Advanced sampling of remote sensing images for area estimation

Classification methods of Earth observation data obtained by remote sensing allow to produce land cover map of every places of the world, each day, a major asset to compute estimations of area of a particular land cover class (e.g. crop area). However, the commission and omission errors occurring during the classification process (partly due to mixed pixels) are not counterbalanced and therefore, bias the results. Because of this constraint, operational systems have to set up a sampling scheme of in-situ data and/or survey to assess area rather than making use of Earth observation solely. For example, the last version of LUCAS, the crop acreage European monitoring system, is a two-stage systematic design scheme of unclustered points which are identified in the field. Remote sensing data are then used to produce a crop mask and a stratification of the territory to improve the efficiency of the sampling. This research explores ways for improving the accuracy of the information provided by Earth observation data classification by applying sampling directly to the satellite image in order to get a subset classifiable more accurately. The selection criteria used to build the sampling has to be as independent as possible to the land cover type, so ideally derivable a priori i.e. before the classification process. Using a series of 3 SPOT5 images acquired during 2013 growing season in Belgium, this research seeks to develop an object-based method for the estimation of a priori misclassification probabilities of (i) the crop land and (ii) the crop type. The validation is done using the Integrated Administration and Control System (IACS), a vectorial and annually updated GIS which contains information on most of the agricultural parcels in the Walloon region (field limits and crop type). First results show that a set of spectral and spatial object features, completely and automatically extractable from one image, are useful to produce a misclassification probabilities map of the crop land. From this output, a sampling of the image object can be performed based on a trade-off between the sampling rate (to maximize) and misclassification rate (to minimize). The influence of the sampling on the classification bias is studied via the evolution of commission and omission errors with the sampling rate. Those results could be used as a significant input to operational agriculture monitoring systems. In the case of unbiased sampling, it could be directly used as an alternative estimator to field sampling for crop area. In any case, it could improve the efficient of field sampling, given information on area where crop type has been identified with a high confidence level and help to prioritize field visit through the identification of areas hard to classify with remote sensing. That information could also be used for crop yield monitoring because it enables to follow parcels with low misclassification probabilities.