New technique to geo-label PV modules in utility-scale solar parks

Scientists in Morocco have developed a method that uses the metadata of PV plants’ infrared images to label them geographically. The automatic database can then be used in deep learning models and significantly reduce the time required for data labeling.

A group of researchers from Morocco has developed a novel technique to geo-label solar modules in large-scale parks.

It utilizes infrared (IR) images from unmanned aerial vehicles (UAVs) as inputs, employing adaptive thresholding, edge refinement, and photogrammetric data to segment and localize solar modules without requiring manual annotation. Furthermore, the automatically labeled dataset can be used to train deep learning models.

“This contribution accelerates the inspection process for large-scale installations by significantly reducing the time required for solar panel annotation, enabling the training of deep learning detectors with minimal human intervention,” the team said. “The proposed workflow also ensures real-time applicability by achieving an optimal trade-off between detection accuracy and inference time.”

The novel labeling method uses UAV metadata such as GPS,  inertial measurement unit (IMU) and camera parameters to calculate ground sampling distance (GSD). Then, for the automatic labeling, it uses the Niblack technique to generate threshold values from the images and find solar panels. Following that, it uses edge refinement and clustering to verify its findings, comparing the dimensions with the actual dimensions of the solar panel. The extraction process includes a script to convert the coordinates into the format required by deep learning models.

This method was tested on two case studies in Morocco. The first was the Green Energy Park platform, which included 22 kW of ground-mounted monocrystalline panels with a tilt of 31°. The second case study was of a plant located on the roof of a Moroccan data center, with 1 MW of monocrystalline panels. Thermal image acquisition was performed using DJI’s Mavic 2 Enterprise Advanced (M2EA), equipped with both thermal and visual cameras, featuring resolutions of 640 x 512 and 8,000 x 6,000, respectively.

“It achieved 91 % recall in the automatic geo-labeling step and significantly reduced the false positives through clustering and geometrical constraints,” the results of the automated dataset showed. “Ultimately, this paper improves the mentation and extraction of single PV modules and accelerates O&M operations by including the geolocation of the extracted modules.”

The labeled images generated by the automatic process were then split into training, validation, and test sets and tested with several deep learning models. Namely, the team tested the SSD ResNet50 V1, SSD MobileNet v2, Faster RCNN ResNet 50 V1, Faster EfficientDet d1, CenterNet hg104, and YOLOv7 models. In all cases, a batch size of 8 was used, while the training was conducted for 500 epochs.

“The automatically geo-labeled data was used to train multiple deep learning detectors. Among them, Yolov7 achieved the best performance, with a mean average precision at 0.5 intersection over union (mAP@0.5) of 98.33% with an inference time of 15 ms per image, proving its adaptability for real-time inspection scenarios,” the team concluded. “Moreover, the geolocation method achieved 2.51 m error, which supports accurate field-level identification of modules and enhances on-site maintenance operations.”

The new methodology was presented in “Fast and automatic solar module geo-labeling for optimized large-scale photovoltaic systems inspection from UAV thermal imagery using deep learning segmentation,” published in Cleaner Engineering and Technology. The research team included Scientists from Morocco’s Agronomic and Veterinary Institute Hassan II (IAV Hassan II), the Green Energy Park Research Platform, and the National Schools of Applied Sciences Oujda.