pith. sign in

arxiv: 2604.23662 · v1 · submitted 2026-04-26 · 💻 cs.CV

SolarFCD: A Large-Scale Dataset and Benchmark for Solar Fault Classification in Photovoltaic Systems

Misbah Ijaz , Saif ur Rehman Khan , Abd Ur Rehman , Arooj Zaib , Sebastian Vollmer , Andreas Dengel , Muhammad Nabeel Asim This is my paper

Pith reviewed 2026-05-08 06:36 UTC · model grok-4.3

classification 💻 cs.CV
keywords solar PVfault classificationdatasetbenchmarkResNetdefect detectionthermal imagingphotovoltaic systems
0
0 comments X

The pith

SolarFCD merges three public PV datasets into a single 4435-image benchmark covering four defect classes in RGB and thermal imagery.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces SolarFCD to overcome the shortage of large annotated datasets for detecting faults in solar photovoltaic panels. It reconciles labels from three existing sources into healthy, surface obstruction, structural fault, and electrical fault categories, then splits and augments the data for training. Sixteen different classification models are evaluated, with ResNet101V2 achieving the highest scores of 86.68% accuracy and similar figures for other metrics while maintaining balance across classes. A reader would care because scalable automated inspection can help maintain the efficiency and safety of expanding solar energy installations worldwide.

Core claim

SolarFCD is a unified dataset of 4,435 images created by combining and reconciling three public datasets of solar panel defects in two modalities. After label mapping, duplicate removal, and augmentation, it is split 80:10:10 and used to benchmark sixteen architectures. ResNet101V2 delivers the top performance with 86.68% accuracy, 88.65% precision, 88.62% recall, and 88.17% F1-score, with per-class results varying by less than 1.2 percentage points.

What carries the argument

The SolarFCD dataset, formed through methodical label mapping and reconciliation across RGB/drone and thermal infrared sources into four unified classes.

If this is right

  • Researchers can train and compare new models against the provided baselines for PV fault detection.
  • Open availability of the dataset, annotations, and code supports reproducible experiments in solar panel inspection.
  • Balanced performance across defect classes indicates the dataset supports reliable multi-class classification.
  • Automated systems using these models could scale to inspect large solar farms efficiently.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Future work could explore fusing RGB and thermal modalities within single models for improved accuracy.
  • Applying the dataset to real-time drone-based monitoring systems might accelerate field deployment.
  • Extending the classes or adding segmentation labels could address more granular fault analysis.

Load-bearing premise

The process of mapping and reconciling labels from the three original datasets yields consistent and unbiased ground truth labels.

What would settle it

Re-annotation of a subset of SolarFCD images by independent experts revealing systematic label disagreements would undermine the benchmark reliability.

read the original abstract

The increasing global deployment of solar photovoltaic (PV) systems needs robust, scalable, and automated inspection technologies capable of detecting a wide range of panel flaws under a variety of operating situations. The lack of large-scale, multi-modal, publicly available annotated datasets is a major obstacle preventing advancement in this field. We introduce SolarFCD, an extensive dataset of solar panel defects created by methodically combining and reconciling three publicly accessible datasets covering two imaging modalities: RGB/Drone images and Thermal Infrared. The dataset consist of 4,435 images arranged under four unified defect classes such as: healthy images, Surface Obstruction, structural fault, and electrical fault. The dataset was divided into training, validation, and test splits at an 80:10:10 ratio through methodical label mapping, near-duplicate removal, and targeted augmentation of minority classes. Sixteen classification architectures from five design families were trained and assessed on the dataset to provide repeatable benchmark baselines. With an accuracy of 86.68%, precision of 88.65%, recall of 88.62%, and F1-score of 88.17%, ResNet101V2 performed the best overall. Per-class results showed balanced detection across all four defect categories within a narrow performance band of less than 1.2 percentage points. To promote open and repeatable research in automated PV inspection and solar energy operations and maintenance, the dataset, annotation files, and baseline code are made openly available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper introduces SolarFCD, a dataset of 4,435 images formed by reconciling three public sources (RGB/drone and thermal) into four unified classes (healthy, surface obstruction, structural fault, electrical fault) via label mapping, near-duplicate removal, and minority-class augmentation. An 80:10:10 train/val/test split is used to benchmark 16 CNN architectures; ResNet101V2 achieves the highest scores (accuracy 86.68%, precision 88.65%, recall 88.62%, F1 88.17%) with per-class performance varying by less than 1.2 percentage points. The dataset, annotations, and baseline code are released publicly.

