Spatio-Temporal Data Mining: From Big Data to Patterns

Technological advances in terms of data acquisition enable to better monitor dynamic phenomena in various domains (areas, fields) including environment. The collected data is more and more complex spatial, temporal, heterogeneous and multi-scale. Exploiting this data requires new data analysis and knowledge discovery methods. In that context, approaches aimed at discovering spatio-temporal patterns are particularly relevant. This paper1 focuses on spatio-temporal data and associated data mining methods. 1 Spatio-temporal Data In recent years, technological advances in data acquisition (satellite images, sensors, etc.) have enabled numerous applications in surveillance and environmental monitoring: detection of abrupt changes (natural disasters, etc.), evolution tracking of natural phenomena (coastal erosion, desertification, wildfires, etc.) or development of models (hydrology, agriculture, etc.). The collected data is usually heterogeneous, multiscale, spatial and temporal (time series of satellite images, aerial or terrestrial photos, digital terrain models, physical ground measurements, qualitative observations, etc. ). This data is used to understand and predict phenomena generated by processes that are complex and of multidisciplinary origin (climatic, geological, etc.). Exploitation by experts of those huge volume of complex data (big data) requires not only to structure it to the best but also and mainly to design data analysis and knowledge discovery methods. In that context, approaches involving pattern mining are particularly relevant. The content of the paper was prepared in collaboration with H. Alatrista Salas, S. Bringay, F. Flouvat, and N. Selmaoui. With the dramatic growth of spatial information and Geographic Information Systems (GIS), many studies have been carried out in the context of spatiotemporal patterns mining. Early work in this area has dealt with spatial and temporal dimensions separately. Extraction of temporal sequences aims at identifying features frequent over time without taking into account spatial relationships. Colocation mining methods extract set of features which frequently appear in close objects without taking into account the temporal aspect. More recently, these works have been extended to simultaneously integrate spatial and temporal dimensions. Examples include the detection of sequences of located events and trajectory mining. A review has been published by the consortium GeoPKDD (Giannotti and Pedreschi, 2008). However, in those approaches, the mined patterns do not match the spatial complexity encountered when dealing with sattelite images. Similarly, the primitive constraints usually used (typically minimum frequency) are not sufficient to express criteria of interest for experts, such as geologists. A spatiotemporal database contains information characterized by a spatial and a temporal dimensions. Two types of spatiotemporal databases are mainly considered: databases containing trajectories of moving objects located in both space and time (e.g. bird or aircraft trajectories); databases storing spatial and temporal dynamics of events (e.g. erosion evolution in a region or epidemic spread in a city). 2 Mining moving object trajectories The emergence of new mobile technologies has facilitated the collection of large amounts of spatiotemporal data, dedicated to the localization of mobile objects in space and time (Perera et al., 2015). These new databases provide opportunities for new applications. The project GeoPKDD (Giannotti and Pedreschi, 2008), for example, studied