Automatic Analysis of Satellite Image Time Series
Abstract
This project tackles the analysis of series of satellite images by scaling up research into time series classification to large quantities of data. Latest Earth observation satellites (Sentinel-2, Landsat-8, Vens) are now imaging Earth every 2-5 days at high resolution. This opens up incredible opportunities to closely monitor the land-use and land-cover of the planet over time. Current capability can not make the most of these opportunities, because state-of-the-art research into time series classification does not scale to this amount of data. This project explored how to scale up one of the state-of-the-art algorithms for time series classification NN-DTW (the Nearest Neighbor algorithm coupled with the Dynamic Time Warping similarity measure) - to large quantities of data. In the first year, the research team focused on the classification time and introduced a novel method that allows classification of time series up to 500x faster; while it would have taken 1 CPU-year to create a temporal map of the Houston area at 10m resolution with previous state-of-the-art approaches, the research team showed an ability to do it in less than 4 hours on a single computer. In the second year, they focused on the training time and introduced a novel method to efficiently learn a classification model from data. They showed that our approach allows to gain up to 3 orders of magnitude in training time. This work was described in two publications in the 2017 and 2018 editions of the SIAM International Conference on Data Mining. The work for the second year was awarded Best Paper.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 15, 2018
- Accession Number
- AD1057269
Entities
People
- Francois Petitjean
Organizations
- Monash University