Skip to content

Concept Note: TraceTheSun

Cost-effective, verifiable, community-driven photovoltaic (PV) capacity mapping from free satellite imagery.

Drafted July 2026. Results in this note reflect the atlas snapshot of 2026-08-20; see the live documentation for the current figures.

Highlights

  • Decision-ready PV maps of installed rooftop and ground-mounted capacity per 0.1° grid cell, for grid planning, battery-storage siting, policy and investment.
  • A transparent basis for debate on data quality, uncertainty and the large discrepancies between today's non-verifiable national estimates.
  • A scalable method for global PV capacity estimation that does not depend on commercial high-resolution imagery.
  • Fully open outputs: reproducible capacity maps, training data and AI models, released for independent verification and reuse.
  • Local ownership: communities can validate, update and maintain the data with limited resources.
  • Faster energy-transition decisions, and the reduced greenhouse-gas emissions that follow from better and cheaper planning.
Pilot country Pakistan (complete); Gujarat, India registered as a worked template
Next target countries Mexico, Japan, Korea, Indonesia, India, Brazil, South Africa, Nigeria
Partners Open Energy Transition; Centre for Water Informatics and Technology (WIT), Lahore University of Management Sciences; Michigan State University; DeepPVMapper
Code and results https://github.com/open-energy-transition/earthpv

Background

The rapid global growth of photovoltaics is a serious problem for planners, policymakers and investors, because nobody can say with confidence how much is installed or where. Pakistan's installed capacity, for example, is reported anywhere from 6.8 GW in official figures to 47 GW in NGO estimates.

Existing mapping methods depend on high-resolution satellite imagery, which drives up the cost and complexity of every such project and puts it out of reach for most countries. Commercially acquired imagery cannot be shared with third parties, and the publicly accessible high-resolution layers from Google, Mapbox and Esri may not be processed with AI. Sentinel-2, the only free, global, regularly updated imagery with enough resolution to see PV at all, had until recently only been used to detect large, isolated utility-scale plants.

The result is a market in which only a company that already holds the imagery can update a PV-distribution dataset. Nobody else can collaborate on it, reproduce it or check it, so each new estimate is a recurring cost that cannot be independently verified or improved. The gap between estimates, and the fact that none can be checked, erodes trust in all of them.

The TraceTheSun pilot, run in Pakistan by Open Energy Transition (OET) with a team of students at the Centre for Water Informatics and Technology (WIT), Lahore University of Management Sciences, has shown that this can change. Using an open geospatial foundation model and a team of human mappers, the pilot reconstructs the distribution of PV across a country from Sentinel-2 alone, bringing the size at which an individual rooftop array can be outlined down to about 400 m2. Below that size, in densely built areas with high adoption, the pilot estimates PV density per building directly instead of trying to outline each installation, which recovers much smaller rooftop systems. The published national result for Pakistan is a best estimate of 24,330 MWp of rooftop and ground-mounted capacity, with a 90% range of 20,822 to 33,582 MWp, built from 15,642 individually hand-mapped and verified OpenStreetMap installations plus the model's own recall-corrected detections.

What makes this work is the community loop. The AI reads only Sentinel-2 and proposes candidates; experienced OpenStreetMap mappers check each one against the licensed high-resolution layers in the OpenStreetMap editor (Esri, Mapbox, Bing) and map what is real; those verified installations become the next round of training data. Local mappers are trained for the task and add context no satellite carries, such as which roof belongs to a factory or a school. Existing PV datasets are folded in and validated the same way. Because Sentinel-2 revisits every five days, the same loop can also track how PV deployment has grown over time.

Objectives

The goal is an open, verifiable, reproducible and globally adaptable workflow for mapping rooftop and ground-mounted solar, with Pakistan as the completed pilot and Mexico, Japan, Korea, Indonesia, India, Brazil, South Africa and Nigeria as the next target countries.

