Abstract:

Mixed pixel is a mainly factor that causes inaccurate classification results. In this paper we extracted the planting information of summer harvesting crops (mainly wheat and rape) in Shigu region of Hubei Province from HJ CCD data using the linear mixed model with an improved stepwise regression method. The detail process was as follow: preprocessing (radiometric correction and geometric correction) three HJ CCD remote sensing dates (spatial resolution is 30 m) in 2013, and created hyperspectrallike images with 15 bands using the layer stacking technique. Extracted four spectral characteristics curves of wheat rape water and uncovered land in hyperspectrallike images using EARMSE (endmember average root mean square error) methods combined with pure pixels index and GPS survey. According to time series characteristics of crop phenology, we extracted spatial distribution of summer harvest crop fields using  NDVI threshold. Finally according to four types of endmember spectral characteristics curves and hyperspectrallike images with 15 bands that only include summer harvest crop, we used the Linear Mixture Spectrum Analysis method that replace multiple regression with improved wise step, and calculated abundance value of four types of endmember in each mixed pixel. The calculated values was compared to ground truth data that were derived from ZY03 image (spatial resolution is 21 m) using maximum likelihood classification method in subarea that area is 522 square kilometer. The result shows that the error rate of rape area is 67% and that of wheat is 95%, indicating that this reversing method is accurate in Shuhu region. The paper provids a new method with high precision to extract crop planting information with low spatial resolution data, especially in southern part of China where most of pixels are mixed pixels