Composited Classification Methods Based On Multi-resolution Satellite Images

In recent years, the integration and composited analysis of multi-resource satellitedata has become one of the most important techniques in the remote sensing field. Forlarge scale land covering classification, it's no surprise that low resolution data hasworse performance due to the mixed-pixel problem, and high resolution data with widecovering range has more limitations such as long period of acquiring cycle, high dataand processing cost. Facing the above problems, a new composited classificationframework for the multi-resolution satellite data with in the same area and differentcovering ranges has been studied in this dissertation. To study the multi-to-single spatialcorrespondence between high resolution image and low resolution image,nonparametric "real" likelihood distribution estimation is adopted and "real" likelihoodfeatures for low resolution pixels are extracted based on cloud theory. An sub-pixellevel enhanced composited classification method based on multi-kernel non-linearregression model and a context level enhanced composited classification method basedon Conditional Random Fields model are proposed. Our experiments on different sets ofsatellite data show that the proposed methods can greatly improve the accuracy for largescale land covering classification applications. The major contributions of thisdissertation are listed as following:A novel integrated frame of the composited classification method at algorithmlevel is proposed and realized. To solve the problems of tradition compositedclassification method based on linear regression model, an enhanced compositedclassification method based on multi-kernel non-linear regression model is proposed.Several key techniques such as multi-to-single spatial relation between the higherand lower resolution satellite data, generation and description for the likelihooddistribution scatter plot and real-likelihood function extraction algorithms are putforward. And the relationships between real-likelihood function models andclassification accuracy, the relationships between the extending principles andclassification accuracy are analyzed in details.A composited classification model which integrating spatial contextual informationbetween pixels and multi-to-single spatial correspondence is proposed. The sequences of "real" likelihood features, which represent relations between spectrum and landcovering types, is integrated into the classifier with the spatial contextual informationbetween pixels by defining two types of potential functions. CRFs based classifieroffers a robust and accurate framework which can support multiple features andrepresents the special continuity of land covering.Using various data combination and different kinds of ground truth data, theaccuracy of the proposed compounded classification algorithms has been analyzed, andthe results show that the proposed method can greatly improve the accuracy for largescale land covering classification applications.The related functional modules and simulation software are developed and tested.An operational process for the Crop Condition Monitoring System is developed.These researches have provided new efficient analysis methods for multi-resourcesatellite data and will extend the applications of composited classification to someextent.