The TanDEM-X Forest/Non-Forest Map, derived from the volume decorrelation factor using a supervised fuzzy clustering algorithm, represents the baseline approach for forest mapping with TanDEM-X data at large/global-scale. Deep learning methods have been demonstrated to be also suitable for mapping forests with TanDEM-X interferometric data, e.g. by utilizing a U-Net convolutional neural network (CNN) on full-resolution images. In this work, we investigate the capabilities of using a U-Net-like architecture with TanDEM-X interferometric data for forest and water mapping on a large scale. An ad-hoc training strategy has been developed to detect forest and water on TanDEM-X images acquired with different acquisition geometries over the Amazon rainforest. In this case, a significant performance improvement with respect to the clustering approach, with a mean f1-score increase of 0.13 on test images has been measured with respect to the baseline clustering technique. The trained U-Net over the Amazon rainforest has been used to extend the forest and water mapping to other tropical forests over Africa and Asia. The classification improvements applying CNN methods on TanDEM-X data allow for the generation of time-tagged mosaics over the tropical forests by utilizing the nominal TanDEM-X acquisitions between 2011 and 2017, skipping the weighted mosaicking of overlapping images used in the clustering approach for achieving a good final accuracy, as well as avoiding the use of external layers to filter out water surfaces. The explotation of such mosaics over extended areas is a key aspect for the detection and monitoring of deforested areas worldwide.