Remote Sensing Data Based Crop Growth Parameters Retrieval And Crop Management Zone Delineation Research

Crop growth parameters are critical indicators for monitoring crop growth. In order to guide agricultural production, predict crop yield and make food security policies, it is important to acquire accurate and timely crop growth information. Considering the current agricultural production conditions in China, there exist different degrees of spatial variations in crop growth, not only between different farmlands but also within the same farmland. They result from different field management methods and climate differences. Traditionally, crop growth parameters estimates are based on destructive sampling which is time-consuming and often not applicable to large areas. Remote sensing techniques provide a cost-effective method for making quantitative estimations of crop growth parameters, researching spatial variations in crop growth. Based on the need of crop growth and its spatial variation information, this dissertation mainly focuses on the key problems in crop plant zone extraction based on hyperspectral image, crop growth parameters retrieval and monitoring based on remote sensed data and crop management zone delineation. The major contents and results in this dissertation are as follows: (1) Crop plant zone extraction based on hyperspectral imageIf the training samples are limited in hyperspectral image classification, a reduction in the classification accuracy of the test data is often observed due to the poor generalization of the training results and this effect is known as the curse of dimensionality. In order to solve this problem, dimensionality reduction was conducted using band selection and feature selection. A new band selection method based on scatter matrix was proposed. This method chooses the class separability-scatter matrix as band selection criterion and extends it from two classes to multi-class. The sequential floating forward search is used to realize the band search. The experimental results showed that comparing with JMI, mRMR, CMIM, DISR and JM, the proposed method named as ScatterMatrix got the best overall classification accuracy (90.1%). This result suggests that the proposed method is effective in dimension reduction. Considering that the features extracted by PCA has no guarantee to achieve good classification accuracy, a hybrid feature extraction approach named as PCA_ScatterMatrix was proposed which combined PCA and ScatterMatrix proposed before. The results of experiment indicated that the overall classification accuracy of PCA_ScatterMatrix improved 2.5%, comparing to PCA. In order to obtain more accurate classification result, a method fusing spectral feature and spatial feature, named as SpeSpaVS_ScatterMatrix, was proposed. The ScatterMatrix method is used to extract spectral features. The Gabor spatial features are extracted from the first two principal components. The ScatterMatrix method is conducted on the vector consisting of extracted spectral and spatial features. The experimental results showed that SpeSpaVS_ScatterMatrix method got the best overall classification accuracy (98.4%). Comparing with the method fusing spectral and spatial features in decision level, named as SpeSpaDF, the overall classification accuracy of the proposed method improved 5.4%. It was also found that the overall classification accuracy of the SpeSpaDF method decreased 3.7%, comparing to that of the method only using Gabor spatial features. The results suggest that the overall classification accuracy could not be improved unless the spectral and spatial features are combined in an appropriate way.(2) Crop growth parameters retrieval based on canopy hyperspectral dataMany spectral vegetation indices (SVIs) suffer the saturation effect which limits the usefulness of optical remote sensing for crop LAI retrieval. Besides, leaf chlorophyll concentration and soil background reflectance are also two main factors to influence crop LAI retrieval using SVIs. In order to make better use of SVIs for crop LAI retrieval, it is significant to evaluate the performances of SVIs under varying conditions, In this context, PROSPECT and SAILH models were used to simulate a wide range of crop canopy reflectance in an attempt to conduct a comparative analysis. The sensitivity function was introduced to investigate the sensitivity of SVIs over the range of LAI. This sensitivity function is capable of quantifying the detailed relationship between SVIs and LAI. It is different with the regression based statistical parameters, such as coefficient of determination and root mean square, can only evaluate the overall performances of SVIs. The experimental results indicated that (a) when LAI was no more than three, the variations of soil background had significant negative effects on SVIs. LAI Determining Index (LAIDI), Optimized Soil-adjusted Vegetation Index (OSVI) and Renormalized Difference Vegetation Index (RDVI) were relatively optimal choices for LAI retrieval; (b) when LAI was larger than three, leaf chlorophyll concentration played an important role in