Information Retrieval Of Rice Nitrogen Concentration At Different Remote Sensing Levels

With more attentions to environmental perturbations, how to balance the factors of high yield and great quality of crop, and the need to provide or minimize environmental perturbations caused by crop production has become a problem needing conscientious consideration and settlement of the governments, the agriculturist and the environmental workers and researchers in the world. So appropriate N fertilizer management, efficient monitoring of plant N status and proper supply of N to crop is significant to economy and ecology.Remote sensing has a proven ability to provide spatial and temporal measurements of surface properties and has been recognized as a reliable and convenient method for the estimation of various variables related to physiology and biochemistry form remote sensing data at different levels. In recent years, quantitatively remote sensing of the vegetation biochemicals has been greatly improved by the development of hyperspectral technology and the use of multivariate statistical methods, particularly the data mining technique, which has become the hot topics in the studies of quantitatively remote sensing of vegetation biochemicals.This paper discussed around the new important problem of data mining technique, and focused on the quantitatively remote sensing of rice nitrogen concentration from different remote sensing levels. Based on the research of ANN and SVM technique and principal components analysis (PCA), the thesis conducted a systemic study on data mining technique of remote sensing information, from the statistical regression method to ANN and SVM technique, and from method analysis to modeling. Various estimation models were developed using remote data at different remote sensing levels and the precision tests of models were performed. Finally, systemic comparisons of the predict capability of LR methods, ANN methods and SVM technique, as well as the performances of models at different remote sensing levels were made.The main contents and conclusions are following:(1) Quantitative Retrieval of nitrogen concentration using leaf spectral variables In the paper, four methods were adopted for modeling, i.e. linear regression (LR models), back propagation neural network (ANN models), radial basis function network (RBF models), and support vector machine technology (SVM models), and two variables were used as input variables for various models, i.e. raw spectral reflectance (R) and the scores of principal components (PC). The rice nitrogen concentration estimation models for different growth stages and different nitrogen fertilizer levels were established and the precision tests of models were done. At the meantime, the validations of models universality were carried out using field rice dataset and rape dataset. The following conclusions were obtained:Comparisons of nitrogen estimation models at different growth stages showed that the models at grain filling stage and milky ripe stage were prior to models at other growth stages. The PC-RBF model at grain filling stage was the best, the RMSE and REP of the models were 0.151 and 6.816% respectively, and the correlation coefficient between the theoretical and the measured nitrogen concentration was 0.977, which achieved significant level;Comparisons of nitrogen estimation models at different nitrogen fertilizer levels indicated that the models at N1 level were normally better than models at NO and N2 levels. Among various models, the R-LR model performance was the best, which with the RMSE=0.720, REP=25.647 and the correlation coefficient r=0.747;The validation of models universality using field rice dataset showed that the application of various models which based on experimental datasets in the field rice dataset was not only feasible and but also satisfying. Meantime, the validation of models universality using rape dataset indicated that, although it was feasible to apply the models in the rape dataset, the whole performances of various models in the rape dataset was not as good as they were in the field rice dataset. Furthermore, due to the narrow value range of rape nitrogen concentration (maximum is 2.84 mg/g, minimum is 1.07 mg/g), which is significantly lower than rice nitrogen concentration (maximum is 4.82 mg/g, minimum is 0.91 mg/g), the various models were normally overestimated the rape nitrogen concentrations.(2) Quantitative Retrieval of nitrogen concentration using canopy spectral variablesIn this section, the same methods to above section were adopted for the discussions of nitrogen estimation models at canopy level. Conclusions are followings:Comparisons of nitrogen estimation models at different growth stages showed that the models at milky ripe stage and ripe stage were prior to models at other growth stages. The R-ANN model at ripe was the best, the RMSE and REP of the models were 0.746 and 48.147% respectively, and the correlation coefficient between the theoretical and the measured nitrogen concentration was 0.912, which achieved significant level;Comparisons of nitrogen estimation models at different nitrogen fertilizer levels indicated that the various models at three nitrogen fertilizer levels all could achieve satisfying results, which is especially true for models at N1 level. The maximum correlation coefficient between the theoretical and the measured nitrogen was 0.962, the minimum was 0.799.The similar results to leaf level model validation were obtained when using field rice dataset for validation of model universality at canopy level. The theoretical nitrogen concentrations retrieved from R-LR model and R-SVM model were significantly correlated with the measured nitrogen concentration, which with r=0.865 and r=0.854, respectively.Besides, the results using rape dataset for the validation of model universality at canopy level were also similar to those at leaf level, i.e. the whole performances of various models in the rape dataset were not as good as they were in the field rice dataset, and the nitrogen concentrations were generally overestimated.(3) Quantitative Retrieval of nitrogen concentration using TM dataIn this section, the TM data and corresponding nitrogen concentration were used as data source, and the LR modeling method, the RBF method and SVM technology were adopted for modeling. On the basis of correlation analysis between nitrogen concentration and spectral variables, the models based on TM2, TM3, NDVI and RVI were established and the precision tests of models were done. Besides, in order to explore whether the models at leaf and canopy levels could be applied in TM data, the models based on simulated TM variables using leaf and canopy spectral reflectance were constructed and verified using TM data. The results indicated that retrieval of nitrogen information using TM data was not only feasible and satisfying. Comparison of various models' precisions showed that the LR method performed worst among the three kinds of modeling methods as a whole, and the further comparison indicated except for RVI-RBF model performed better than RVI-SVM model, the SVM models based on other three variables were generally performed better than RBF models. Among all models, the TM2-SVM model could achieve the best results, the theoretical nitrogen concentration derived from it were significantly correlated with the measured nitrogen concentration, with the correlation coefficient r=0.751.