Skip to content

Concept Note

TraceTheSun - PV Density Estimation Applying PV data mapping methodologies that are cost effective, verifiable, community driven and localized.

July 24, 2026

Highlights Providing decision-ready PV density maps for grid planning, BESS deployment, policy and investment. Enabling transparent debate on data quality, uncertainty and discrepancies between non-verifiable estimates. Developing a scalable, cost-effective and verifiable methodology for global PV density estimation. Releasing open, reproducible PV maps, training data and AI models for independent verification and reuse. Empowering local communities to validate, update and maintain the data with limited resources. Accelerating the energy transition and reducing GHG emissions through faster, better and more cost-effective decisions. Target Country: Pakistan, Mexico, Vietnam, Bangladesh, Indonesia, India, Brazil, Philippines Partners: Lahore University of Management Sciences and Michigan State University

Background

The significant growth of photovoltaics around the world poses serious challenges for planners, policymakers and investors. For example, Pakistan's installed capacity is reported to range from 6.8 GW according to official figures to 47 GW according to NGO estimates. Currently, mapping methodologies rely heavily on high-resolution satellite imagery, which significantly drives the cost, complexity and inaccessibility of such projects. Commercial acquired satellite imagery cannot be shared with others, and most publicly accessible satellite imagery from sources such as Google Maps, Mapbox and ESRI are not allowed to be processed using AI methods. Previous methods using Sentinel 2, the only globally available, free, regularly updated imagery with sufficient resolution to detect PV installations, could only detect large-scale, isolated utility-scale PV installations.

This has created an environment in which only a single company with access to the imagery can update a dataset on PV distribution. Collaboration was not possible, and the results were not verifiable and therefore debatable. As a result, new estimates for PV distribution create recurring costs without the opportunity to discuss or verify this data with a third party. The lack of verifiability and the wide variation between estimates reduces trust in the data and hinders further improvements.

As part of a pilot project called TraceTheSun with 4 internships of the Lahore University of Management Sciences in Pakistan funded by Open Energy Transition (OET), a new AI assisted community approach has shown significant improvement in the estimation of PV installations. With the help of a state of the art open source geospatial foundation model and a team of human data annotators, it has been shown that the relative distribution of PV installation can be reconstructed with Sentinel 2, significantly increasing the detection of rooftop PV area detection threshold down to 400 m². For densely populated areas with high PV adoption it has been shown that even smaller rooftop solar installations can be detected by estimating the PV density directly rather than detecting individual installations.

This development was in particular possible by a community approach where multiple experienced mappers are regularly improving and validating the training data for the AI with the help of the various free satellite imagery accessible via the OpenStreetMap editor allowing for manual mapping and data annotation like ESRI, Mapbox, Bing. This leads to an iterative approach where AI is constantly suggesting new installations using low-resolution imagery and humans validate this data. This creates a nationwide high quality training dataset of utility scale, commercial and large residential PV installations, using OpenStreetMap as a mapping environment. Local Mappers are trained for this task and further add local context and knowledge. Furthermore various existing datasets are combined and validated against the free high resolution imagery in the OpenStreetMap editor. This imagery also enables more frequent observation every 15 days, which allows the growth of PV to be derived over time.

Objectives
The main objective of this project is to create an open verifiable, reproducible and globally adaptable workflow of rooftop solar installation mapping, targeting the following countries: Mexico, Japan, Korea, Indonesia, India, Brazil, South Africa and Nigeria.

  1. Create small, high-quality training datasets that represent different urban regions and contain many smaller installations. Use acquired high-resolution imagery or free high-resolution national imagery, if available. This dataset is used for training PV density estimations and to validate detections from low resolution imagery (1). In contrast to the high-resolution satellite images available by default in OpenStreetMap, this data may be processed using AI.

  2. Training of geospatial foundation model data from regions with exceptional data quality like Germany or UK, afterwards retraining with locally created training data. All detections and training will use global building datasets to limit the regions of detections.

