The IRPVS dataset

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a benchmark for long-tail segmentation and few-shot learning in industrial PV soiling detection.

🏢実務担当者:Enables precise soiling ratio computation for predictive maintenance of rooftop PV systems, reducing downtime and cleaning costs.

🏛政策担当者:Highlights the importance of AI-enabled monitoring for renewable energy infrastructure, informing policies on solar asset management.

1. Research Background and Purpose of Data GenerationSurface soiling of photovoltaic (PV) modules is a major cause of power output loss and irreversible degradation. In industrial parks, contaminants originate from diverse sources—ferrous dust from metal fabrication, atomised coating particles from paint shops, and corrosive soot from chemical facilities. These industrial pollutants adhere strongly to glass surfaces due to electrostatic bonding, oily matrices, or partial curing, making natural rainfall ineffective for self‑cleaning. Most existing public PV imagery datasets focus on installation detection in residential or commercial areas, with soiling types limited to natural dust, bird droppings, pollen, and leaves; annotations are typically at the bounding‑box or image‑level, lacking pixel‑level instance segmentation for process‑specific industrial contaminants.To bridge this gap, we constructed the Industrial Rooftop PV Soiling (IRPVS) dataset—the first publicly available benchmark that provides pixel‑level instance segmentation masks for process‑specific industrial soiling in a rooftop industrial setting. This dataset supports the training and evaluation of semantic segmentation models, enabling precise pixel‑wise computation of the soiling ratio, which can be further integrated with physics‑based hot‑spot risk models to facilitate predictive maintenance strategies for industrial PV systems.2. Data Acquisition MethodsTime and location: All imagery was acquired on 20 March 2025, between 10:00 and 14:00 local time (UTC+8), over an industrial facility at No. 2 Chunzhou Road, Sanshan District Economic Development Zone, Wuhu, Anhui Province, China (31.23°N, 118.27°E). The site hosts a mixed cluster of manufacturing plants—including chemical processing, mechanical fabrication, and coating/painting operations—with rooftop grid‑connected monocrystalline silicon PV arrays totalling approximately 35 MWp.Meteorological conditions: Clear‑sky conditions prevailed during acquisition, with ambient temperatures ranging from 8 °C to 21 °C, a westerly wind of about force 4 on the Beaufort scale (~7.9 m s⁻¹ at 10 m reference height), and no precipitation in the 72 hours prior to the flight, ensuring that soiling deposits remained in their naturally accumulated, undisturbed state.Equipment and flight parameters: A DJI Mavic 3 UAV, equipped with a 4/3 CMOS sensor (effective resolution: 20 megapixels; output: 3840 × 2160 pixels; lens focal length: 24 mm) and a mechanical shutter, was used. The flight was fully autonomous with a double‑grid (cross‑hatch) pattern. Key parameters: flight altitude 30 m, ground sampling distance (GSD) ≈0.55 cm/pixel, gimbal pitch 30°, forward overlap 80%, side overlap 75%, shutter speed 1/1800 s, ISO 100, white balance locked at mission start, and JPEG format. Approximately 8,012 raw frames were acquired, and after ortho‑mosaicking, they were tiled into 8,012 images of 3840 × 2160 pixels.3. Annotation MethodsTaxonomy of contaminant classes: Three classes were defined based on visual appearance, industrial provenance, and hazard potential:Dust (Class 1): Visually characterised by a diffuse, low‑to‑moderate contrast grey veil with gradual edge transitions. It originates from grinding, crushing, material handling, and general industrial activity. Its hazard features include reduced broadband transmittance and acting as a substrate for moisture retention.Redpaint (Class 2): Visually characterised by sharply delineated, high‑chroma orange‑red patches with a semi‑gloss surface texture. It originates from spray painting, coating, and surface‑finishing operations. Its hazard features include high opacity (complete photon blockage), strong adhesion, and potential chemical reactions with anti‑reflective coatings.Soot (Class 3): Visually characterised by white‑to‑light‑grey, moderately absorbing, matte deposits with intermediate edge gradients (sharper than Dust, more gradual than Redpaint). It originates from combustion exhaust, furnace flues, and foundry vents. Its hazard features include strong light scattering due to high specific surface area, and acidic components that accelerate glass surface corrosion under humid conditions.Annotation tool and format: Annotations were produced using Labelme (v5.1.1), an open‑source graphical tool for polygon delineation. The output format is COCO JSON (compliant with COCO 1.0 specification). Each