Spatial Estimation of Classification Accuracy Using Indicator Kriging with an Image-Derived Ambiguity Index

Traditional classification accuracy assessments based on summary statistics from a confusion matrix furnish a global (location invariant) view of classification accuracy. To estimate the spatial distribution of classification accuracy, a geostatistical integration approach is presented in this paper. Indicator kriging with local means is combined with logistic regression to integrate an image-derived ambiguity index with classification accuracy values at reference data locations. As for the ambiguity measure, a novel discrimination capability index (DCI) is defined from per class posteriori probabilities and then calibrated via logistic regression to derive soft probabilities. Integration of indicator-coded reference data with soft probabilities is finally carried out for mapping classification accuracy. It is demonstrated via a case study involving classification of multi-temporal and multi-sensor SAR datasets, that the proposed approach can provide a map of locally-varying accuracy values, while respecting the overall accuracy derived from the confusion matrix. It can also highlight areas where the benefit of data fusion was significant. It is expected that the indicator approach presented in this paper could be a useful methodology for assessing the spatial quality of classification results in a probabilistic way.