Remote sensing

Earth Observation Information

The remote sensing measured parameter is the annual biomass productivity (dry matter) for each crop (100 kg/ha), at agricultural circumscriptions and regional scales.

The estimated dry matter production is derived from NOAA-AVHRR and/or SPOT-VEGETATION sensors. Since 1995, a set of 10-day period satellite images obtained by these sensors has been collected with a 1 x 1 km²  spatial resolution.

The image-based crop monitoring and yield forecasting procedures are always a combination of elements described below.

1. Multitemporal image set: The starting point is a series of geographically congruent and periodic (mostly 10-daily) images over the area of interest. The NDVI-values (Normalised Difference Vegetation Index) vary from 0.15 for bare soils to ±0.80 for full green vegetations, with all gradations in-between. In the multitemporal image set, each pixel is thus characterised by a specific NDVI-time profile. However, since the raw profiles are still disturbed by cloudy measurements, the composites images are first submitted to a cleaning procedure (figure 3).

remote_sensing_1Figure 3: NDVI time series of a random pixel in the 10-daily composites of SPOT-VGT. The local minima in the original profile are mainly due to clouds. The envelope curve is the result of the cleaning procedure.

2. Spectral Unmixing: The pixel-profiles originate from 1km²-surfaces, which mostly cover different parcels and crops. The measured 1km-signal (reflectance or derived NDVI) is thus a mixture of the pure signals of the individual crops. In principle these pure signals can be reconstructed, if the sub-pixel acreage distributions of the different crops are known – as it is the case in Belgium thanks to the IACS (see figure 4). The analysis can then be performed separately for each crop. However, this spectral unmixing is not yet operational as the outcomes are often unpredictable. So far, we thus continue to work with the 1km²-profiles, and accept their mixed nature.

remote_sensing_2

 Figure 4: Crop-specific NDVI-profiles derived from the 1km-imagery of SPOT-VEGETATION by means of spectral unmixing.

3. Dry Matter Production (DM): The (optical) NDVI-values are also combined with meteorological information in order to assess the daily Dry Matter production per pixel (kgDM/ha/day). For this, we use a Monteith approach, which only requires the input of solar radiation and temperature (per pixel x time step). Technically, in the resulting DM-images all sensor differences are removed (AVHRR vs. VGT)

4. Cumulative values: There is in general no direct correlation between the (satellite-registered) optical response or DM-productivity of a crop at a given moment and the yield of the storage organs at harvest time. The final yield rather results from the history of the crop. This history can be quantified in some way by means of the cumulative NDVI or cumulative DM-production, starting from the emergence date of the crop.

5. Differencing: A widely used technique consists in the computation of new images which represent the absolute or relative difference between the actual state of a parameter (NDVI or DM, cumulative or not) and its state at the same date in the previous year (or in a standard reference year). In spite of the simplicity, these difference images are a powerful tool for the detection of zones where crop growth is reduced due to extreme weather conditions.

6. Regional means: Regional mean values of each of the previously defined parameters can be computed by superposing the concerned image(s) with the map of the agro-statistical regions. If the land use is known, the averaging can even be based on the subset of pixels which are mainly covered by the concerned crops. For the MARS-project, mean NDVI’s are computed for each EU-NUTS2 region, with selection of pixels from the CORINE land cover map. For Belgium, the selection is based on the IACS and we compute regional means of the DM-production (figure 5).

 remote_sensing_3
Figure 5 : Regional means of the Dry Matter production for 2 Belgian regions and years, derived from NOAA-AVHRR (mixed 1km²-pixels).

7. Calibration and Integration: Regional means are in the best case only indicative for the final yield. In view of quantitative predictions, crop growth indicators resulting from the model must be calibrated against official yield statistics. For Belgium, crop-specific relationships between the official yields and the indicators are established using neural networks. Independent Jackknife validation pointed out significant R²-values of about 60%, 50% and 35% for winter wheat, sugar-beets and fodder maize. The last improvement concerns the integration of the image-based estimators with the ones provided by the technological trend function and the agrometeorological model. By this way the final yield of winter wheat can already be predicted accurately (R²=70%) at the end of May.

In spite of their low resolution (pixels of 1km), periodic composites images of NOAA-AVHRR and SPOT-VEGETATION are undoubtedly useful for qualitative analyses, in particular for the monitoring of general growth patterns and the detection of short term changes (growth anomalies, floods, disasters, etc.).