  1. Build small, exhaustively mapped training areas that cover a range of urban and rural landscapes and include many small installations, using acquired or free national high-resolution imagery where it exists. These areas train the PV-density estimator and validate the detections made from Sentinel-2. Unlike the high-resolution basemaps inside the OpenStreetMap editor, this imagery may be processed with AI.

  2. Pre-train the geospatial foundation model on regions with near-complete public data, such as Germany or the UK, then retrain on the locally created data. Detection is restricted to built-up areas using global building-footprint datasets.

  3. Iterate between human mappers and AI retraining until detection quality on the representative validation areas from objective 2 is good enough. Every large PV installation above roughly 2,000 m2 is validated and mapped in OpenStreetMap along the way.

  4. Combine the AI detections with the human-verified OpenStreetMap data to estimate national PV density, calibrated against countries with complete official high-resolution registers (Germany, Switzerland, the UK) and cross-checked against national PV-module import and trade data.

  5. (Optional) Convert relative density into operational capacity per 0.1° × 0.1° grid cell, using local surface irradiance and typical installation characteristics.

  6. (Optional) Estimate how PV adoption is distributed across income classes, using local surveys, housing density and size datasets, and poverty data.

  7. (Optional) Map the growth of PV across Pakistan from 2021 to 2026 in quarterly steps, feeding the same OpenStreetMap mapping loop.

  8. (Optional) Build a public platform for exploring the dataset, including PV growth over time.

First results for Pakistan, produced with the PV-density method described in EarthPV from free, low-resolution Sentinel-2 imagery. (Right) After many iterations of human mapping and AI retraining, 15,642 larger rooftop and ground-mounted installations have been verified in OpenStreetMap and only a few thousand large arrays remain unmapped. (Left) The PV-density estimate recovers even very small rooftop systems in dense urban areas.

The project's solar-glint method independently confirms PV installations as small as 100 m2 and recovers their tilt and orientation across a whole country. (Left) Tilt and orientation estimated from Sentinel-2 for 290 installations across Pakistan. (Right) The glint reflection as it appears in Sentinel-2 imagery.

Deliverables

  1. Training data for high-resolution, low-resolution and PV-density estimation, released under an open licence and, where possible, directly in OpenStreetMap.
  2. AI models, released under an open licence with all preprocessing, training and postprocessing code.
  3. Educational and capacity-building material on building the whole pipeline end to end, with regional workflows, imagery and datasets.
  4. A fully reproducible capacity map, combining human-verified installations, AI detections and estimated PV density, with all software and data released under an open licence, including the calibrations against import data, surveys and net-metered systems.
  5. (Optional) Operational capacity under local weather patterns.
  6. (Optional) PV adoption by household income class for residential rooftop solar.

Intended impact and sustainability

The project is built with the TraceTheSun and OpenStreetMap communities. TraceTheSun is an emerging community that brings together the leading open-source PV-detection projects and the most experienced PV mappers in OpenStreetMap to tag and map solar worldwide in an open, verifiable and cost-effective way. It currently includes:

  • Open Energy Transition -- runs EarthPV and funded the Pakistan pilot; seeking funding to continue and scale.
  • Muhammad Awais and the student team at the Centre for Water Informatics and Technology (WIT), Lahore University of Management Sciences -- delivered the Pakistan pilot's mapping, validation and ground-truth work under an OET-funded internship that ran through August 2026.
  • Jake Stid at Michigan State University, creator of GMSEUS -- funded by a US grant, with a regional focus on North America.
  • Gabriel Kasmi, creator of DeepPVMapper -- volunteer.

The intended impact is to give policymakers, utilities, energy modellers and investors PV data they can actually verify, so they can plan, operate and invest with less uncertainty about where solar is and how fast it is growing, and move the energy transition along faster.

To keep the project sustainable, maintenance cost is held as low as possible and the OpenStreetMap community is equipped to reuse the tools directly, with new PV leads pushed to volunteer platforms such as Rapid, MapRoulette and StreetComplete. Establishing TraceTheSun as a global community of local data users, academic institutions, non-profits and volunteers lets the partners share funding, staff, data and methods, which lowers the long-term cost for everyone.