Deep learning for cloud detection

Matthieu Le Goff (1, 2) , Jean-Yves Tourneret (3, 4) , Herwig Wendt (5, 6) , Mathias Ortner (1) , Marc Spigai (1)

1 IRT Saint Exupéry - Institut de Recherche Technologique
2 IRIT-TCI - Traitement et Compréhension d’Images
3 IRIT-MINDS - CoMputational imagINg anD viSion
4 Toulouse INP - Institut National Polytechnique (Toulouse)
5 IRIT-SC - Signal et Communications
6 CNRS - Centre National de la Recherche Scientifique

The SPOT 6-7 satellite ground segment includes a systematic and automatic cloud detection step in order to feed a catalogue with a binary cloud mask and an appropriate condence measure. However, current approaches for cloud detection, that are mostly based on machine learning and hand crafted features, have shown lack of robustness. In other tasks such as image recognition, deep learning methods have shown outstanding results outperforming many state-of-the-art methods. These methods are known to produce a powerful representation that can capture texture, shape and contextual information. This paper studies the potential of deep learning methods for cloud detection in order to achieve state-of-the-art performance. A comparison between deep learning methods used with classical handcrafted features and classical convolutional neural networks is performed for cloud detection. Experiments are conducted on a SPOT 6 image database with various landscapes and cloud coverage and show promising results.

Mots clés: Deep Learning; Feature engineering; Patch-based classication; Superpixels

Domaines: Technologies Émergeantes [cs.ET]; Multimédia [cs.MM]; Autre

Matthieu Le Goff, Jean-Yves Tourneret, Herwig Wendt, Mathias Ortner, Marc Spigai. Deep learning for cloud detection. ICPRS (8th International Conference of Pattern Recognition Systems), Jul 2017, Madrid, Spain. pp. 1-6, ⟨10.1049/cp.2017.0139⟩. ⟨hal-01783857⟩