Satellite imagery, often freely available can be used for research purposes such as the monitoring of the evolution of forest cover. The European program Copernicus gives public access to the images captured by the Sentinel satellites. Radar imagery produced by Sentinel-1 and optical imagery produced by Sentinel-2 can be fused to produced land cover maps with the classes of interest (forest/not-forest). For deforestation detection, the changes occurring within forests can be detected automatically, in particular with deep learning for the segmentation of satellite images. This process also allows the analysis of the evolution of the phenomenon over time when it is applied to time series. Deep neural networks have been very successful in computer vision tasks, outperforming other methods from the state of the art. Convolutional neural networks (CNN) play a prominent role in the architectures used in various tasks such as classification and detection. The U-net architecture has shown good performances for segmentation tasks on biomedical images. This architecture can also be adapted for the classification of satellite images. We aim to propose a model for detecting deforestation using deep learning, with a U-net type architecture on fused optical and radar images. The challenges involved are diverse such as coping with the volume, the variety, the multi-dimensionality and the noise of the data.