Detecting and Segmenting Solar Farms in Satellite Imagery: A Study of Deep Neural Network Architectures

Abstract

In line with global sustainability goals, such as the Paris Agreement, accurate mapping, monitoring, and management of solar farms are critical for achieving net zero emissions by 2050. However, many solar installations remain undocumented, posing a challenge. This paper studies semantic segmentation using deep neural networks, including networks constructed using network architecture search (NAS), for solar farm detection. Semantic segmentation has evolved through technologies like Fully Convolutional Networks and U-Net, which have shown strong performance on satellite imagery. For NAS, Differentiable Architecture Search and its variants like Auto-DeepLab have become efficient ways to automate the creation of neural network architectures. This work compares models generated using Auto-DeepLab to Solis-seg, a Deep Neural Network optimized for detecting solar farms in satellite imagery. Solis-seg achieves a mean Intersection over Union (IoU) of 96.26% on a European Sentinel-2 dataset, with Auto-DeepLab models lagging slightly behind. Our results for Solis-seg also challenge the prevailing method of using transfer learning from classification tasks for semantic segmentation. Thus, this work contributes to both the field of earth observation machine learning and the global transition to renewable energy by studying an efficient, scalable approach to tracking solar installations. We believe that this paper offers valuable insights into applying advanced machine learning techniques to solar farm detection and can be useful for further research in earth observation and sustainability.