Significance. If label quality is confirmed, the work supplies a much-needed public multi-modal benchmark for PV fault classification, an area where large annotated datasets have been scarce. Explicit release of the full dataset, annotation files, and reproducible training code is a clear strength that enables direct follow-on research and comparison.

major comments (2)
  1. [§3] §3 (Dataset Construction / Label Reconciliation): The description of 'methodical label mapping' across the three heterogeneous sources supplies neither an explicit mapping table, inter-source agreement statistics, nor any quantitative audit of how original defect categories were collapsed. Because the sources differ in imaging modality, annotation granularity, and environmental conditions, unquantified misalignment would directly bias the training distribution and the headline metrics (86.68% accuracy, balanced per-class band <1.2 pp). This validation step is load-bearing for all empirical claims.
  2. [§4] §4 (Experiments): The per-class balance claim and overall performance figures rest on the assumption that the 4,435 unified labels are accurate ground truth; without the missing reconciliation audit, the reported narrow performance band cannot be interpreted as evidence of robust detection across defect types.
minor comments (3)
  1. [Abstract] Abstract: 'The dataset consist of' should read 'consists of'.
  2. [§3] The three source datasets are referenced but their original papers and exact class mappings are not cited in a dedicated table; adding this would improve traceability.
  3. [§3] Figure captions for example images could explicitly note the original source and modality for each panel to aid reader inspection of cross-dataset consistency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough and constructive review of our manuscript. The comments on dataset construction and experimental validation are well-taken, and we agree that greater transparency in the label reconciliation process will strengthen the paper. We address each major comment below and will revise the manuscript to incorporate the requested details.

read point-by-point responses

  1. Referee: [§3] §3 (Dataset Construction / Label Reconciliation): The description of 'methodical label mapping' across the three heterogeneous sources supplies neither an explicit mapping table, inter-source agreement statistics, nor any quantitative audit of how original defect categories were collapsed. Because the sources differ in imaging modality, annotation granularity, and environmental conditions, unquantified misalignment would directly bias the training distribution and the headline metrics (86.68% accuracy, balanced per-class band <1.2 pp). This validation step is load-bearing for all empirical claims.

    Authors: We appreciate the referee highlighting the need for more explicit documentation of the reconciliation process. The manuscript describes the high-level steps (methodical label mapping, near-duplicate removal, and targeted augmentation), but we acknowledge that an explicit mapping table and quantitative audit were omitted. In the revised version we will add: (1) a table listing the original class labels from each of the three source datasets and their mappings to the four unified classes; (2) the number of samples contributed by each source to each class before and after processing; and (3) a description of the criteria used to collapse categories, including examples of how defect descriptions were aligned across modalities. Because the sources were annotated independently under different protocols, we did not compute inter-annotator agreement metrics such as Cohen’s kappa prior to mapping; we will add a limitations paragraph discussing this and the mitigation steps (manual review of a random subset of mapped images). These additions will allow readers to assess potential label noise directly. revision: yes

  2. Referee: [§4] §4 (Experiments): The per-class balance claim and overall performance figures rest on the assumption that the 4,435 unified labels are accurate ground truth; without the missing reconciliation audit, the reported narrow performance band cannot be interpreted as evidence of robust detection across defect types.

    Authors: We agree that the narrow performance band (<1.2 pp) and overall metrics can only be interpreted as evidence of consistent detection once the label quality is substantiated. The revisions described in response to the §3 comment—explicit mapping table, source-contribution statistics, and discussion of reconciliation limitations—will directly address this concern. In the revised experiments section we will cross-reference the new audit material and qualify the interpretation of the balanced per-class results accordingly, while retaining the reported numbers as the benchmark baselines. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical dataset construction and benchmarking

full rationale

The paper performs dataset reconciliation from three public sources followed by standard supervised training of 16 off-the-shelf architectures and direct reporting of test-set metrics. No equations, predictions, or derivations are present that could reduce to self-defined quantities, fitted parameters renamed as predictions, or load-bearing self-citations. The reported accuracies are empirical outcomes on a held-out split and remain independent of any author-internal definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the semantic fidelity of label mapping from three heterogeneous sources and on standard assumptions of supervised image classification; no new physical entities or fitted constants are introduced.

axioms (2)
  • domain assumption Label mapping across source datasets preserves consistent semantic meaning for the four unified classes
    Invoked when reconciling the three public datasets into healthy, surface obstruction, structural fault, and electrical fault categories.
  • ad hoc to paper Near-duplicate removal and targeted augmentation do not distort the underlying data distribution
    Applied during dataset preparation and minority-class balancing.

