Empirical models for estimating land cover areas from remotely sensed imagery

The mapping of land cover and land use is a key application of remotely sensed data. Traditionally, classification techniques are used to assign every pixel of an image to one of a number of mutually exclusive land cover classes. Alternatively, a modelling approach assigns to every pixel the area proportion containing each land cover class. This paper examines the hypothesis that the area modelling, or area estimation, approach can offer a richer and qualitatively more accurate representation of the true land cover than can be provided by the traditional classification approach. The paper describes the empirical, non-linear classifiers and area estimation models, based on neural networks and nearest neighbour algorithms, that have been developed to investigate this hypothesis. The algorithms were applied to an area-labelled Landsat TM data set produced as part of the EU FLIERS Project. The results demonstrated that a better representation of the true land cover was obtained using the area estimation models compared to the representation produced by the classification algorithms when the size of the land cover objects on the ground was less than the resolution of the sensor. These results are presented with a discussion of the evaluation issues involved with area estimation