Automatic Boundary Extraction of Large-Scale Photovoltaic Plants Using a Fully Convolutional Network on Aerial Imagery

Abstract

This article presents a novel method for boundary extraction of photovoltaic (PV) plants using a fully convolutional network (FCN). Extracting the boundaries of PV plants is essential in the process of aerial inspection and autonomous monitoring by aerial robots. This method provides a clear delineation of the utility-scale PV plants' boundaries for PV developers, operation and maintenance service providers for use in aerial photogrammetry, flight mapping, and path planning during the autonomous monitoring of PV plants. For this purpose, as a prerequisite, the 'Amir' dataset consisting of aerial imagery of PV plants from different countries, has been collected. A Mask-RCNN architecture is employed as a deep network with VGG16 as a backbone to detect the boundaries precisely. As comparison, the results of another framework based on classical image processing are compared with the FCN performance in PV plants boundary detection. The results of the FCN demonstrate that the trained model is able to detect the boundaries of PV plants with an accuracy of 96.99% and site-specific tuning of boundary parameters is no longer required.