SPID: Solar Panel Image Dataset Overview

• SPID is a curated collection of annotated solar panel images from aerial, satellite, and UAV sources, enabling precise PV detection and mapping.
• It integrates diverse imaging modalities and rigorous annotation protocols to support semantic segmentation, object detection, and performance analytics.
• The dataset’s enriched metadata and benchmarking standards facilitate robust research in solar energy monitoring, defect detection, and infrastructure analysis.

A Solar Panel Image Dataset (SPID) is a curated collection of remote sensing imagery—often aerial, satellite, or close-range RGB and multispectral images—containing annotated photovoltaic (PV) panels, arrays, or farms. SPIDs are foundational for computer vision research in solar energy infrastructure monitoring, asset mapping, performance analytics, and defect detection. Recent SPIDs are characterized by rigorous annotation, extensive metadata, geospatial referencing, and benchmarking protocols targeting both methodologic advancement and applied energy deployment (Parhar et al., 2022).

1. Dataset Typology and Domain Scope

SPIDs span multiple image acquisition modalities and spatial scales:

• High-resolution aerial RGB datasets (e.g., HyperionSolarNet (Parhar et al., 2022); SPID-Fresno (Malof et al., 2016)) with sub-meter ground sampling distance (GSD ≈ 0.07–0.30 m/px), typically focused on urban or suburban regions for rooftop PV asset detection and mapping.
• Multispectral satellite-image datasets (e.g., GloSoFarID (Yang et al., 2024); Global Renewables Watch (Robinson et al., 19 Mar 2025)) covering continental or global extents, with per-pixel ground truth or regional polygon labels for utility-scale solar farm identification.
• Crowdsourced dual-provider aerial datasets (Kasmi et al. (Kasmi et al., 2022)) exploiting human annotation consensus on images from different acquisition chains (Google Earth Engine, national orthophotos) to study cross-domain generalization and calibration.
• UAV-obtained defect datasets (SPID-Solar Inspection (Rodrigo et al., 2 Sep 2025)) designed for close-up panel defect/contamination detection under controlled lighting, altitude, and environmental conditions.
• Time-series panel performance datasets (PV-Net (Mehta et al., 2017)) combining RGB panel imagery with environmental and power-output telemetry for soiling/defect impact analysis.

Spatial coverage ranges from small-city tests (e.g., Berkeley, CA in HyperionSolarNet) to global inventories (Global Renewables Watch, GloSoFarID) with temporal sampling from single campaigns to quarterly revisits across multiple years.

2. Annotation Methodologies and Data Schemas

Annotation protocols vary substantially by dataset objective and domain:

• Polygonal and pixel-wise masks: Per-pixel binary segmentation masks are standard for semantic segmentation tasks (Parhar et al., 2022, Kasmi et al., 2022, Yang et al., 2024, Malof et al., 2016), typically encoded as PNG (aerial) or GeoTIFF (satellite) with precise georegistration. Polygons are manually drawn (LabelBox, GIS GUI, crowdsourcing) or subject to consensus fusion (PAC method in Kasmi et al.).
• Classification labels: Binary ("solar" vs. "no_solar") assignments per image/tile for asset presence/absence (HyperionSolarNet, SPID-Fresno) or multi-class for panel state/defect type (SPID-Solar Inspection).
• Object detection (bounding boxes): COCO-style JSON entries indicate defects, contamination, or array units (Rodrigo et al., 2 Sep 2025), supporting evaluation at varying IoU thresholds.
• Metadata enrichment: Solar PV datasets increasingly include installation-level metadata—surface area, installed capacity (kWp), tilt/azimuth, installation date, integration type, self-consumption flag—matched to imagery and annotation (Kasmi et al., Global Renewables Watch).
• Weak supervision: Certain datasets (PV-Net (Mehta et al., 2017)) avoid explicit localization in favor of weakly supervised soiling impact through global power-loss labels and pseudo-mask fusion using network feature maps.
• Consensus-based quality assurance: Dual-phase and multi-annotator protocols maximize annotation fidelity, with kernel-density fusion and relative consensus thresholds (PAC ≥ 2.0, pixel inclusion ≥ 45%) (Kasmi et al., 2022).

3. Dataset Structure, Splits, and Access

SPIDs are generally organized by split type (training, validation, test) with rigorous statistics on class balance and distribution:

Dataset	Train	Validation	Test	Image Type	Classes	Geographic Focus
HyperionSolarNet (Parhar et al., 2022)	1,963	492	2,243	416/600px RGB	Solar/no_solar	14 US states, Berkeley CA
PV-Net (Mehta et al., 2017)	27,537	18,217	241 (segm.)	RGB + telemetry	Power-loss bins	Lab (panels)
SPID (defects) (Rodrigo et al., 2 Sep 2025)	700	150	100	UAV RGB 224/256px	Bird, Clean, Dusty, Electrical, Physical	Multi-regional
GloSoFarID (Yang et al., 2024)	≈9,592	≈2,055	≈2,056	256px, 13 bands	Solar farm mask	USA, Global
Crowdsource SPID (Kasmi et al., 2022)	—	—	—	400px, RGB	Segmentation mask	France
Fresno SPID (Malof et al., 2016)	40 tiles	—	20 tiles	5,000px, RGB	Per-pixel mask	Fresno CA
Global Renewables (Robinson et al., 19 Mar 2025)	—	—	—	4,096px, RGB	Polygon, area, capacity	Global

Licensing is generally open (CC BY 4.0, MIT), with some imagery subject to commercial restrictions (PlanetScope, Google Maps API).

Access modes include GitHub repositories (Yang et al., 2024, Rodrigo et al., 2 Sep 2025), Zenodo DOIs (Kasmi et al., 2022), dedicated project portals, or author request (Parhar et al., 2022).

