A Swedish research group has found that using deep machine learning to identify solar energy systems in aerial images may not be so accurate in non-densely populated countries such as Sweden. They have also found, however, that this technique may be trained via an iterative process and achieve satisfying results.
According to the research group, the proposed framework achieved an accuracy of 63.9% when used over a Swedish data set. That is lower than previous research conducted with the same framework in other countries. For example, a US group of researchers attained an accuracy of 91%, and a study done in Germany achieved 87.3%. However, the Swedish-trained CNN has achieved more competitive results regarding the recall rate. While the precision rate refers to the method’s ability not to make mistakes, the recall rate refers to its ability not to let positive information slip through. In that recall metric, the Swedish DeepSolar has achieved 81.8%, compared to 98.1% in the US and 87.5% in Germany.
The scientists said the algorithm was first trained with a data set from North-Rhine Westphalia state in Germany and was then fine-tuned to Sweden with pictures from eight municipalities. It was then used to scan the whole spatial area of three Swedish municipalities – Uppvidinge, Falun, and Knivsta. These data were compared with other data collected with onsite inspections.
This iterative process involved multiple scans, with the CNN algorithm being retrained after each municipality scan, resulting in progressively enhanced accuracy. In the initial scan, the algorithm detected 89% of the detectable PV systems (excluding BIPV and vertical installations) and 59% of the ST systems,” the scientists emphasized. “Remarkably, by the fourth and final scan, these detection rates improved to 95% for PV systems and 80% for ST systems.”
They also specified that most undetected PV systems were frameless modules, typically installed on darker-colored roofs. In addition, shading from trees or structures, image reflections, and systems with high tilt angles impeded the classification algorithm’s detection efficacy.
