Crop classification using MODIS EVI series in North China

We studied the crop classification in North China using multi-bands MODIS data with time resolution of 8 days and spatial resolution of 500m in year 2007. Vegetation Index EVI was seen as a robust vegetation indicator and its layers were stacked in the time dimension to detect the phenology of various vegetation types including our targets crops. Before classification, a series of data processing steps were performed: first, a comprehensive use of time-frequency analysis methods such as iterated Savitzky-Golay filtering, multi-resolution analysis and energy threshold based algorithm was conducted to reduce noises in the EVI series data; second, crop/non-crop boundary was obtained from the noise reduced data using a binary encoding based algorithm, in which we introduced the concept of "effective width" as the threshold for crop/non-crop vegetation; third, we analyzed the wave structures including starting/ending/maximum curvature/minimum curvature/half wave height points and matched them to the typical crops' phenology in North China to form the training sample sets. The classification methods include ISODATA (unsupervised), SAM (Spectral Angle Mapper), Minimum Distance and SVM (Support Vector Machine). The results showed that the SVM method had the highest accuracy: 82.3% in the double-cropping area and 93.4% in the single-cropping area.