Crop Area Assessment Using Sub-pixel Classification with a Neural Network Trained for a Reference Year

Reliable and timely information on location and extent of crop areas is important for crop production forecasting, for decision support related to agri-environmental land use planning and management and for various scientific purposes. We investigated the possibility to estimate cropped areas from short and long time series of NDVI-imagery of low spatial resolution (SPOT-VEGETATION), by means of sub-pixel classification using neural networks (NN). We paid particular attention to verifying whether NN, trained for a particular agricultural season can be used to soft-classify time series of imagery of other seasons. To this end, separate networks were trained using reference data of 2003, 2004, 2005 and 2006 for March-to-May and March-to-October VEGETATION-NDVI-time series. Each of the resulting calibrated networks was then applied to the three other seasons. The region of Flanders in Belgium was selected as the test zone, because of the availability of excellent reference data in the form of a vectorial GIS with the boundaries and cover type of the large majority of agricultural fields. The result of the subpixel classification are a set of Area Fraction Images (AFIs). Each AFI contains for each 1 km-pixel the estimated area proportion occupied by one cover type (crops or other land use). The AFIs were validated at three different levels: the 1 km-pixel, the municipality and the agro-statistical district. In general, the neural networks trained for a certain year, provide more accurate overall results when applied on the image data of the same year, as compared to the time series of a year other then the training year. This is true at the pixel, municipal and regional level, although accuracy measures strongly improve with increasing aggregation. There is no major variability of the obtained accuracies over the various years. Long post-harvest time series outperform the short pre-harvest ones. Performance for individual land use classes and crops is variable too. Application of once-calibrated NN for other seasons cannot be recommended, even at higher spatial aggregation levels.