This paper presents a principled approach for detecting shadows in satellite imagery at a large scale. We propose a two-stage process, starting with the generation of a training set based on an airborne LiDAR point cloud and a target satellite image. This point cloud is first rasterised onto the spatial grid of the image (0.5m). Shadows are then geometrically projected using a standard hillshade algorithm. Finally we improve the quality of the training set by handling small objects and removing "bright" shadows that appear due to temporal misalignment between image and point cloud. The second stage leverages Deep Learning (DL) to automatically detect shadows in areas where airborne LiDAR is not available, or out of date. Our experiments on manually labeled shadows demonstrate that introducing variance to solar azimuth helps to reduce confusions with water and dark surfaces (asphalt). We further show that the learned model is able to replicate the quality of the LiDAR-based shadow maps on small and medium-sized shadows, in a geographical area unseen during training. We believe increasing the size of the training data set to more diverse cities would improve generalization on large shadows and on vegetation.