A.J.W. de Wit

A review on reflective remote sensing and data assimilation techniques for enhanced agroecosystem modeling

Abstract During the last 50 years, the management of agroecosystems has been undergoing major changes to meet the growing demand for food, timber, fibre and fuel. As a result of this intensified use, the ecological status of many agroecosystems has been severely deteriorated. Modeling the behavior of agroecosystems is, therefore, of great help since it allows the definition of management strategies that maximize (crop) production while minimizing the environmental impacts. Remote sensing can support such modeling by offering information on the spatial and temporal variation of important canopy state variables which would be very difficult to obtain otherwise. In this paper, we present an overview of different methods that can be used to derive biophysical and biochemical canopy state variables from optical remote sensing data in the VNIR-SWIR regions. The overview is based on an extensive literature review where both statistical–empirical and physically based methods are discussed. Subsequently, the prevailing techniques of assimilating remote sensing data into agroecosystem models are outlined. The increasing complexity of data assimilation methods and of models describing agroecosystem functioning has significantly increased computational demands. For this reason, we include a short section on the potential of parallel processing to deal with the complex and computationally intensive algorithms described in the preceding sections. The studied literature reveals that many valuable techniques have been developed both for the retrieval of canopy state variables from reflective remote sensing data as for assimilating the retrieved variables in agroecosystem models. However, for agroecosystem modeling and remote sensing data assimilation to be commonly employed on a global operational basis, emphasis will have to be put on bridging the mismatch between data availability and accuracy on one hand, and model and user requirements on the other. This could be achieved by integrating imagery with different spatial, temporal, spectral, and angular resolutions, and the fusion of optical data with data of different origin, such as LIDAR and radar/microwave.

Efficiency and accuracy of per-field classification for operational crop mapping

A crop map of The Netherlands was created using a methodology that integrates multi-temporal and multi-sensor satellite imagery, statistical data on crop area and parcel boundaries from a 1 : 10 000 digital topographic map. In the first phase a crop field database was created by extracting static parcel boundaries from the digital topographic map and by adding dynamic crop boundaries using on-screen digitizing. In the next phase the crop type was determined from the spectral and phenological properties of each field. The resulting crop map has an accuracy larger than 80% for most individual crops and an overall accuracy of 90%. By comparing cost and man-hours it was demonstrated that per-field classification is more efficient than per-pixel classification and decreased the effort for classification from 1500 to 500 man-hours, but the effort for creating the crop field database was estimated at 2300 man-hours. The use of image segmentation techniques for deriving the crop field database was discussed. It was concluded that image segmentation cannot replace the use of a large-scale topographic map but, in the future, image segmentation may be used to map the dynamic crop boundaries within the topographic parcels.

Crop growth modelling and crop yield forecasting using satellite-derived meteorological inputs

Distributed crop simulation models are typically confronted with considerable uncertainty in weather variables. In this paper the use of MeteoSat-derived meteorological products to replace weather variables interpolated from weather stations (temperature, reference evapotranspiration and radiation) is explored. Simulations for winter-wheat were carried for Spain, Poland and Belgium using both interpolated and MeteoSat-derived weather variables. The results were spatially aggregated to national and regional level and were evaluated by comparing the simulation results of both approaches and by assessing the relationships with crop yield statistics over the periods 1995–2003 from EUROSTAT. The results indicate that potential crop yield can be simulated well usingMeteoSat-derivedmeteorological variables, but thatwater-stress hardly occurs in thewater-limited simulations. This is caused by a difference in reference evapotranspiration whichwas 20–30%smaller in case ofMeteoSat. As a result, the simulations usingMeteoSat-derivedmeteorological variables performed considerably poorer in a regression analyseswith crop yield statistics on national and regional level. Our results indicate that a recalibration of the model parameters is necessary before the MeteoSatderived meteorological variables can be used operationally in the system.

Using MERIS on Envisat for land cover mapping in the Netherlands

This paper describes the results of a feasibility study to test the usefulness of MERIS for land cover mapping. The Netherlands was used as a test site because of its highly fragmented landscape. Results showed that the geometric and radiometric properties of the studied MERIS images of the Netherlands are suitable for land applications. Calculation of principal components and correlation coefficients revealed that the 15 MERIS bands provided a lot of redundant spectral information. For land applications, information came from the visible part of the spectrum on the one hand and from the near‐infrared part on the other hand. In addition, the red‐edge slope of the reflectance curve (in particular MERIS band 9 at about 708 nm) provided supplementary information. The Dutch land use database LGN5 was used as a reference for classifications in this study after aggregation from 25 m to 300 m and recoding to 7 relevant land cover classes. For land cover classification best results in terms of classification accuracies were obtained for the image of 14 July 2003. For the seven land cover classes selected the overall classification accuracy was 67.2%. A multitemporal classification did not improve the overall classification accuracy.

