Data fusion of multi-source satellite data sets for cost-effective disaster management studies

Abstract

A common approach to multisource data fusion is to aggregate the information in a stacked vector and treat it as a unique dataset. The statistical classifiers used in these data fusion approaches are also transformed by integrating the contextual information from neighboring pixels, to improve the accuracy of a fuzzy-logic-based fusion scheme. A decision level fusion approach was developed by Gokaraju et. al, 2012, which combines statistical methods and machine learning techniques. Here, each data sample is integrated through separate classifiers such as empirical methods and support vector machines (SVMs) and then used a Probabilistic Neural Network (PNN) to fuse the decisions for a unified consensus decision. The data fusion approach consists of either pixel-level or feature-level data fusion in combination with machine learning techniques for classification. The intermediate results of the disaster management studies, such as levee land-slide and tornado debris assessment using data fusion techniques, are presented in this paper. For levee landslide studies, we used the multi-temporal datasets of air-borne synthetic aperture radar sensor (UAVSAR). For Tornado disaster studies, we used multi-source and multi-temporal datasets of both synthetic aperture radar sensor (RADARSAT-2) and multispectral sensor (RapiEye) datasets. The results of data fusion approach outperformed the non-data fusion techniques in both studies with kappa accuracies of 82.8% and 72%.