A Classification Strategy for Multi-Sensor Remote Sensing Data. An analysis and implementation of Meta-Gaussian classification schemes

Abstract

In integrated remote sensing, one of the objectives is to create reliable services by combining information from various data sources. The combination of multiple data sources is often denoted "data fusion", and is a topic that has high interest in remote sensing applications. In this thesis, we devise a classification strategy for multi-sensor remote sensing data, based on the strategy presented in the paper "On the Combination of Multisensor Data Using Meta- Gaussian Distributions" \cite{storvik}. The classification method uses data fusion through a transformation of variables into a multivariate Meta-Gaussian distribution, and correct assumptions or estimates of the marginal probability density functions is an important key in this transform. We found that using general parametric probability density functions, or kernel estimates were valid in a supervised classification setting, with no need to specify individual marginals based on the true underlying distribution. Further, we found that classification based on the Meta-Gaussian function, using transformed variables, surpassed that of a standard multivariate Gaussian function. Unsupervised classification based on the same strategy was implemented in a generalized mixture decomposition algorithmic scheme framework. Current results are positive, and indicate that this method has potential when it comes to combining multi-sensor remote sensing data.