Real-time Semantic Segmentation of Solar Photovoltaic Arrays for Autonomous UAV Flights

This paper considers the real-time semantic segmentation of solar PhotoVoltaic (PV) arrays using down-facing cameras onboard Unmanned Aerial Vehicles (UAVs) to support autonomous navigation for ultra-low altitude industrial inspection. Various semantic segmentation algorithms, including SegNet, UNet, Deeplabv3+, and their respective variants, are compared to identify the most suitable one that strikes a balance between segmentation and real-time performance. Subsequently, each video frame is processed by the selected segmentation network to create a solar array mask. This mask is then optimally fitted to diverse solar array shapes to determine angular and horizontal deviations between the image center and the solar array center-line. The algorithms are validated on a self-constructed dataset of 594 solar PV array UAV images, encompassing varied environmental conditions and solar farm layouts. The top-performing network, lightweight vgg 16bn-Unet, achieves 99.4% overall accuracy and 98.2% Intersection over Union (IoU), while the most balanced network, mobilev3-Deeplabv3+, attains 98.9% accuracy and 96.5% IoU, with a 27.59 fps (using an NVIDIA GeForce MX150 GPU) suitable for real-time UAV flights.