NASA/JPL polarimetric AIRSAR data and KOMPSAT-1 (KOrea Multi-Purpose SATellite-1) EOC (Electro Optical Camera) panchromatic data and obtained noticeably better land-cover classification results.