Fully automatic multi-temporal land cover classification using Sentinel-2 image data

• Published in: Procedia computer science 2019, Vol.159, p.650-657
• Main Authors: Baamonde, Sergio, Cabana, Martiño, Sillero, Neftalí, Penedo, Manuel G., Naveira, Horacio, Novo, Jorge
• Format: Article
• Language: English
• Subjects: Land cover classification; Machine learning; Remote sensing; Sentinel-2
• ISSN: 1877-0509; 1877-0509
• DOI: 10.1016/j.procs.2019.09.220

The analysis of remote sensing images represents a highly important issue to be performed in many relevant fields such as climate change studies or land cover mapping. Traditional proposals usually identify the land cover classes from general related groups such as different tree species or different crop varieties. Additionally, these proposals commonly use information from a precise time span or season, not accounting for the variability of the data over the entire year, specially in regions with several seasons. In this work, we propose a multi-temporal classification system to identify and represent diverse land cover classes over any period of the entire year by using Sentinel-2 satellite image data. To this end, 526 representative samples were labelled from 5 complex and variable different land cover types over the Special Area of Conservation (SAC) Betanzos-Mandeo in the northwest of the Iberian Peninsula. The method achieves a satisfactory mean accuracy value of 84.0% for the testing set using the best configuration with a radial Support Vector Machine classifier. This system will be used in the study of the population connectivity of two threatened herptiles, but it can be easily extended to other species of interest in the future.