By: Ernesto Bastidas Obando
Abstract:
The need for regular information on crop area distribution in European landscapes makes of MERIS a good candidate for temporal monitoring of agricultural systems. Nevertheless, MERIS 300 m spatial resolution might still be too coarse for a hard classification approach because crop fields can be smaller on size. Therefore, the potential for retrieving sub-pixel information from temporal MERIS datasets over crop areas using feedforward artificial neural networks (FFNNs) was evaluated in this study. Two kinds of network structures, the first one using single pixel information and the second one using neighbor information from a 3 by 3 pixel window, were evaluated to estimate sub-pixel fractions based on five MERIS level 2 full resolution images. The five MERIS level 2 datasets were distributed over the growing season and covered the province of Noord-Holland in The Netherlands. A thematically and spatially aggregated version of the Dutch land use database (LGN5), with the main economically agricultural land cover types for The Netherlands (grassland, potatoes, sugar beet and cereals), was used to reference the MERIS datasets. Also the LGN5 was utilized for selection of training, validation and testing samples. Sub-pixel estimations coming from the trained 3 by 3 window FFNN were more precise than the estimations coming from a trained single pixel FFNN. The precision for potato estimates with a 3 by 3 window FFNN for different spatial scales ranged with coefficient of correlation (R) values from 0.25 at 9 ha to 0.85 at ground areas over 81 ha. For grassland, area estimates ranged with R values from 0.75 at 9 ha to 0.95 at ground areas over 81 ha. Further studies should include a definition of the temporal profiles with more temporal datasets and adoption of look-up tables for training the neural networks.