Rooftop Solar Panel Mapping from Drone and Aerial Imagery for Electricity Distribution Planners

The problem: solar nobody can see

Rooftop solar has spread rapidly across Pakistan in the last few years as households and businesses respond to high tariffs and load-shedding. Some systems are registered under net metering; many are not. For distribution engineers this creates a blind spot. At midday, a street full of panels can push power back towards the transformer, raise voltages and confuse load forecasts. Planners need to know where the panels are, and how much capacity sits behind each transformer, before they can plan upgrades or protection.

Counting panels door to door is slow. From above, panels are easy to see.

What already exists and the gap

• Google Project Sunroof estimates rooftop solar potential, in the United States and some other countries, rather than mapping installed panels.
• Research systems map installed panels from satellite imagery at national scale in the United States and elsewhere.
• Public datasets of labelled solar panels exist at several image resolutions.

The gap is an open pipeline that a distribution company, housing society or researcher in Pakistan can run on drone or aerial imagery of a defined area, producing panel counts and capacity ranges per building and per transformer area, with honest uncertainty.

Inspiration

"DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States" by Jiafan Yu, Zhecheng Wang, Arun Majumdar and Ram Rajagopal at Stanford, published in Joule in 2018, located solar installations across the contiguous United States from satellite imagery and estimated their size. For training data, "Multi-resolution dataset for photovoltaic panel segmentation from satellite and aerial imagery" by Hou Jiang and colleagues, published in Earth System Science Data in 2021, provides labelled panels at 0.8 m, 0.3 m and 0.1 m resolution, including rooftop panels on several roof types.

What you build

• Distribution planners draw a transformer service area on the map and see the number of solar buildings, total panel area and a capacity range.
• Housing societies and campuses get an inventory of their own roofs.
• Researchers use the pipeline to study how fast solar is spreading in a neighbourhood.

Every result shows the image date, because a map of panels is only true for the day the images were taken. Each detection carries a confidence score, and low-confidence shapes are shown differently so a planner can check them on the image before counting them. Totals are always given as ranges with the assumption written next to them, so nobody mistakes an estimate for a meter reading.

The technically interesting part

• Resolution and domain. Public datasets come from other countries and roof types. Flat concrete roofs with water tanks, satellite dishes and shade structures are common here and easily confused with panels. Fine-tune on local labels and report the confusions.
• From pixels to kilowatts. Panel area is measurable; capacity is not, directly. Convert area to a capacity range using typical module power per square metre and show the assumption, instead of a single precise-looking number.
• Buildings and areas. Joining panels to footprints and summing by area seems simple until footprints are missing or misaligned. Handle it explicitly.
• Imagery. Your results depend on the images you are allowed to use. Plan this first.

How to evaluate it

• Segmentation quality: intersection over union and per-building detection on held-out local tiles.
• Count accuracy: compare the system with a hand survey of buildings with panels in at least one neighbourhood.
• Capacity check: for a few buildings whose owners share their system size, compare the estimated range with the real value.
• Processing time for one square kilometre.

Suggested plan

• Weeks 1 to 3: secure imagery, for example drone flights over your campus or a housing society with written permission, read both papers.
• Weeks 4 to 7: train on the public dataset, label local tiles, fine-tune.
• Weeks 8 to 10: polygons, footprints, capacity ranges and area totals.
• Weeks 11 to 13: web map and the hand survey evaluation.
• Weeks 14 to 15: report and demo.

Path to a product

Distribution companies, regulators, solar installers and researchers all want to know where solar is installed. A mapping service for defined areas, sold per survey, is realistic, and the same pipeline helps installers find leads and housing societies plan their own grids. Energy and geospatial machine learning skills are in growing demand.

Scope warning

Imagery terms matter: do not train on or redistribute images from map services whose terms forbid it. Use your own drone flights with permission or openly licensed imagery. Do not try to cover a city. Do not publish building-level results for private homes; share totals by area.

Learn what you need on Sefism

• Machine learning roadmap: segmentation training and evaluation.
• Data science roadmap: turning detections into estimates with honest uncertainty.
• PyTorch: Learn the Basics: training the segmentation model.
• Computer Vision: Algorithms and Applications, 2nd Edition: segmentation background.
• Label Studio: drawing panel polygons on local tiles.

Get help with this idea

The risk is stalling on imagery permissions or overclaiming precision. Tauseef can help you secure an area to survey and frame results that planners will trust. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session