An Evidential Reasoning Contextual Approach to Tropical Forest Classification Using Very High Spatial Resolution Earth Observation Data

Earth observation has been widely recognised as having considerable potential to aid an improved understanding of tropical forest biodiversity. However, studies that have explored the use of the multispectral response recorded by Earth observation satellites to infer climatically and edaphically determined floristically distinct tropical forest types have on the whole reported disappointing results, finding that many distinct forest types that differ in species composition, structure and habitat are poorly distinguished. In order to address this issue, this paper reports the results of employing a contextual evidential reasoning approach to classify lowland Peruvian Amazonia primary forest. Ikonos-2 multispectral images and elevation data, collected during the Shuttle Radar Topography Mission (SRTM) interferometer mission, were employed to derive fine spatial scale tropical forest classes employed in field-based bio-diversity studies of the study area. A contextual evidential reasoning classification approach was employed with forest-type evidence generated using kernel smoothing, while weights for each source was obtained using a genetic algorithm. The standard Dempster-Shafter assignment rule was employed to assign each pixel in the image to one of the forest types. Overall, an accuracy of 81% was achieved compared to a standard reflectance only maximum likelihood classification of 65%. Several classes with very low standard classification accuracy experienced a dramatic increase for the contextual evidential reasoning method (<30% for reflectance only to >83% evidential reasoning contextual).