Predicting crop reflectances using satellite data observing mixed pixels
Résumé
The use of satellite remote sensing for crop management is conditioned by the repeatability of the measurements and by the spatial resolution of the sensor: the better resolution, the less frequent is the observation. To have a regular follow-up, we need to use a satellite with median resolution (around 1 km2 corresponding to a weekly temporal resolution). Pixels with such a resolution correspond to different spatial components (cultures) and are named mixed pixels. We propose a statistical modeling of such satellite data that will enable us to predict information relative to crops observed through mixed pixels. At a given time and restricted to a homogeneous agro-climatic region, this model assumes that reflectances of the same crop (e.g. wheat, barley, and forests) are distributed as Gaussian with parameters depending on the crop. Conditional on the percentage of land occupation, we write a linear model with random parameters. We use the best linear unbiased prediction to predict the individual variations of reflectances. We apply this model to a SPOT image (pixel size 20 x 20 m) with a degraded resolution such that new pixels are sized 400 x 400 m. We can validate the results given by our model with the original SPOT data. With such hypotheses, the method gives good results