Authors: Zhihao Wang*, The Ohio State University, Desheng Liu, The Ohio State University
Topics: Remote Sensing, Land Use and Land Cover Change, Environmental Science
Keywords: spatial-temporal modeling, Markov Random Field, MODIS time series data, Land cover classification
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Buchanan, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Monitoring the dynamics of land cover is immensely essential to the change science in both regional and global scales. However, the accuracy of current multi-temporal land cover products is limited by lacking spatial-temporal consistency, which yields unrobust classification in year-to-year label changes that rarely happen in the real world. A lot of efforts have been made to solve this illogical transition by incorporating the spatially and temporally neighboring information into the classifications. Few studies, however, focus on the contexts of all the previous and subsequent information; in other words, the optimal labelling sequence is rarely considered in most spatial-temporal models. Thus, we propose a novel framework that fully exploits the temporal sequence information coupled with the spatial context. Specifically, we employ a spatial-temporal Markov Random Filed model as a postprocessing step to the supervised classification, and we optimize the time sequence labelling using the idea from the existing hidden Markov model. The MODIS images are used to evaluate the effectiveness of our method, and the results show that the classification accuracy is greatly improved and the spatial and temporal information is consistent with ecological rules.