Charlotte Pelletier (IRISA, Université Bretagne Sud)
Time-series data, which are ordered sequences of numerical or symbolic values, are nowadays ubiquitous. They are growing in quantity and velocity as the number of sensors (weather stations, surgical robots, body sensors, and many more) increases.
Their analysis is fundamental in a variety of applications including food security, environment, medicine, and human activity recognition. Among possible analysis tasks, time series classification (TSC) consists of associating a time series with a label.
As traditional classification approaches (e.g., random forests) fail to exploit the temporal structure of these data and their particularities (temporal relationships between consecutive observations, irregular sampling, high volume, etc.), specific methods have been proposed to automatically classify unlabelled time series in a reasonable amount of time.
In this talk I will present the different families of TSC approaches with a focus on recent advances, which are looking for a good tradeoff between accuracy and scalability. I will detail novel approaches based on decision trees and deep learning techniques. I will also briefly outline some applications and challenges of TSC to remote sensing data in the context of land cover mapping.