Sub-Pixel Methods in Remote Sensing

Abstract

Th eoretically, remote sensing is well-suited as a source of information on environmental features such as land cover. Remotely sensed imagery are available at a range of spatial and temporal scales, have a map-like format and, due to the nature of radiation interactions with the Earth’s surface, depict variation in land surface properties that is linked to land cover. Indeed, as most general land cover classes (e.g. forest, grassland etc.) diff er in their interaction with the radiations commonly used in remote sensing, it should be possible to accurately map and monitor land cover from remotely sensed data. Although the literature contains an enormous wealth of articles that demonstrate that it is possible to accurately map land cover from remotely sensed data it is often, in practice, an extremely diffi cult task (Townshend, ). Th ose studies that have succeeded in accurately mapping land cover have also commonly been based upon small test sites, often with ideal conditions (e.g. large relatively homogenous land cover parcels, insignifi cant topographic variability etc.) and used carefully selected and processed imagery. Th ese successes do not mean it is generally easy to extract land cover information accurately from remotely sensed data. Th ere are many instances in which an approach to land cover mapping developed successfully at one site fails to map land cover accurately when applied to another (Wilkinson, ; van Collie et al., ). Consequently, it is unsurprising that land cover maps often contain considerable error. For example, the IGBP global land cover map derived at the end of the last century provides a representation of the Earth’s land cover that is estimated to have an area-weighted accuracy of ~, substantially below its target accuracy of  (Scepan, ). Th us despite a long history of research into the topic and a widely held view that land cover mapping is a basic and simple task it is apparent that the potential of remote sensing as a source of land cover data is unfulfi lled (Townshend, ; Estes & Mooneyhan, ).