Toward global rooftop PV detection with Deep Active Learning

It is crucial to know the location of rooftop PV systems to monitor the regional progress toward sustainable societies and to ensure the integration of decentralized energy resources into the electricity grid. However, locations of PV are often unknown, which is why a large number of studies have proposed variants of Deep Learning to detect PV panels in remote sensing data using supervised Deep Learning. However, these methods are based on annotating datasets and therefore often require relabeling when fine-tuned or extended to a different region. Recent advances in Deep Active Learning offer the opportunity to significantly reduce the number of required annotated images by intelligently selecting the images to label next based on their informative value for the model. In this study, we compare different Deep Active Learning algorithms using a variety of datasets from different regions and compare different model training variants. In the simulations, the entropy-based acquisition function shows the highest performance with only 3% of the data needed in case-imbalanced data, while remaining simple to implement. We believe that Deep Active Learning provides an elegant solution to maintain high model accuracy while reducing annotation effort substantially. This facilitates the development of generalizable models for worldwide rooftop PV detection. • Labeling is hindering generalizable, global rooftop PV detection. • Deep Active Learning greatly facilitates model training for PV detection. • In highly imbalanced data, only 3% of the data are needed. • Entropy-based sampling acquisition function showed best performance. • Deep Active Learning is reliable for model extension and model fine-tuning.