Information Extraction Of Rape Nitrogen Concentration Using Remotely Sensed Data At Different Levels

The purpose of this study is to use traditional regression methods and AI methods for remote sensing estimation and inversion of rape nitrogen concentrations to obtain more convenient and effective methods for nitrogen diagnose of rape in leaf, canopy, and satellite platform. The AI methods conclude the back-propagation neural network (BP), radial basis function neural network, (RBF), support vector machine (SVM). The estimation models of rape nitrogen concentration based on different methods provide the basis for nitrogen concentration inversion at large scales. It has an important practical significance for inversion of rape yield in large scale. The research results and discussions are as follows. (1) Based on the analysis of the hyperspectral data of rape at leaf and canopy levels, obtained the change trends of rape spectral under different nitrogen levels, different developmental stages, and different band position and width. The results show there is a reflectance peak around in 550 nm, which is the reason of green for rape leaf. There is a sudden increase reflectance in 700 nm, and form highly reflective platform in near-infrared region. For different levels of nitrogen, the spectral characteristics of rape leaf and canopy are significant differences in the near-infrared band. The red edge position of rape leaf has a "red-shift" phenomenon with rise of the nitrogen levels. There are 'two peak' and 'platform of red edge' phenomena for the red edge of canopy spectra of rape. The position of red edgeλ_red is between 690nm and 720nm. There were 'red shift' and 'blue shift' phenomena for the slope of red edge Dλ_red and area of red edge S_red for the canopy spectra, which is obvious different with rice, corn and other crops. (2) The validation results of rape nitrogen concentration remote sensing estimation at leaf level show that, AI methods improve estimation capability of the regression methods. The AI-based estimation model using hyperspectral data eliminated the platform effect in the regression method, and expanding the scope of estimation. The RBF neural network gets the best validation result among the three AI methods. (3) The validation results of rape nitrogen concentration remote sensing estimation at canopy level show that, AI methods improve estimation capability of the regression methods. The BP neural network gets the best validation result among the three AI methods. (4) The results of the rape nitrogen concentration remote sensing inversion based on satellite remote sensing images show that, due to the large number of mixed-pixel presence in the study area, the most important issue in the rape acreage extraction is the accuracy of the classification results and certification error caused by the mixed-pixel. Five kinds of hard classification for the capacity of the classification of mixed pixel, from SVM, ARTMAP, KNN, BPN, and MXL were lower. Purity of pixel changes and the purity index of various classifications changes have the same trend, that is, the higher purity of one category, the higher total accuracy of this category. Multiple classifiers voting law with the classification can significantly improve the total accuracy of the classification. Fuzzy classification used for the study area TM image has higher accuracy than the hard classification alone as a whole. Because of its great proportion of mixed pixel, the fuzzy classification can be avoided the hard classification which hard to choose suitable and adequate pure pixel for classification training. TM band reflectance based rape nitrogen concentration estimation model prior to the TM vegetation indices based rape nitrogen concentration estimation model. It confirmed that ensure the accuracy of the geometric correction and image reconstruction premise spectrum, it can use the TM vegetation indices based rape nitrogen concentration estimation model for large-scale rape nitrogen concentration mapping.In conclusion, this study has some new progress or innovation: the systematic study of rape nitrogen concentration remote sensing estimation model at different levels; The AI technology have been introduced to the entire process of rape nitrogen concentration remote sensing estimation. The application of fuzzy classification and the mixed pixel influence on the classification introduce to the rape covered area extraction, and given a new and effective exploration.