Solar Panel Mapping via Oriented Object Detection

TL;DR

This paper tackles scalable, fine-grained mapping of individual solar panels from aerial imagery by formulating panel detection as an arbitrarily oriented object-detection problem. It introduces a Faster R-CNN-based RRPN/RRoI pipeline with rotated bounding boxes $(x, y, w, h, \\theta)$ and rotated anchors spanning $[-90^\circ, -60^\circ, -30^\circ, 0^\circ, 30^\circ, 60^\circ, 90^\circ]$, enabling end-to-end prediction of panel vertices. The dataset comprises 121 high-resolution orthomosaics at about $2.5$ cm GSD, with patch-based processing and a 80/10/10 train/val/test split; evaluation uses $mAP$/$mAR$ and high-IoU metrics like $ ext{AP}_{75}$. Results show the ResNet-50-FPN backbone achieves the best validation performance ($ ext{AP}=83.3\%$, $ ext{AP}_{75}=94.1\%$), supporting accurate, georeferenced mapping of panels across diverse solar farms. This approach has practical impact by enabling automated, scalable monitoring and potentially recovering significant solar capacity by accelerating fault detection.

Abstract

Maintaining the integrity of solar power plants is a vital component in dealing with the current climate crisis. This process begins with analysts creating a detailed map of a plant with the coordinates of every solar panel, making it possible to quickly locate and mitigate potential faulty solar panels. However, this task is extremely tedious and is not scalable for the ever increasing capacity of solar power across the globe. Therefore, we propose an end-to-end deep learning framework for detecting individual solar panels using a rotated object detection architecture. We evaluate our approach on a diverse dataset of solar power plants collected from across the United States and report a mAP score of 83.3%.

Figures (3)

• Figure 1: Oriented Panel Detection Architecture.
• Figure 2: Sample Test Set Ortho Panel Predictions.
• Figure 3: USA Solar farm locations.