Retrieval of biophysical parameters from multi-sensoral remote sensing data, assimilated into the crop growth model CERES-Wheat

Abstract

This study investigated the possibilities and constraints for an integrated use of a crop growth model (CERES-Wheat) and earth observation techniques. The assimilation of information derived from earth observation sensors into crop growth models enables regional applications and may also help to improve the profound knowledge of the different involved processes and interactions. Both techniques can contribute to improved use of resources, reduced crop production risks, minimised environmental degradation, and increased farm income.<br /> Up to now, crop growth modelling and remote sensing techniquices mostly have been used separately for the assessment of agricultural applications. Crop growth models have made valuable contributions to, e.g., yield forecasting or to management decision support systems. Likewise, remote sensing techniques were successfully utilized in classification of agricultural areas or in the quantification of vegetation characteristics at various spatial and temporal scales. Multisensoral remote sensing approaches for the quantification biophysical variables are rarely realized. Normally the fusion of the data sources is based on the use of one sensor for classification purposes and the other one for the extraction of the desired parameters, based on the map classified previously. Pixel-based fusions between multispectral and SAR data is seldom realised for the assessment of quantitative parameters.<br /> The integration of crop growth models and remote sensing techniques by assimilating remotely sensed parameters into the models, is also still an issue of research. Especially, the integration of, e.g., multi-sensor biophysical parameter time-series for the improvement of the model performance, might feature a high potential.<br /> The starting point of the presented study was the question, if it is possible to derive the values of important crop variables from various remote sensing data? For the retrieval of these quantitative parameters by the use of various multispectral remote sensing sensors, intercalibration issues between the different retrieved vegetation indices had to be taken into account, in order to assure the comparability. Features influencing the vegetation indices are, e.g., the sensor geometry (like viewing- and solar-angle), atmospherical conditions, topography and spatial or radiometric resolution. However, the factors taken into account within this study are the spectral characteristics of the different sensors, like band position, bandwidth and centre wavelengths, which are described by the relative spectral response functions. Due to different RSR functions of the sensor bands, measured spectral differences occur, because the sensors record different components of the reflectance’s spectra from the monitored targets. These are then also introduced into the derived vegetation indices. The chosen cross-calibration method, intercalibrated the assessed Normalized Difference Vegetation Index and the Weighted Difference Vegetation Index between the various sensor pairs by regression, based on simulated multispectral sensors. Differences between the various assessed remote sensing sensors decreased form around 7% to below 1%. The intercalibration also had a positive impact on the later biophysical retrieval performance, producing sounder retrieval results.<br /> For the retrieval of the biophysical parameters empirical and semi-empirical models were assessed. The results indicate that the semi-empirical CLAIR model outperforms the empirical approaches. Not only for the Leaf Area Index retrieval, but also in the cases of all other assessed parameters. Concerning the other remote sensing data type used, the SAR data, it was analysed what potential different polarizations and incidence angles have for the extraction of the quantitative parameters. It became obvious that especially high incidence angles, as provided by the satellite Envisat ASAR, produce sounder retrieval results than lower incidence angles, due to a smaller amount of received soil signal. In the context of the assessed polarizations, sound results for the VV polarization could only be achieved for the retrieval of fresh biomass and the plant water content. For the ASAR sensor modelling fresh biomass and LAI using the HV polarization or the dry biomass using the ratio (HH/HV) was appropriate.<br /> As roughness aspects also have an influence on the retrieval performance from biophysical parameters using SAR data, the impact of soil surface and vegetation roughness was additionally considered. Best results were achieved, when also considering roughness features, however due to the need of regional modelling it is more appropriate not to consider them.<br /> For the calibration and re-tuning of crop growth models information about important phenological events such as heading/flowering is rather important. After this stage reproductive growth begins, whereby the number of kernels per plant is often calculated from plant weight at flowering and kernel weight is calculated from time and temperature available for dry matter distribution. By the use of the SAR VV time-series this important stage could be successfully extracted. Further methods for pixel-based fused biophysical parameter estimations, using SAR and multispectral data were analysed. By this approach the different features, being monitored of the two systems, are combined for sounder parameter retrieval. The assessed method of combining the multi-sensoral information by linear regression did not bring sound results and was outperformed by single sensor use, only taking into account the multispectral information. Only for the parameter fresh biomass, modelling based on the NDIV and the ASAR ratio slightly outperformed the single sensor modelling approaches. The complex combined modelling by the use of the CLAIR and the Water Cloud Model featured no valid results. For the combination, by using the CLAIR model and multiple regression slight improvements, in contrast to the single multispectral sensor use, were achieved. Especially, during late phenological stages, the assessed VV information improved the modelling results, in comparison to only using the CLAIR model.<br /> All the findings could finally be successfully applied for regional estimations. Only the roughness features could not be applied, due to the fact, that it is hard to regionally assess this needed model input parameter. Regional parameter on the basis of remote sensing data, is the major advantage of this technique, due to the large spatial overview given.<br /> The second main question was, if it is possible to integrate the crop variables gained from multisensoral data into a crop growth model, increasing the final yield estimation accuracy. Thus far, beneficial linkages between both techniques have been often limited to land use classification via remote sensing for choosing the adequate model and quantification of crop growth and development curves using biophysical parameters derived from remote sensing images for model calibration. Only a few studies actually considered the potentials of remote sensing for model re-initialization of growth and development characteristics of a specific crop, as the here studied winter wheat. Overall, the integration of remotely sensed variables into the crop growth model CERES-Wheat led to an improved final yield estimation accuracy in comparison to an automatic input parameter setting. The assessed final yield bias for the automatic input parameter setting summed up to 6.6%. When re-initializing the most sensitive input parameters (sowing date and fertilizer application date) by the use of remotely sensed biophysical variables the biases ranged from 0.56% overestimation to 5.4% understimation, in dependence of the data series used for assimilation. Whereby, it was assessed that the combined dense data series, considering SAR and multispectral information, slightly outperformed the performance of the full multispectral data series. However, when analysing the assimilation of the multispectral data series in further detail, it became clear that the actually information from the phenological stage ripening declines the modelling performance and thus the final yield estimation accuracy. When neglecting the information from this phenological stage the reduced multispectral data series performed as sound as the dense data series containing SAR and multispectral information. Thus, when the appropriate phenological stages are monitored by multispectral data, additional SAR information does not lead to a model improvement. However, when important dates are not monitored by multispectral images, e.g., due to cloud coverage, the additionally considered SAR information was not able to appropriatly fill these important multispectral time gaps. They even had a more negeative influence on the modelling performance. Overall, the best results could be obtained by assimilating a multispectral data series, covering the crop development during the important phenological stages stem elongation and flowering (without ripening stage), into the CERES-Wheat model.<br /> Finally, the integration of remote sensing data in the point-based crop growth model allowed it‘s spatial application for prediction of wheat production at a more regional scale. This approach also outperformed another evaluated method of direct multi-sensoral regional yield estimation. This study has demonstrated that biophysical parameters can be retrieved from remote sensing data and led, when assimilated into a crop growth model, to an improved final yield estimation. However, overall the SAR information did not really have a significant positive effect on the multi-sensoral biophysical parameter retrieval and on the later assimilation process. Thus, overall SAR information should only be considered, when multispectral data acquisitions are tremendously hampered by cloud coverage. The assessed assimilation of remote sensing information into a crop growth model had a positive effect on the final yield estimation performance. The analysed method, combining remote sensing and crop growth model techniques, was succsessfully demonstrated and will gain even more importance in the future for, e.g., decision support systems fine-tuning fertilizer regimes and thus contributing to more environmentally sound and sustained agricultural production.