Application of Data-Driven Decomposition to Landsat-TM Images for Crop Area Estimation

An accurate crop area estimation method based on satellite remote sensing imagery is needed to manage the agricultural subsidy system of the European Union. The area estima-tor can use either classiication, which allocates a pixel to a single class, or decomposition, which divides a pixel between several classes, to determine the ground cover type(s) a pixel is composed of. While in early days classiication was much used, recently the decomposition approach has gained more interest, however, only on a per pixel basis. In a previous study, we developed the data-driven decomposition method, which used spatial information to guide the decomposition process; on artiicial Landsat-TM images this method proved to be far more accurate than techniques based on classiication or pixel-based decomposition. To investigate whether data-driven decomposition also results in an improved area estimation when using real satellite images, the area of 17 agricultural lots was determined from a large scale to-pographical map. After co-registration with a corresponding Landsat-TM image, application of data-driven decomposition gave an estimation that was equally or more accurate than the estimates of a similar method based on classiication in 14 of the 17 cases. Furthermore, data-driven classiication also showed to be better suited for handling the small boundary structures that separated the agricultural elds. These results suggest that the accuracy of data-driven decomposition is higher than that of an area estimator based on classiication when dealing with agricultural elds.