Recognition of corn acreage in Jilin province based on mixed pixels decomposition

To improve the application ability of the low resolution remote sensing image in crop area remote sensing monitoring, the corn of Jilin province is selected as the object of study. Based on the remote sensing zoning of corn planting area in the research area, and based on middle and low resolution remote sensing data, the corn area mixed-pixel decomposition model based on neural network function is constructed by using Levenberg-Marquard network optimization algorithms and by taking the smallest root-mean-square error as the condition of iteration termination. The input of the model is the EOS-MODIS NDVI time series data of 18 ten-days during the corn growing season from April to September of 2011, with the spatial resolution of 250m; the output is the corn area value of corresponding resolution. The endmember of corn area comes from the maximum likelihood classification result of the SPOT4 image, with the spatial resolution of 20m. Comparing the decomposition results of mixed pixels from the 3 counties of Dehui district with the the results of the background investigation made in 2011, the precision reached 86.1%. The paper attempts to use high resolution images to directly extract crop endmember. Compared with the endmember extraction methods such as Maximum Noise Fraction (MNF), Pixel Purity Index, Principal Component Analysis, and convex cone geometry theory, etc, it has more intuitive effect, more definite physical significance, stronger operability, and larger business operation potential.