Assessing the Pattern of Error in a Predictive Soil Map
Coakley, Benjamin M.
MetadataShow full item record
SoLIM is a predictive soil mapping application employing fuzzy logic and expert knowledge to model the relationship between landscape and soil type. The end result of SoLIM's predictive approach is a hardened map of soil types and two confidence statistics describing the uncertainty associated with their prediction. This research valuated SoLIM's performance using a validation dataset consisting of 495 manually-classified soil samples randomly located within a test watershed. The soil types identified manually in each sample point also contained two assessments of measurement uncertainty. The patterns of SoLIM's mispredictions were investigated by comparing the manual and SoLIM soil type predictions: 1) In the spatial domain, using point pattern and geostatistical methods; 2) In the attribute domain; and 3) In relation to SoLIM's internal confidence measures. Because measurement uncertainty was assessed within the validation dataset, the analysis was repeated for different subsets of the validation dataset defined using different levels of measurement uncertainty. No spatial trends were found in the location of errors. Association was found between some input variables used in SoLIM's predictive model and prediction error for some subsets of the validation dataset, but no overall trends were visible. Association was found between SoLIM's internal uncertainty measures and error.