  3. Multiple iterations between human mappers and AI retraining to constantly improve the training data and AI implementation towards better detection metrics. This is done until reasonable detection quality of representative validation regions created within objective (2) is achieved. All larger PV installations above a threshold >2000m² will be validated and mapped in OpenStreetMap.

  4. Combining AI generated data with manually validated data in OpenStreetMap to estimate nationwide PV density. This relative density estimation will be calibrated using regions where official comprehensive high resolution PV installation data is available like Germany, Switzerland or the United Kingdom. Validations and calibrations will also incorporate PV cells import/trading data into the country.

  5. (Optional) Considering local surface irradiance and typical characteristics for PV installations, the estimated local operational capacity is resolved for local grid cells of 0.1°x0.1°.

  6. (Optional) By using local surveys, housing density/ size datasets and poverty data, the distribution of PV based on different income classes is estimated.

  7. (Optional) Map the growth of PV across Pakistan from 2021 to 2026 in quarterly timesteps, improving the AI training data with OpenStreetMap mapping workflow.

  8. (Optional) Creating a public facing platform to explore the dataset under different considerations, including PV growth over time.


These are the first results of the PV distribution across Pakistan, derived from our new PV density methodology, defined in EarthPV, using free but low resolution Sentinel 2 imagery. (Right) After multiple iterations between human annotation and AI retraining, only 3191 larger rooftop solar installations are unmapped. (Left) Using our new PV density estimation, even very small rooftop solar installations could be identified in dense urban areas.

Our in-house developed solar glint method enables us to validate the existence of PV installations of up to 100 m² and estimate their tilt and orientation across a country. (Left) The tilt and orientation estimated with Sentinel 2 for 96 PV installations across Pakistan. (Right) The high-reflection glint effect is visible in Sentinel 2 imagery.

Deliverables

  1. AI Training Data for High Resolution, Low Resolution and PV density estimation. All training data will be released under an open license and where possible directly in OpenStreetMap.
  2. AI Models will be released under an open license with all code for preprocessing, training and post processing included.
  3. Educational and capacity building material on how to develop the complete mapping pipeline end-to-end, with specific regional workflows, imagery and datasets included.
  4. Fully reproducible PV maximum capacity map, derived from a combined dataset of human verified installations, AI detected installations, and estimated PV density for urban areas. All software and data to create the capacity maps will be released under an open licence. This includes detailed calibrations against import data, surveys, net-metered systems, and other data sources.
  5. (Optional) Estimated operational capacity under local weather trends.
  6. (Optional) Distribution between household income classes for residential rooftop solar.

Intended Impact and Project Sustainability

The project will be developed in close partnership with the TraceTheSun and OpenStreetMap communities. TraceTheSun is an emerging community that brings together the most prominent open-source projects in PV detection and the most skilled PV mappers in OpenStreetMap. Together, they address the challenges of tagging PV data worldwide in an open, verifiable and cost effective way. This community that is currently forming includes:

  • Open Energy Transition (currently unfunded)
  • Muhammad Awais of Lahore University of Management Sciences in Pakistan (currently funded 4 interns by Open Energy Transition until August)
  • Jake Stid, creator of GMSEUS, from the Michigan State University (currently funded by US grant with regional focus on North America)
  • Gabriel Kasmi, volunteer and creator of DeepPVMapper.

Our intended impact is to provide trustworthy, verifiable and cost-effective data that is the best available. This data can be used by policymakers, utilities, modellers and investors to make better operational decisions, as well as better planning and investment decisions. This will enable them to understand the financial and operational opportunity/risks of the energy transition powered by PV - reducing the uncertainty and unknowns surrounding the energy transition, accelerating it further.

In order to ensure the long-term sustainability of the project, maintenance costs will be reduced as much as possible. Additionally, the OpenStreetMap community will be empowered to reuse our tools. New PV leads will be added to various volunteering mapping platforms, such as Rapid, MapRoulette and StreetComplete. By establishing TraceTheSun as a global community comprising local data users, academic institutions, non-profit organisations and volunteers, we will be able to share funding, resources, data sources and methodologies more effectively, thereby reducing long-term costs.