Using quadtree segmentation to support error modelling in categorical raster data

This paper explores the use of quadtree segmentation of a land-cover map to improve error modelling by (1) accounting for variation in classification accuracy among differently sized homogeneous map regions and (2) improving the statistical properties of map realizations generated by sequential indicator simulation (SIS). The latter was accomplished by locating the first simulation nodes—which affect many subsequent local simulations—within the largest quadtree leaves. These represent the largest homogeneous and hypothetically most accurately classified areas on the map. A case study showed that, indeed, the overall accuracy of the land-cover map increased with the quadtree leaf size, ranging from 67% for single cells in heterogeneous areas to 98% for homogeneous blocks of 256 cells. A map of the prior probability of each land-cover class was prepared on the basis of quadtree level-specific confusion matrices. Next, two unconditional SIS algorithms were used to generate sets of 50 realizations of the map, thereby accounting for spatial continuity of the residuals between indicator-transformed reference data and the priors. The proposed quadtree-guided SIS outperformed the more common multiple-steps method as judged by the reproduction of target proportions of map categories, class-specific accuracy levels and variogram reproduction.

Representing Uncertainty in Continental-Scale Gridded Precipitation Fields for Agrometeorological Modeling

This work proposes a relatively simple methodology for creating ensembles of precipitation inputs that are consistent with the spatial and temporal scale necessary for regional crop modeling. A high-quality reference precipitation dataset (the European Land Data Assimilation System (ELDAS)) was used as a basis to define the uncertainty in an operational precipitation database (the Crop Growth Monitoring System (CGMS)). The distributions of precipitation residuals (CGMSELDAS) were determined for classes of CGMS precipitation and transformed to a Gaussian distribution using normal score transforma- tions. In cases of zero CGMS precipitation, the occurrence of rainfall was controlled by an indicator variable. The resulting normal-score-transformed precipitation residuals appeared to be approximately multivariate Gaussian and exhibited strong spatial correlation; however, temporal correlation was very weak. An ensemble of 100 precipitation realizations was created based on back-transformed spatially correlated Gaussian residuals and indicator realizations. Quantile-quantile plots of 100 realizations against the ELDAS reference data for selected sites revealed similar distributions (except for the 100th percentile, owing to some large residuals in the realizations). The semivariograms of realizations for sampled days showed considerable variability in the overall variance; the range of the spatial correlation was similar to that of the ELDAS reference dataset. The intermittency characteristics of wet and dry periods were reproduced well for most of the selected sites, but the method failed to reproduce the dry period statistics in semiarid areas (e.g., southern Spain). Finally, a case study demonstrates how rainfall ensembles can be used in operational crop modeling and crop yield forecasting.

Grain Yield Observations Constrain Cropland CO2 Fluxes Over Europe

Carbon exchange over croplands plays an important role in the European carbon cycle over daily to seasonal time scales. A better description of this exchange in terrestrial biosphere models -- most of which currently treat crops as unmanaged grasslands -- is needed to improve atmospheric CO~2~ simulations. In the framework we present here, we model gross European cropland CO~2~ fluxes with a crop growth model constrained by grain yield observations. Our approach follows a two-step procedure. In the first step, we calculate day-to-day crop carbon fluxes and pools with the WOrld FOod STudies (WOFOST) model. A scaling factor of crop growth is optimised regionally by minimizing the final grain carbon pool difference to crop yield observations from the Statistical Office of the European Union. In a second step, we re-run our WOFOST model for the full European 25 x 25 km gridded domain using the optimized scaling factors. We combine our optimized crop CO~2~ fluxes with a simple soil respiration model to obtain the net cropland CO~2~ exchange. We assess our model’s ability to represent cropland CO~2~ exchange using 40 years of observations at 7 European FluxNet sites and compare it with carbon fluxes produced by a typical terrestrial biosphere model. We conclude that our new model framework provides a more realistic and strongly observation-driven estimate of carbon exchange over European croplands. Its products will be made available to the scientific community through the ICOS Carbon Portal, and serve as a new cropland component in the CarbonTracker Europe inverse model.

Stochastic simulation of large grids using free and public domain software

This paper proposes a tiled map procedure enabling sequential indicator simulation on grids consisting of several tens of millions of cells, without putting excessive memory requirements. Spatial continuity across map tiles is handled by conditioning adjacent tiles on their shared boundaries. Tiles across the area can be characterized by dissimilar models of spatial continuity (semi-variograms) thus relieving the requirement of a global stationarity decision. Additionally, the approach provides a simple mechanism for reseeding the pseudo random number generator. Implementation of the algorithm involved small modifications to a GSLIB program and Bash and awk scripting. The software was stable on several platforms, including 32-bit systems with a 4Gb memory addressing limit. In an experiment we simulated 25 realizations of a 11,274x13,000 grid representing local uncertainty in the Dutch land cover at 25m resolution. With the objective of mimicking the typical absence of well-distributed hard reference data, the simulations were only conditioned on local prior class probabilities and semi-variograms. Output was evaluated on the basis of reproduction of target levels of (1) cover type proportions, (2) overall class label accuracy and (3) spatially averaged local Shannon entropy. As expected, the realized statistics differed significantly from the target levels. However, the differences were consistent over the borders and the insides of map tiles. Thus, they did not result from the tiled map procedure but rather should be attributed to the used semi-conditional sequential indicator simulator. The current implementation can easily be adapted to accept other simulation algorithms.