pith-pipeline@v0.9.0 · 5589 in / 1332 out tokens · 53005 ms · 2026-05-08T06:36:32.163470+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Environmental Research Letters16(6), 064017 (2021) https://doi.org/10.1088/1748-9326/ abfba1

    Johnson, N., Gross, R., Staffell, I.: Stabilisation wedges: measuring progress towards transforming the global energy and land use systems. Environmen- tal Research Letters16(6), 064011 (2021) https://doi.org/10.1088/1748-9326/ abec06

    work page doi:10.1088/1748-9326/ 2021

  2. [2]

    Sustainable Cities and Society 65, 102527

    Zhang, M., Zhang, D., Xie, T.: Balancing urban energy considering economic growth and environmental sustainability through integration of renewable energy. Sustainable Cities and Society101, 105178 (2024) https://doi.org/10.1016/j.scs. 2024.105178

    work page doi:10.1016/j.scs 2024

  3. [3]

    Technical report, International Energy Agency, Paris (2025)

    International Energy Agency (IEA): Renewables 2025. Technical report, International Energy Agency, Paris (2025). https://www.iea.org/reports/ renewables-2025

    work page 2025

  4. [4]

    Renewable and sustainable energy reviews 208, 115073 (2025) https://doi.org/10.1016/j.rser.2024.115073

    Bamisile, O., Acen, C., Cai, D., Huang, Q., Staffell, I.: The environmental factors affecting solar photovoltaic output. Renewable and sustainable energy reviews 208, 115073 (2025) https://doi.org/10.1016/j.rser.2024.115073

    work page doi:10.1016/j.rser.2024.115073 2025

  5. [5]

    Renewable and Sustainable Energy Reviews82, 743–760 (2018) https://doi.org/10.1016/j.rser.2017.09.042

    Said, S.A., Hassan, G., Walwil, H.M., Al-Aqeeli, N.: The effect of environmental factors and dust accumulation on photovoltaic modules and dust-accumulation mitigation strategies. Renewable and Sustainable Energy Reviews82, 743–760 (2018) https://doi.org/10.1016/j.rser.2017.09.042

    work page doi:10.1016/j.rser.2017.09.042 2018

  6. [6]

    Journal of Infrastructure Intelligence and Resilience2(3), 100051 (2023) https://doi.org/10.1016/j.iintel.2023.100051

    Malek, K., Rodr´ ıguez, E.O., Lee, Y.-C., Murillo, J., Mohammadkhorasani, A., Vigil, L., Zhang, S., Moreu, F.: Design and implementation of sustainable solar energy harvesting for low-cost remote sensors equipped with real-time monitoring systems. Journal of Infrastructure Intelligence and Resilience2(3), 100051 (2023) https://doi.org/10.1016/j.iintel.2023.100051

    work page doi:10.1016/j.iintel.2023.100051 2023

  7. [7]

    Renewable energy118, 452–467 (2018) https://doi.org/10.1016/j.renene.2017.10.053

    Das, S., Hazra, A., Basu, M.: Metaheuristic optimization based fault diagnosis 19 strategy for solar photovoltaic systems under non-uniform irradiance. Renewable energy118, 452–467 (2018) https://doi.org/10.1016/j.renene.2017.10.053

    work page doi:10.1016/j.renene.2017.10.053 2018

  8. [8]

    Energies14(22), 7770 (2021) https://doi.org/10

    Dhanraj, J.A., Mostafaeipour, A., Velmurugan, K., Techato, K., Chaurasiya, P.K., Solomon, J.M., Gopalan, A., Phoungthong, K.: An effective evaluation on fault detection in solar panels. Energies14(22), 7770 (2021) https://doi.org/10. 3390/en14227770

    work page 2021

  9. [9]

    Science of The Total Environment955, 176911 (2024) https://doi.org/10.1016/j.scitotenv.2024

    Wang, Y., Liu, B., Peng, H., Jiang, Y.: Locating the suitable large-scale solar farms in china’s deserts with environmental considerations. Science of The Total Environment955, 176911 (2024) https://doi.org/10.1016/j.scitotenv.2024. 176911

    work page doi:10.1016/j.scitotenv.2024 2024

  10. [10]

    Neural Computing and Applications36(27), 16769–16796 (2024)

    Shaban, W.M.: Detection and classification of photovoltaic module defects based on artificial intelligence. Neural Computing and Applications36(27), 16769–16796 (2024)

    work page 2024

  11. [11]

