Skip to main content
Conference abstract, 2005

Probabilistic classifier with applications to quality assessment and surface water management

Ranneby Bo, Yu Jun


Accurate and quality assured detection of spatial and temporal variation in land use or landscape patterns is needed as direct or indirect input for the assessment of ecosystem health. For these purposes remote sensing methods have potential to be an efficient tool. Unfortunately, several of the existing remote sensing methods will not give satisfactory results. Although, it is possible to improve traditional remote sensing classification methods, for several applications it is necessary to introduce a new concept, pixel-wise probabilistic classifiers. Instead of classifying each pixel to a specific class, each pixel is given a probability distribution describing how likely the different classes are. The pixel-wise vectors of probabilities can be used to judge how reliable a traditional classification is and to derive measures of the uncertainty (entropy) for the individual pixels. It is extremely important that proper probability distributions allowing frequency interpretation are derived; otherwise misleading results are obtained. As the probabilistic classifier gives unbiased area estimates over arbitrary areas they are very useful in source apportionment models. In these models type concentrations for different classes are multiplied with the corresponding area estimates. Then it will be possible to estimate leakage within the drainage basin. In addition, different scenarios are modelled, e.g. "no forestry" and "no forestry, agriculture or point sources" in the drainage basin, in order to demonstrate effects of forestry in relation to other human impact on nitrogen concentrations in surface waters

Published in

Proceedings of the 90th Annual Meeting of Ecological Society of America (ESA), Montreal, Canada
Publisher: ESA


The 90th Annual Meeting of Ecological Society of America (ESA)