A Markov Chain Geostatistical Framework for Land-Cover Classification With Uncertainty Assessment Based on Expert-Interpreted Pixels From Remotely Sensed Imagery

This paper introduces an expert interpretation-based Markov chain geostatistical (MCG) framework for classifying land-use/land-cover (LULC) classes from remotely sensed imagery. The framework uses the MCG method to classify uninformed pixels based on the informed pixels and quantify the associated uncertainty. The method consists of the following steps: 1) decide the number of LULC classes and define the physical meaning of each class; 2) obtain a data set of class labels from one or a time series of remotely sensed images through expert interpretation; 3) estimate transiogram models from the data set; and 4) use the Markov chain sequential simulation algorithm to conduct simulations that are conditional to the data set. The simulated results not only provide classified LULC maps but also quantify the uncertainty associated with the classification. A case study with three LULC classes shows that, with increasing number of informed pixels from 0.45% to 1.81% of the total pixels at the resolution of 4.8 m × 4.8 m, the optimal classification accuracy based on maximum probabilities increases from 88.13% to 99.23% and the averaged classification accuracy of realization maps increases from 81.84% to 97.18%. Although it is relatively labor intensive, such an expert interpretation and geostatistical simulation-based approach may provide a useful LULC classification method complementary to existing image processing methods, which usually account for limited expert knowledge and may not incorporate ground observation data or assess the uncertainty associated with classified data.