Automated 3D Vegetation Detection Along Power Lines using Monocular Satellite Imagery and Deep Learning

Vegetation is one of the primary causes of outages in electricity transmission and distribution networks and represents a significant expense in maintaining a power grid. While LiDAR or multi-view images can be used for detecting vegetation along power lines, such technologies are costly and difficult to acquire to cover widespread electricity networks. This paper proposes a framework for 3D mapping of trees along power lines using monocular high-resolution satellite images. Such type of imagery has become nowadays affordable and easy to acquire. Furthermore, single snapshots can cover a large portion of the grid in high revisiting time. We train and test different state-of-the-art models to map the contextual information from images into a height prediction. We validate our proposed satellite-based framework for an electricity distribution network in the western part of Norway using actual LiDAR data.