New image dataset aims to improve AI-based navigation of PV cleaning robots

An international research team has released an open-access dataset of 6,200 annotated images of solar panels to improve visual navigation for autonomous photovoltaic cleaning robots. The dataset includes clean images, synthetically generated contamination at five severity levels, and laboratory images of physically soiled modules. According to corresponding author Kumeresan A. Danapalasingam, the dataset focuses on navigation cues rather than panel defect inspection, preserving consistent reference-line geometry across contamination levels. This allows researchers to systematically evaluate whether robotic vision algorithms can reliably identify navigation markers despite surface contamination.

The dataset was constructed using top-down images captured by an HF868-2 camera mounted 15 cm above PV modules on a cleaning robot. The camera used a 2.7 mm wide-angle lens with a 135-degree field of view, recording video at 320 × 240 pixels and 60 frames per second. Images were geometrically calibrated and resized to 160 × 120 pixels to meet onboard computational constraints. Researchers manually annotated 800 clean images, identifying three structural reference lines: white cell-gap lines, horizontal frame-cell border lines, and vertical frame-cell border lines.

The team processed each image through a synthetic augmentation pipeline to generate five additional versions with increasing soiling severity, including bird droppings, water stains, puddles, glare, and frame dust. Simulated contamination obscured reference lines from 1% to 50% depending on severity, while maintaining unchanged underlying geometry. Co-author Xin Li noted that different contaminants systematically obscure the same reference lines, enabling isolation of visual degradation effects from changes in viewing geometry.

The resulting dataset includes 4,800 training and validation images, divided into 3,846 training and 954 validation images, plus 1,200 testing images covering six clean and synthetic soiling conditions. It also features 200 additional images of physically contaminated panels with pigment droplets, sand, wastewater, and glare. The dataset is accompanied by annotation files detailing reference-line positions, orientations, endpoints, and class identifiers, totaling 6,200 images.