Deep learning for photovoltaic panels segmentation

Due to advanced sensor technology, satellites and unmanned aerial vehicles (UAV) are producing a huge amount of data allowing advancement in all different kinds of earth observation applications. Thanks to this source of information, and driven by climate change concerns, renewable energy assessment became an increasing necessity among researchers and companies. Solar power, going from household rooftops to utility-scale farms, is reshaping the energy markets around the globe. However, the automatic identification of photovoltaic (PV) panels and solar farms' status is still an open question that, if answered properly, will help gauge solar power development and fulfill energy demands. Recently deep learning (DL) methods proved to be suitable to deal with remotely sensed data, hence allowing many opportunities to push further research regarding solar energy assessment. The coordination between the availability of remotely sensed data and the computer vision capabilities of deep learning has enabled researchers to provide possible solutions to the global mapping of solar farms and residential photovoltaic panels. However, the scores obtained by previous studies are questionable when it comes to dealing with the scarcity of photovoltaic systems. In this paper, we closely highlight and investigate the potential of remote sensing-driven DL approaches to cope with solar energy assessment. Given that many works have been recently released addressing such a challenge, reviewing and discussing them, it is highly motivated to keep its sustainable progress in future contributions. Then, we present a quick study highlighting how semantic segmentation models can be biased and yield significantly higher scores when inference is not sufficient. We provide a simulation of a leading semantic segmentation architecture U-Net and achieve performance scores as high as 99.78%. Nevertheless, further improvements should be made to increase the model's capability to achieve real photovoltaic units.