A Study On Land Cover Classification Based On SVM And Multi-source Data Fusion

The research on land use/cover change plays an important role in the global change project, and it will have a great impact on Earth's surface albedo, biochemical cycle, regional water quality and quantity, biodiversity, the primary productivity and the environmental adaptive ability of terrestrial ecosystems, so , timely and accurate acquirement of land use/cover information is significant to governmental decision-making and scientific research. With the diversification of remote sensing platform and the improvement of spatial resolution, remote sensing has already become an indispensable technique in land use/cover research.In the research on the land use/land cover classification, interpretation signature is used as classification feature, and at the same time, it has become an important factor in automatic and accurate classification as the linkage between image information and ecological characters of land use/cover. In addition, the selection of classification method is the precondition to realize accuracy and automatic information extraction of remote sensing either.With the Landsat TM image obtained in June, 2005, as the main data source, land use classification was realized according to the common three-level classification hierarchy of land use/cover investigation. The main researches are:(1) We analyzed the spectrum bands of the original TM image, and selected the best multi-band combination as the classification feature; at the same time, on account of the impact on wetlands and water bodies of vegetation index, we also added vegetation index to the classification feature. We completed the land use classification by using the combination of spectral feature and vegetation index, the results show that the spectral feature is the dominant feature sources, and the combination of bands and vegetation index could get the best classification accuracy, and succeed in separating wetlands from the water bodies.(2) Considering the research of classification methods, SVM was used, and then be compared with other traditional classification methods. The results show that SVM has a unique advantage of resolving the limited-samples and high-dimensional problem, it could have higher classification accuracy, so it could be widely used in the classification of remote sensing images.(3) We succeeded in merging the high-resolution images and multi-spectral imaging by integrated the SPOT Image data, and completed the fusion image's classification by using SVM. The results show that the integration of remote sensing data fusion and the intelligent method of land use classification can get the highest classification accuracy.