Improving MODIS Land Cover Classification Using NDVI Time-Series and Support Vector Machine in the Poyang Lake Basin, China

MODIS data play an important role in global and regional environment and resources researches. Remote image classification often fails to separate a large number of land cover classes that may present similar spectral reflectance. To improve the classification accuracy in such situations, multi-temporal data has been proved as valuable auxiliary information. In this paper, we used 250m MODIS/NDVI time-series datasets as main data source and reconstructed the NDVI time series datasets with higher quality by Harmonic Analysis algorithm. The results of NDVI and MODIS spectral band1-7 reflectance were both used as classification input data. With Support Vector Machine, a recently proposed classification method, we obtained the land cover classification results of the Poyang Lake Basin in 2009. Using field observation data, we validated the land cover results. In the matrix of land cover classification error, the overall accuracy was 82.21%, and the kappa coefficient was 0.779. For the MODIS land cover product (MCD12Q1), it was only 66.71% and 0.5846 respectively. Compared with the land cover product data, the spatial pattern of the classification was similar. However, the classification result was essentially better at the aspects of land cover identification and spatial accuracy. It can be conclude that the method of data processing and classification used in this paper is feasible in the study of land cover classification in the research region.