Remote sensing image classification and fusion for terrain reconstruction

This thesis presents a system for three-dimensional terrain reconstruction of a remotely sensed earth scene using multi-sensor image data. This system encompasses techniques of pattern classification and data fusion theory for creating either a realistic terrain model, or a labeled target map of the sensed scene.; The first section of this thesis investigates the problem of scene classification, a common application of pattern recognition in remote sensing data analysis. Several techniques are developed and implemented in this work for the analysis of hyperspectral and multispectral data, incorporating statistical probability models, fuzzy logic, and neural networks. An overview and analysis of the strengths and weaknesses of each classifier is presented, as well as a quantitative comparison of their performance. The results of tests involving both real and simulated data show that the novel fuzzy neural-network classifier developed in this thesis performs the best overall when accuracy and computation time are considered.; Next the work details steps needed to enhance what can be understood about a scene through an effective fusion of multi-sensor imagery. Image registration by an affine transformation achieves a one-to-one correspondence between data sets, permitting direct fusion of data from various modes of surface imagery. Once registered, data fusion of elevation, color-composite images, and classification results produce realistic and accurate scene reconstructions of the sensed area. Approaches for combining multisensor data of various resolutions are explored, and the quality of the resulting models is assessed using human perception and judgment. From testing, high-resolution image data seems to benefit the most from the fusion process, while fusion of low-resolution data gains very little detail over a single source.