4. Evaluation Metrics and Benchmarking

SPIDs support benchmarking across several canonical metrics aligned to the annotation schema:

• Semantic segmentation: Intersection over Union (IoU), pixel-wise F1 (Dice) score, precision, recall; for a predicted mask $P$ and ground truth $G$, $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}$ (Parhar et al., 2022, Kasmi et al., 2022, Yang et al., 2024).
• Object detection: Mean Average Precision (mAP) across classes and IoU thresholds (0.50:0.05:0.95); $$F_1 = 2 \frac{\text{precision} \times \text{recall}}{\text{precision}+\text{recall}}$$; Average Recall (AR) (Rodrigo et al., 2 Sep 2025).
• Classification: Accuracy, precision, recall, and class-specific $F_1$ scores (Parhar et al., 2022).
• Soiling/defect analysis: Power-loss quantification $$\text{SoilingImpact} = 100 \times ( P_{\text{reference}} - P_{\text{soiled}} ) / P_{\text{reference}}$$ (Mehta et al., 2017).
• Capacity estimation: Empirical or module-efficiency-based formulas, e.g., $$P_i = A_i \times \eta_\text{module} \times G_\text{ref}$$ for solar farm area-to-capacity conversion (Robinson et al., 19 Mar 2025).
• Baselines: A simple U-Net on Google images in Kasmi et al. achieves IoU ≈ 0.70–0.80; object-level and pixel-level PR curves are standard (Malof et al., 2016).

Reported benchmark results allow technical comparison of segmentation backbones (e.g., U-Net IoU=0.793 (Yang et al., 2024)) and object detectors (YOLOv3, Faster R-CNN, Swin Transformer in (Rodrigo et al., 2 Sep 2025)), with precise performance stratified by classes and test domains.

5. Technical Challenges and Robustness Considerations

SPID development and utilization entail specific challenges:

• Domain shift: Cross-provider imagery (Google vs. IGN in (Kasmi et al., 2022)), sensor type, and acquisition conditions lead to distributional variance requiring normalization and data augmentation (channel norm, random crops, flips, rotations, brightness adjustment).
• Class imbalance and sparsity: "solar" class sparsity in real-world test sets (14% solar in HyperionSolarNet Berkeley test), managed with hard-negative mining and augmentation strategies.
• Annotation error and uncertainty: Manual annotation is subject to false positives (e.g., roof windows), ambiguous panel appearances, and limited inter-annotator agreement statistics; consensus fusion and post-processing filters mitigate these.
• Spatial/geographic bias: Urban-centric datasets (Fresno, Berkeley), national coverage (France), or global scale (GloSoFarID, Global Renewables); limitations addressed by extension to new regions/sensors.
• Multispectral and 3D expansion: Most SPIDs are RGB-only; incorporation of Sentinel-2 bands or LiDAR data proposed for enhanced rooftop/type discrimination and domain adaptation.
• Temporal variability: Longitudinal datasets (Global Renewables Watch, GloSoFarID) enable construction date assignment and land-use change analysis. Panel defects/soiling impact can be tracked in close-range time series (PV-Net).

6. Applications and Future Directions

SPIDs underpin multiple research and policy endeavors:

• Automated rooftop/ground PV registry construction for grid integration planning and TSOs, utilizing segmentation and registry metadata (area, orientation, capacity) (Kasmi et al., 2022).
• Renewable energy analytics including global deployment tracking, capacity estimation, and land-use transition mapping at sub-national to planetary scales (Robinson et al., 19 Mar 2025).
• Defect detection and maintenance automation (object detection, UAV inspection) improving asset reliability (Rodrigo et al., 2 Sep 2025).
• Transfer learning and domain adaptation exploiting dual-provider and cross-modal datasets (Kasmi et al., 2022).
• Methodological research: Benchmarking novel architectures (e.g., Half-UNet (Yang et al., 2024)), multispectral fusion, weakly supervised localization (Mehta et al., 2017).
• Crowdsourcing strategy analysis: Annotation protocol evaluation and consensus approaches (Kasmi et al., 2022).

A plausible implication is that future SPIDs will more frequently incorporate multisensor fusion, richer temporal labeling, and expanded global coverage, including new annotation modalities (e.g., active learning) and cross-linked registries for performance tracking and verification.

7. Representative Datasets in the Literature

The following table organizes salient SPIDs and related resources:

Name / Paper	Coverage / Domain	Annotations	Licensing / Access
HyperionSolarNet (Parhar et al., 2022)	14 US states + Berkeley CA	RGB, solar/no_solar, masks	UC Berkeley: upon request, CC BY 4.0
PV-Net (Mehta et al., 2017)	Panel testbeds, lab imagery	RGB, power-loss, weak mask	https://deep-solar-eye.github.io/
GloSoFarID (Yang et al., 2024)	Global, Sentinel-2, 2021-23	13 bands, masks	GitHub/AWS S3, MIT
SPID defect (Rodrigo et al., 2 Sep 2025)	UAV over solar fields	RGB, COCO bbox/mask, 5 cls	GitHub, CC BY 4.0
Crowdsource SPID (Kasmi et al., 2022)	France, Google/IGN aerial	RGB, masks, meta	Zenodo, CC BY 4.0, CC0
SPID-Fresno (Malof et al., 2016)	Fresno CA, aerial RGB	Pixelwise masks, polygons	Figshare, CC BY 4.0
Global Renewables (Robinson et al., 19 Mar 2025)	Global, PlanetScope 2017-24	RGB, polygons, meta	Author portal, open for research

Each dataset is referenced as originally described, with strict adherence to licensing and access protocols, and is directly benchmarked in supporting literature.