A deep learning based approach for detecting panels in photovoltaic plants

Photovoltaic (PV) panels are a clean and widespread way to produce renewable energy from sunlight; at the same time, such plants require maintenance, since solar panels can be affected by many types of damaging factors and have a limited yet variable lifespan. With the impressive growth of such PV installations, it is in the public eye the need of a cheap and effective way to continuously monitor the state of the plants and a standard technique designed to promptly replace broken modules, in order to prevent drops in the energy production. Since the faults mainly appear as Hot Spots on the surface of the PV panels, aerial thermal imaging can be used to diagnose such problems and also locate them in huge plants. To this aim, dedicated automatic Computer Vision methods are able to automatically find hot spots from thermal images, where they appear as white stains. In these methods a fundamental step is the segmentation of the PV panels, which allows to automatically detect each module. In this paper, we address the problem of PV Panel Detection using a Convolutional Neural Network framework called YOLO. We demonstrate that it is able to effectively and efficiently segment panels from an image. The method is quantitatively evaluated and compared to existing PV panel detection approaches on the biggest publicly available benchmark dataset; the experimental results confirm its robustness.