influencing the performances of SVIs. Enhanced Vegetation Index 2(EVI2), LAIDI, RDVI, Soil Adjusted Vegetation Index (SAVI), Modified Triangular Vegetation Index 2(MTVI2) and Modified Chlorophyll Absorption Ratio Index 2 (MCARI2) were less affected by leaf chlorophyll concentration and had better performances due to their higher sensitivity to LAI even when LAI reached seven. The analytical results could be used to guide the selection of optimal SVIs for crop LAI retrieval in different phenology periods.In order to ascertain the optimal methods for winter wheat biomass estimation, this study compared the utility of univariate techniques involving narrow band vegetation indices and red-edge position (REP), as well as multivariate calibration techniques involving the partial least square regression (PLSR) analyses using band depth parameters, and the combination of band depth parameters and hyperspectral indices including narrow band indices and REP. Narrow band indices were calculated in the form of normalized difference vegetation index (NDVI) and soil adjusted vegetation index (SAVI) using all possible two-band combinations for selecting optimal narrow band indices. Band depth, band depth ratio (BDR), normalized band depth index, and band depth normalized to area extracted from a red absorption region (550nm-750nm) were utilized as band depth parameters. The results indicated that:(a) Compared with the traditional NDVI and SAVI constructed with bands at 670 nm and 800 nm and REP, the selected narrow band indices (optimal NDVI-like and optimal SAVI-like) produced higher estimation accuracy of the winter wheat biomass; (b) the PLSR models based on band depth parameters produced lower root mean square error, relative to the models based on the selected narrow band indices; and (c) the PLSR model based on the combination of optimal NDVI-like and BDR produced the best estimated result of the winter wheat biomass(R2= 0.84, RMSE= 0.177kg/m2). The results of this study suggest that PLSR analysis using the combination of optimal NDVI-like and band depth parameters could significantly improve estimation accuracy of winter wheat biomass.The correlation analyses between nitrogen nutrient diagnosis indices and wheat grain protein content (GPC) were conducted based on the data obtained in field experiment in 2013. The results of analyses showed that nitrogen nutrient index (NNI) in wheat booting stage had the higher correlation with GPC compared with the other nitrogen nutrient diagnosis indices. The NNI in this stage was retrieved based on double-peak canopy nitrogen index (DCNI) for plant nitrogen concentration estimation, the proposed biomass retrieval method in chapter four and critical nitrogen dilution curve. The NNI retrieval accuracy is high (R2=0.822, RMSE=0.116). Then wheat GPC prediction model was established based on retrieved NNI. This model got high prediction accuracy (R2cv=0.723, RMSEcv=0.883%) in leave-one-out cross validation. The experimental results suggest that it is feasible to using NNI as a bridge between hyperspectral and GPC for GPC prediction. (3) Crop growth parameters retrieval and monitoring based on CASI hyperspectral imageIn order to upscale the crop growth parameters retrieval to regional scale and realize crop growth parameters monitoring over large area, LAI retrieval was conducted based on CASI hyperspectral image. First, crop plant zone was extracted based on the ScatterMatrix method proposed in chapter three. Due to lack of enough ground measured samples, look-up-table (LUT) method was used to retrieve LAI based on PROSAIL radiative transfer model. In order to get more robust results, the average value of multi-solutions was used as the final inversion result. The experimental results indicated the method based on the first 100 solutions got the highest retrieval accuracy (R2=0.551, RMSE=0.496). The results suggest that when lack of enough ground measured samples, LUT inversion using multi-solutions is a good choice for crop growth parameters retrieval. Finally, LAI map over the whole study area was generated for monitoring use. (4) Crop management zone delineationAfter analyzing the spatial variation in soil nutrient parameters and crop growth, crop management delineation was conducted based on soil nutrient data and Quickbird remotely sensed image, respectively. The experimental results showed that the coefficient of variation in each management zone decreased after management zone delineation based on the two data sources. From the view of cost and timeliness, remotely sensed image has more advantages over soil nutrient data. In order to obtain management zones with better spatial coherence, spatial information was involved in management zone delineation. The experimental results showed that fuzzy C means cluster method involving spatial location information (LocationFCM) could generate more spatial coherent management zones, compared with traditional fuzzy C means clusters method and Lark’s method. The results suggest that it is a good choice to delineate management zone based on remotely sensed image using LocationFCM method.