Deep constrained clustering applied to satellite image time series

Baptiste Lafabregue (1, 2) , Jonathan Weber (2) , Pierre Gançarski (1) , Germain Forestier (2)

1 ICube - Laboratoire des sciences de l'ingénieur, de l'informatique et de l'imagerie
2 IRIMAS - Institut de Recherche en Informatique Mathématiques Automatique Signal

The advent of satellite imagery is generating an unprecedented amount of remote sensing images. Current satellites now achieve frequent revisits and high mission availability and provide series of images of the Earth captured at different dates that can be seen as time series. Analyzing satellite image time series allows to perform continuous wide range Earth observation with applications in agricultural mapping , environmental disaster monitoring, etc. However, the lack of large quantity of labeled data generally prevents from easily applying supervised methods. On the contrary, unsupervised methods do not require expert knowledge but sometimes provide poor results. In this context, constrained clustering, which is a class of semi-supervised learning algorithms , is an alternative and offers a good trade-off of supervision. In this paper, we explore the use of constraints with deep clustering approaches to process satellite image time series. Our experimental study relies on deep embedded clustering and the deep constrained framework using pairwise constraints (must-link and cannot-link). Experiments on a real dataset composed of 11 satellite images show promising results and open many perspectives for applying deep constrained clustering to satellite image time series.

Mots clés: time series; satellite images; remote sensing; clustering; con- straints; deep embedded clustering

Domaines: Intelligence artificielle [cs.AI]

Baptiste Lafabregue, Jonathan Weber, Pierre Gançarski, Germain Forestier. Deep constrained clustering applied to satellite image time series. ECML/PKDD Workshop on Machine Learning for Earth Observation Data (MACLEAN), Sep 2019, Würzburg, Germany. ⟨hal-02326219⟩