A Proposal for Vegetation Detection on Distribution Networks using Deep Learning for Satellite Image Recognition

Electricity interruptions caused by vegetation close to the distribution network are common in several energy distributors that use overhead networks. The growth of vegetation in contact with the grid can result in temporary or permanent faults, interrupting the supply of electrical energy, impacting indicators and system reliability. This work aims to develop a system for detecting trees that are close to the distribution network, so that preventive maintenance plans can be generated for tree pruning, avoiding contact with the network. To achieve this, computer vision and deep learning techniques are used to identify trees in satellite images taken from Google Earth. A YOLO model is trained with 105 images, 80 for training and 25 for validation, and cross-referenced with data referring to the georeferenced distribution network. A case study carried out in an urban setting is presented. The trees and the distribution network are identified in an image area. After this step, it is possible to extract the regions of the identified elements and find where the scenario is close to the network.