Researchers have created a dataset of 6,200 solar panel images designed to support AI-based navigation of cleaning robots. The dataset features annotated panel reference lines under clean, synthetically soiled, and laboratory contamination conditions, with the aim of helping developers assess the robustness of visual navigation algorithms.
An international research team developed an open-access image dataset to support visual navigation for autonomous PV cleaning robots under different soiling conditions. The dataset comprises 6,200 annotated images of solar panels, including clean images, synthetically generated contamination at five severity levels, and laboratory images of physically soiled modules.
“The novelty of our research lies in developing a photovoltaic image dataset specifically for the visual navigation of cleaning robots, rather than conventional panel defect inspection,” corresponding author Kumeresan A. Danapalasingam told pv magazine. “By preserving the same reference-line geometry across different levels of simulated soiling, we provide researchers with a systematic way to evaluate whether robotic vision algorithms can reliably identify navigation cues even when the panel surface is contaminated.”
To construct the dataset, the researchers captured top-down images of PV panels using an HF868-2 camera mounted at the front of a solar panel cleaning robot, approximately 15 cm above the module surface. The camera featured a 2.7 mm wide-angle lens with a 135-degree field of view and recorded video at 320 × 240 pixels and 60 frames per second. The images were geometrically calibrated and resized to 160 × 120 pixels to accommodate the computational constraints of onboard robotic vision systems.
The team manually annotated 800 clean images, identifying three types of structural reference lines used for robot navigation: white cell-gap lines, horizontal frame-cell border lines, and vertical frame-cell border lines. We then processed each image using a synthetic augmentation pipeline that generated five additional versions with increasing soiling, from very light to extreme. The simulated effects included bird droppings, water stains, puddles, glare, water accumulation along module frames, and frame dust, with targeted reference-line coverage ranging from 1% to 50% depending on the severity level.
“One particularly interesting aspect of our research is that different forms of contamination can systematically obscure the same photovoltaic reference lines while their underlying geometry remains unchanged,” co-author Xin Li said. “This allows us to isolate the effects of visual degradation from changes in the robot’s viewing geometry. We believe this provides a useful foundation for understanding how visual navigation algorithms fail under challenging conditions and how their robustness can be improved.”
This process produced 4,800 images for training and validation, divided into 3,846 training images and 954 validation images. For testing, the researchers assembled 1,200 images across six clean and synthetic soiling conditions, supplemented by 200 additional images captured on two separate PV panels physically contaminated with pigment droplets, sand, wastewater, and clean water, as well as exposed to high-intensity lighting to simulate glare. The resulting dataset comprises 6,200 images with corresponding annotation files containing reference-line positions, orientations, endpoints, and class identifiers.
“Our next objective is to move beyond dataset development towards reliable autonomous navigation of photovoltaic cleaning robots,” the academics concluded. “We are investigating lightweight artificial intelligence models for estimating the robot’s heading relative to panel reference lines, as well as localization methods that combine visual information with inertial and wheel-encoder measurements. Ultimately, we aim to validate these approaches on physical robotic platforms under realistic operating conditions.”
The paper “A robot-view image dataset of photovoltaic panel reference lines with paired synthetic soiling for cleaning robot navigation” has appeared in Data in Brief. Researchers from the University of Technology Malaysia, Canada’s University of Prince Edward Island, Belgium’s Thomas More University of Applied Sciences, Western Norway University, and the Indian university Amrita Vishwa Vidyapeetham have participated in the study.
This content is protected by copyright and may not be reused. If you want to cooperate with us and would like to reuse some of our content, please contact: [email protected].