Abstract & Overview
Following the dramatic qualitative and quantitative developments of remote sensing systems and digital imaging technologies, numerous kinds of images have been captured, processed and analyzed in several aspects of earth applications. Almost all of image sensors are designed for specific purposes therefore the utilization of remotely sensed data is very diverse. While these datasets are highly useful for some applications, for some others are incomplete, imprecise, and redundant. For example the information contained in hyperspectral and multispectral datasets provide a valuable basis for environmental studies while the low spatial resolution characteristics of these datasets reduce their performances in many applications. For several remote sensing applications, data fusion is the process of alleviating the shortages of data sources. Data fusion is a formal framework that provides means and tools for the alliance of data originating from different sources. It aims at obtaining information of greater quality; the exact definition of greater quality will depend upon the application. Data fusion is usually carried out in three levels of processing: pixel, feature and decision. In an ideal data fusion framework the level of process will be adapted based on data characteristics and the user requirements.This PhD. work explores the concepts and techniques of multi-source remotely sensed data fusion for the purpose of visual spatial resolution enhancement and improvement the accuracy of land cover classification. The investigation datasets were selected from the Earth Observation-One satellite imager which has Hyperion, hyperspectral, ALI, multispectral, and ALI, panchromatic sensors.For the first part of work, 10 of the most common pixel level DF techniques and two innovated ones, named Radon and fanbeam, were investigated and obtained results were compared. Results showed that these two innovated techniques have good abilities in spectral preservation but their ability in spatial quality of fused images was weak. More than the spectral quality of fused images a new technique in spatial quality assessment called coefficient of variance of image autocorrelation was developed. The qualities of fused images were evaluated using this new methodology and its ability was efficient and informative in fused imagery quality assessment. For hyperspectral feature (i.e. spectral band) reduction an innovative methodology called Maximum Spectral and Spatial Information Indicator (MSSI) was introduced and its ability evaluated in comparison to two common feature reduction techniques i.e. Transformed Divergence (TD) and Bhattacharyya Distances (BD). Outcome results showed that MSSI has almost the same ability as for TD and BD. Another evaluated aspect was the comparison of hyperspectral and multispectral datasets fused with panchromatic image. Using the investigated datasets, the measured qualities of fused images showed that in pixel level data fusion MS has higher ability then HS datasets.For the second part of this work, decision level data fusion, two procedures for decision level fusion were innovated. These methodologies called hyperspectral Wavelength Based Decision Fusion (WBDF) and Class Based Decision Fusion (CBDF). The ability of these two new procedures was evaluated for hyperspectral data fusion in land cover mapping accuracy improvement. Final produced maps showed an about 4 % overall accuracy improvement. Also another methodology for decision level data fusion called Multi Classifier Decision Fusion (MCDF) was evaluated which could improve the accuracy of results to about 8%. More than these experimental and technical investigations and innovations the literature of DF and its usability for several applications like environmental, agriculture, mining, urban, medicine studies, etc was studied. Finally, some experimental recommendations for future works in data fusion are drawn.