In this work, we face the problem of forest mapping from TanDEM-X data by means of Convolutional Neural Networks (CNNs). Our study aims to highlight the relevance of domain-related features for the extraction of the information of interest thanks to their joint nonlinear processing through CNN. In particular, we focus on the main InSAR features as the backscatter, coherence, and volume decorrelation, as well as the acquisition geometry through the local incidence angle. By using different state-of-the-art CNN architectures, our experiments consistently demonstrate the great potential of deep learning in data fusion for information extraction in the context of synthetic aperture radar signal processing and specifically for the task of forest mapping from TanDEM-X images. We compare three state-of-the-art CNN architectures, such as ResNet, DenseNet, and U-Net, obtaining a large performance gain over the baseline approach for all of them, with the U-Net solution being the most effective one.2 of 18 between the master and slave images, enabling high-resolution interferometric measurements with an unprecedented quality. The constellation comprises two twin satellites flying in a bistatic close-orbit configuration, which allows for a flexible selection of the acquisition geometries and, in particular, of the interferometric baselines [8]. The main goal of the mission was the generation of a global consistent high-resolution digital elevation model (DEM) with unprecedented accuracy, which has been successfully completed in 2016 [9]. Besides the nominal DEM product, for each bistatic interferometric TanDEM-X acquisition, additional quantities can be computed as by-pass products. Indeed, the bistatic acquisition is not affected by temporal decorrelation, allowing for an accurate isolation of volume scattering phenomena from the interferometric coherence. This feature was exploited in References [10,11], where the authors presented a framework for the development of a global TanDEM-X forest/non-forest map [12] as described more in details in Section 2.1.Deep learning approaches and, in particular, Convolutional Neural Networks (CNNs) have been massively used in computer vision and image processing in the last few years, since the publication of the breakthrough work of Krizhevsky et al. on image classification in 2012 [13]. Thanks to the CNNs capability to learn very complex nonlinear relationships from huge labeled datasets with the help of commercial GPUs, unprecedented results have been obtained for many typical tasks such as super-resolution [14,15], segmentation [16], denoising [17], object detection [18,19], classification [20][21][22], and many others.Recently, deep learning has started to significantly impact remote sensing applications as well, as testified by the recent survey of Zhu et al. [23]. Established techniques in remote sensing concern, e.g., pansharpening [24,25], vehicle detection [26] with optical images, crop classification [27,28], anomaly detection with hyperspectral data [29], despeckling [30,31],...