Land cover classification with MODIS data in China

摘要

In this paper, Moderate Resolution Image Spectroradiometer (MODIS) data with high spectral and temporal resolutions were used-as input parameters for Chinese-regional-scale land cover classification. Firstly, Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI) and Nonnalized Difference Soil Index (NDSI) were calculated as input spectral features relies on an annual time series of twelve MODIS 8-day composite reflectance images (MOD09) acquired during the year of 2007. The monthly EVI was produced by the maximum value composite; the three indices were added in the image to form a 10-spectral-bands image. In order to reduce the input feature space dimension, we resort to the mean Jeffries-Matusita distance as a statistical separability criterion to select the best spectral feature combination according -to their ability of separating the land cover classes. Once we achieved, the monthly best combination spectral bands were dealt with Principal Component Analysis (PCA) method and their first three principal components were used as input parameters for decision tree classification. The result showed that the best combination of spectral bands added temporal information as input parameters can reach a certain high classification accuracy (81.16%) at moderate spatial scales without other accessorial data.