Automatic Land Cover Classification Using Multitemporal Satellite Images

Land Cover plays a significant role in the earth system science, which reflects the influence of human activities and environmental changes. The remote sensing characterized by the high temporal resolution and the large scale cover has been mainly approach of getting the Land Cover data.The study proposed the automatic classification of multi-temporal remote sensing images using relative radiation normalization in order to overcome the weakness that the training field needed to be repeatedly selected in the traditional Land Cover classification. The method can make the different temporal images have the same radiance or reflectance for the same land cover type, thus the error generated by the repeated selection of training filed can be attenuated. Specially, the supervised classification of Land Cover can be implemented based on the reference image which was selected from the multi-temporal images of the same region, and then the classified characteristics were extracted; the relative radiation correction were carried out for the other images based on the relative radiation normalization of multivariate change detection; the automatic classification can be realized by using the classified characteristics and the maximum likelihood classifier at last. What's more, the study analyzed the characteristics of NDVI for the different land cover types, and the prior probability of maximum likelihood classification can be calculated linking with NDVI, which can improve the classification accuracy.Based on the 2007-08-13 Landsat TM5 (reference image) and 1990-08-30 Landsat TM5(target image), the method of automatic classification was tested. The results showed:1. The radiation difference was attenuated or eliminated by the relative radiation normalization correction, thus the images of different time had the same radiation scale.2. The same classification regulation can be used for the images of the different time after the relative radiation normalization process for the images.3. The classification accuracy of the maximum likelihood method can be increased by 4.69% based on the floating prior probability of NDVI distribution. The innovative approach in this thesis include as following:1. The relative radiometric normalization was applied to reduce the numeric differences between multitemporal images, after that, this thesis develop an automatic classification for multitemporal images based on sharing classification regulation.2. This thesis present a method of modifying the prior probabilities based on NDVI, which can improve the accuracy efficiently.