    Solar Energy 287, 113240 (2025) https://doi.org/10.1016/j.solener.2025.113240

    Hossain, S., Arika, A.M., Fahim, I.N., Uddin, J., Ahmed, A., Apon, H.J., Hoque, M.A.: Enhancing solar panel performance: A machine learning approach to dust detection and automated water sprinkle-based cleaning strategy. Solar Energy 287, 113240 (2025) https://doi.org/10.1016/j.solener.2025.113240

    work page doi:10.1016/j.solener.2025.113240 2025

  12. [12]

    Renewable Energy255, 123774 (2025) https://doi.org/10.1016/j.renene.2025

    Jonathan, A.L., Bamisile, O., Cai, D., Ejiyi, C.J., Nkou, J.J.N., Victor, K., Ukwuoma, C.C., Wei, L., Huang, Q.: A multimodal deep learning approach for very short-term solar forecasts using sky images and historical numerical data. Renewable Energy255, 123774 (2025) https://doi.org/10.1016/j.renene.2025. 123774

    work page doi:10.1016/j.renene.2025 2025

  13. [13]

    In: 2024 5th International Conference on Smart Electronics and Communication (ICOSEC), pp

    Mehta, S., Kundra, D.: Utilizing cnn-gan for enhanced detection and classification of dust on solar panels. In: 2024 5th International Conference on Smart Electronics and Communication (ICOSEC), pp. 915–919 (2024). https://doi.org/10.1109/ ICOSEC61587.2024.10722527 . IEEE

    work page arXiv 2024

  14. [14]

    Masset, R

    Ejiyi, C.J., Qin, Z., Ejiyi, M.B., Ukwuoma, C., Ejiyi, T.U., Muoka, G.W., Gyarteng, E.S., Bamisile, O.O.: Maccom: A multiple attention and convolutional cross-mixer framework for detailed 2d biomedical image segmentation. Com- puters in Biology and Medicine179, 108847 (2024) https://doi.org/10.1016/j. compbiomed.2024.108847

    work page doi:10.1016/j 2024

  15. [15]

    Case Studies in Thermal Engineering66, 105749 (2025) https://doi

    Tella, H., Hussein, A., Rehman, S., Liu, B., Balghonaim, A., Mohandes, M.: Solar photovoltaic panel cells defects classification using deep learning ensemble methods. Case Studies in Thermal Engineering66, 105749 (2025) https://doi. org/10.1016/j.csite.2025.105749

    work page doi:10.1016/j.csite.2025.105749 2025

  16. [16]

    Applied Energy 375, 124201 (2024) https://doi.org/10.1016/j.apenergy.2024.124201 20

    Araji, M.T., Waqas, A., Ali, R.: Utilizing deep learning towards real-time snow cover detection and energy loss estimation for solar modules. Applied Energy 375, 124201 (2024) https://doi.org/10.1016/j.apenergy.2024.124201 20

    work page doi:10.1016/j.apenergy.2024.124201 2024

  17. [17]

    Renewable Energy219, 119471 (2023) https://doi.org/10.1016/j.renene.2023.119471

    Guo, Z., Zhuang, Z., Tan, H., Liu, Z., Li, P., Lin, Z., Shang, W.-L., Zhang, H., Yan, J.: Accurate and generalizable photovoltaic panel segmentation using deep learning for imbalanced datasets. Renewable Energy219, 119471 (2023) https://doi.org/10.1016/j.renene.2023.119471

    work page doi:10.1016/j.renene.2023.119471 2023

  18. [18]

    Applied Energy395, 126132 (2025)

    Chen, X., Li, B., Braid, J.L., Byford, B., Colvin, D.J., Glaws, A., Jost, N., Pierce, B., Rabade, S., Springer, M.,et al.: Open data sets for assessing photovoltaic system reliability. Applied Energy395, 126132 (2025)

    work page 2025

  19. [19]

    Garladinne, H.S.: Solar Panel Dust Detection. Kaggle. Accessed: 2026-04-25 (2023)

    work page 2026

  20. [20]

    Afroz, P.: Solar Panel Images Clean and Faulty Images. Kaggle. Accessed: 2026- 04-25 (2023)

    work page 2026

  21. [21]

    Data in Brief57, 111184 (2024) https://doi

    Bello, R.-W., Owolawi, P.A., Wyk, E.A., Du, C.: Photovoltaic module dataset for automated fault detection and analysis in large photovoltaic systems using photovoltaic module fault detection. Data in Brief57, 111184 (2024) https://doi. org/10.1016/j.dib.2024.111184 21

    work page doi:10.1016/j.dib.2024.111184 2024