Scientists in Morocco have developed a method that uses the metadata of the infrared images of PV plants to label them geographically. The automatic database can then be used in in -depth learning models and the time required for data labeling is considerably shortened.
A group of researchers from Morocco has developed a new technique for geo-label solar modules in large-scale parks.
It uses infrared (IR) images of unmanned aircraft (UAVs) as inputs, using adaptive threshold, peripheral and photogrammetric data to segment and locate solar panels without manual annotation. In addition, the automatically labeled data set can be used to train deep learning models.
“This contribution speeds up the inspection process for large -scale installations by significantly reducing the time it is required for annotation of solar panels, making the training of deep learning detectors with minimal human intervention possible,” the team said. “The proposed workflow also ensures real -time applicability by achieving an optimal assessment between detection and inference time.”
The new labeling method uses UAV metadata such as GPS, Inertial Measurement Unit (IMU) and camera parameters to calculate the ground sample distance (GSD). It then uses the NIBLAC technique for automatic labeling to generate threshold values from the images and find solar panels. After this, the peripheral and clustering uses to verify the findings, comparing the dimensions with the actual dimensions of the solar panel. The extraction process contains a script to convert the coordinates into the format required by in -depth learning models.
This method was tested on two case studies in Morocco. The first was the Green Energy Park platform, including 22 kW-mounted monocrystalline panels with a tilt of 31 °. The second case study was from a plant on the roof of a Moroccan data center, with 1 MW monocrystalline panels. Thermal image acquisition was carried out using DJI’s Mavic 2 Enterprise Advanced (M2EA), equipped with both thermal and visual cameras, with resolutions of 640 x 512 and 8,000 x 6,000 respectively.
“It reached 91 % recall in the automatic geo-labeling step and considerably reduced false positives due to clustering and geometric limitations,” showed the results of the automated dataset. “Ultimately, this article improves the mentation and extraction of a few PV modules and accelerates O&M operations by taking up the geolocation of the modules extracted.”
The labeled images generated by the automatic process were subsequently split into training, validation and test sets and tested with different depleting models. The team tested the SSD RESNET50 V1, SSD Mobilenet V2, faster RCNN RESNET 50 V1, faster efficient D1, CenterNet HG104 and YOLOV7 models. In all cases, a batch size of 8 was used, while the training was given for 500 periods.
“The automatically geo-lanced data was used to train multiple depression detectors. YOLOV7 achieved the best performance among them, with an average average precision at 0.5 intersection over Union ([email protected]) from 98.33% with an inference time of 15 ms per image, which proves the adaptability for real -time inspection scenarios, “the team concluded.” In addition, the geolocation method reached the 2.51 m error, which supports accurate identification of the field level and supports the maintenance operations on the site. “
The new methodology was presented in “Fast and automatic solar module Geo-Labeling for optimized large-scale photovoltaic systems Inspection from UAV-Thermal images using the deep learning segmentation“Published in Cleaner engineering and technology. The research team included scientists from the Agronomic and Veterinary Institute Hassan II (IAV Hassan II) of Morocco, the research platform of Green Energy Park and the National Schools of Applied Sciences Oujda.
This content is protected by copyright and may not be reused. If you want to work with us and reuse part of our content, please contact: editors@pv-magazine.com.
