Spatiotemporal Mining of Time-Series Remote Sensing Images Based on Sequential Pattern Mining
- 1Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100094, China
- 2University of Chinese Academy of Sciences, Beijing 100049, China
Keywords: Remote Sensing, Satellite Image, Time Series, Data Mining, Sequential Pattern Mining, Land Cover Change
Abstract. With the continuous development of satellite techniques, it is now possible to acquire a regular series of images concerning a given geographical zone with both high accuracy and low cost. Research on how best to effectively process huge volumes of observational data obtained on different dates for a specific geographical zone, and to exploit the valuable information regarding land cover contained in these images has received increasing interest from the remote sensing community. In contrast to traditional land cover change measures using pair-wise comparisons that emphasize the compositional or configurational changes between dates, this research focuses on the analysis of the temporal sequence of land cover dynamics, which refers to the succession of land cover types for a given area over more than two observational periods. Using a time series of classified Landsat images, ranging from 2006 to 2011, a sequential pattern mining method was extended to this spatiotemporal context to extract sets of connected pixels sharing similar temporal evolutions. The resultant sequential patterns could be selected (or not) based on the range of support values. These selected patterns were used to explore the spatial compositions and temporal evolutions of land cover change within the study region. Experimental results showed that continuous patterns that represent consistent land cover over time appeared as quite homogeneous zones, which agreed with our domain knowledge. Discontinuous patterns that represent land cover change trajectories were dominated by the transition from vegetation to bare land, especially during 2009–2010. This approach quantified land cover changes in terms of the percentage area affected and mapped the spatial distribution of these changes. Sequential pattern mining has been used for string mining or itemset mining in transactions analysis. The expected novel significance of this study is the generalization of the application of the sequential pattern mining method for capturing the spatial variability of landscape patterns, and their trajectories of change, to reveal information regarding process regularities with satellite imagery.