Remote Sensing Image Classification Based On Data Fusion

As an important application of pattern recognition in the remote sensing area, land cover classification has been successfully applied to both military and civil fileds. Traditional methods for remote sensing image classification are implemented based on a single spectrum, which makes it difficult to satisfy different practical requirements. In this paper, the idea of information fusion is introduced to seek higher classification accuracy, and three new classification algorithms are presented based on AdaBoost Algorithm, Evidence Theory and Fuzzy Integral, respectively.Firstly, a remote sensing image classification method is presented based on AdaBoost Algorithm, which is used to boost the performance of basic K-means classifier. To solve the resampling of patterns, a weighted version is provided. Classification results produced by the boosted K-means present an obvious advantage on the elimination of isolated points and recognition of slim objects, when compared with the basic K-means.Secondly, an Evidence Theory(DS)-based remote sensing image classification method is proposed, which combines the results using features in HCI color space and inverse-gradient spatial information. A multi-ratio evaluation for the basic probability function is employed here for the fusion problem of unreliable sources. This evaluation strategy achieves a higher performance than the single-ratio one because recognition accuracy for a given pattern of each classifier is considered.The land cover classification experiments prove the validity of DS in the fusion of information from multi-features.Finally, a Fuzzy Integral-based remote sensing image classification method is proposed. In this new method, a search algorithm for the parameter of fuzzy measure X is designed and Choquet fuzzy integral is selected to combine the classification results using wavelet package textures of R, G, B channels. Experimental results show that the new classification method integrates the advantages produced using textures of three channels and generates a satisfying mapping.