A spatial–temporal contextual Markovian kernel method for multi-temporal land cover mapping

Due to a lack of spatial–temporal consistency, the current generation of multi-temporal land cover products is subject to significant error propagation in change detection results. To address the evolving needs of land change science, the next generation of land cover products must be derived from new classification methods that are designed specifically for multi-temporal land cover mapping. In this paper, a next generation classifier is proposed that fully exploits contextual information by combining results born from the machine learning paradigm in remote sensing with domain knowledge from multi-temporal land cover mapping. This classifier, the Spatial–Temporal Markovian Support Vector Classifier, exhibits an entirely new level of accuracy of change detection when evaluated for the classification of seven Landsat images from an Appalachian Ohio study area. It exceeds previous leading techniques employing machine learning kernel methods and Markov Random Field models of image context on all accuracy metrics for the creation of a spatial–temporally consistent land cover product. It owes its performance to the greatly improved decision-making about contextual information afforded by the extension and integration of these previous techniques. With such a classifier, substantially more accurate and spatial–temporally consistent multi-temporal land cover products are possible that are suitable for the detailed study of land cover change.