annotation entry records the image file identifier, category ID, and an ordered list of (x, y) polygon vertex coordinates in pixel units.4. Data File StructureThe dataset root directory is named datasets and is organised as follows:datasets/├── img/             # Raw image files (JPEG)├── json/             # Original COCO JSON annotation files└── seg/             # Core data for segmentation tasks  ├── images/          # Image files  │  ├── train/        # Training set (6,409 images)  │  ├── val/         # Validation set (801 images)  │  └── test/         # Test set (802 images)  └── labels/          # YOLO‑format polygon segmentation labels (TXT)    ├── train/        # Paired with train/ images    ├── val/         # Paired with val/ images    └── test/         # Paired with test/ imagesFile naming conventions:Image files: Named as [sequence].jpg (e.g., 000001.jpg), where the sequence is a five‑digit number from 000001 to 008012. All images are JPEG, 3840 × 2160 pixels.YOLO label files: Each has the same base name as its corresponding image but with the .txt extension (e.g., 000001.txt), and is placed in the corresponding subdirectory under seg/labels/.COCO JSON annotation files: Stored in the json/ directory, containing complete instance‑level polygon information.5. Dataset StatisticsThe dataset comprises 8,012 images with a total of 39,212 soiling instances. The distribution per class is as follows:Dust: 37,811 instances, appearing in 7,842 images (96.4% of all images).Redpaint: 738 instances, appearing in 316 images (1.9% of all images).Soot: 663 instances, appearing in 287 images (1.7% of all images).This distribution is not a sampling artefact but reflects the real physical phenomenology: Dust is a pervasive background aerosol in manufacturing environments, while Redpaint and Soot are episodic and process‑proximate. This long‑tailed distribution mirrors real‑world inspection challenges, where rare but high‑consequence contaminant types must be detected against a ubiquitous background.The dataset is split into training (6,409 images, 80.0%), validation (801 images, 10.0%), and test (802 images, 10.0%) via stratified random sampling at the image level based on class presence, ensuring that tail classes (Redpaint, Soot) are represented in all splits.6. Data Formats and Software CompatibilityImage format: JPEG (.jpg), 3840 × 2160 pixelsAnnotation formats: COCO JSON (.json) and YOLO‑format (.txt)Compatible frameworks: Directly usable with YOLO26‑seg, Mask R‑CNN, Detectron2, MMDetection, and other mainstream instance segmentation frameworksRecommended tools: Labelme (v5.1.1 or later) for viewing/editing polygons; PyTorch, Ultralytics YOLO, etc., for model training7. Data Quality ValidationWe benchmarked the IRPVS dataset using the YOLO26‑seg instance segmentation architecture. Experiments were performed on a workstation with an Intel Core i9‑14900K CPU, 64 GB DDR5‑5600 RAM, and a single NVIDIA GeForce RTX 5090 GPU (32 GB GDDR7 VRAM), using PyTorch 2.5.0. The model was initialised with COCO‑pretrained weights and trained for 300 epochs at an input size of 640 × 640 pixels.Segmentation performance on the test set (IoU threshold = 0.70) is as follows (all values in percentage):Dust: Precision 87.95%, Recall 85.05%, IoU 76.17%, Dice 86.48%, mAP50 83.87%.Redpaint: Precision 96.11%, Recall 93.24%, IoU 89.85%, Dice 94.65%, mAP50 94.26%.Soot: Precision 95.64%, Recall 96.74%, IoU 92.65%, Dice 96.19%, mAP50 97.06%.Macro‑average: Precision 93.23%, Recall 91.68%, IoU 86.22%, Dice 92.44%, mAP50 91.73%.The results show that Soot achieves the highest performance due to its distinctive features and relatively large instance areas; Redpaint, though sharply bounded, suffers from limited training samples (738 instances); Dust, despite abundant data, is the most challenging because of its small area and low contrast with the aluminium frame edges of PV panels.8. Data Value and Reuse PotentialThe IRPVS dataset offers significant value in three domains:(i) Scientific contribution: It is the first public resource providing pixel‑level instance segmentation masks for process‑specific industrial soiling in a rooftop industrial environment, filling a critical gap in existing PV imagery datasets.(ii) Technical enablement: It serves as a standardised benchmark for long‑tail semantic segmentation, few‑shot learning, domain adaptation, and physics‑informed predictive maintenance modelling. In particular, the Dust category—characterised by small areas and low contrast—presents a challenging testbed for few‑shot and hard‑example mining methods.(iii) Practical applicability: The pixel‑wise soiling masks output by segmentation models enable precise computation of the soiling ratio (SR = contaminated area / total module area) for arb

• scidb https://doi.org/10.57760/sciencedb.44796first seen 2026-08-04 06:00:36