Where Are the Solar Panels, Actually? ― Trying to Draw Their Outlines with AI

Hello from the MAPRISE team. This is a record of the solar panel detection work we've carried out across the Kyushu 7 prefectures.

The FIT program, and what happened over the last decade

Japan's Feed-in Tariff (FIT) program, launched in 2012, guarantees a fixed purchase price for electricity generated from renewable sources, with the aim of spreading renewable energy adoption. Under this program, a huge number of solar power facilities have been installed nationwide.

Looking at the certification data we've loaded from Japan's Ministry of Economy, Trade and Industry (METI), the Kyushu 7 prefectures alone have 63,416 registered small-scale facilities (10–50kW) and 176 registered large-scale mega-solar facilities (2MW and above). Even these numbers alone convey the scale of it.

We started wondering where these things actually are

We don't intend to debate the merits of the program itself. But checking locations one by one on the map, you notice the installation sites fall into a genuinely wide range of patterns.

Rooftops of factories and buildings, corners of reclaimed land, farmland that's been taken out of cultivation, former golf course sites — in these cases, it's easy to see them as legitimate land reuse.

On the other hand, when you spot panels cut into a mountainside or lined up neatly partway up a hillside, it's the kind of thing that makes you pause.

The image above shows a mountainous area in Kagoshima Prefecture via aerial photo. The area shaded in purple on the right is what the AI we'll describe below estimated as "panels are here." A sizeable area follows the valley, covering the slope.

Landslide risk, visual impact on the landscape, and what happens to the equipment 20 years from now — this isn't a question of good or bad. We think there's inherent value simply in knowing what exists somewhere. That's especially true if you're the one conducting real-estate due diligence on the area.

How we implemented this in MAPRISE ― three layers

In MAPRISE, we represent solar power facilities across three distinct layers.

First, there's point data for FIT-certified facilities. We display certification information published by METI, color-coded by output scale. This location data does have a limitation, though.

The registered address is often accurate only to the neighborhood (oaza) level, meaning the actual panel could be anywhere within a radius of a few hundred meters. We state this explicitly in the popup. We've also added a risk rating for whether the location falls within a forest area.

Second, there are polygons for panel installation footprints. We import outlines hand-drawn in OpenStreetMap and classify them as either ground-mounted or building-related.

These outlines are accurate, but only exist where someone has drawn them by hand. Coverage remains a real limitation.

Third is the AI-estimated coverage area — this is the part we actually want to talk about here.

When multiple pieces of this information overlap at the same location, we let you choose which one to view from a selection popup.

Drawing outlines with AI ― a roundabout development story

We feed aerial photography published by Japan's Geospatial Information Authority (GSI) — the orthophoto imagery from its digital base map — into a deep-learning model to estimate panel locations.

The prototype for this technique was something we sharpened back in 2019, when we entered a business plan competition hosted by a regional bank in the Kansai area. We'd already been interested in the direction of "reading ground features from satellite and aerial imagery" at that point, and that accumulated experience is what carried over into this project.

That said, the implementation was anything but smooth.

Initially, we planned to use a solar panel detection model that had been published in France and adapt it for Japan through transfer learning. That didn't work out. French aerial photography is captured at roughly 20cm ground resolution, a different condition from Japan's 50cm. On top of that, roof shapes and surrounding land use also differ. When we actually tested it, even tiles directly over real, existing mega-solar facilities were judged "not a panel" more than 98% of the time.

So we switched course and retrained the model from scratch on Japanese aerial photography. We collected, one by one, the ground features prone to false positives — forest canopy, tidal flats, aquaculture ponds, plastic greenhouses, bare land under development, quarries — added them to the training data, and re-measured accuracy each time.

Even so, we still needed a human eye at the end

Even after repeated training, the model alone never reached a usable level of accuracy on its own.

Taking the raw detection results at face value, only a few percent actually turned out to be real panels. Narrowing candidates down by area and shape regularity improved that to roughly 50%, but that still meant half were wrong.

The method we ultimately settled on was, honestly, quite primitive: visually checking each narrowed-down candidate against aerial photography, one by one. We manually judged several hundred cases across the Kyushu 7 prefectures this way.

Once we actually reviewed them, plenty of what the model flagged as "panel-like" turned out to be genuinely confusable things — orchard rows, the reflection off greenhouse plastic, ripples on the sea surface. Conversely, there were also plenty of cases where real panels were only partially captured.

We've been upfront about these limitations in the on-map popup too. The disclaimer "AI-estimated outlines are reference information, less accurate than hand-drawn ones" exists for exactly this reason.

Not just using open data as-is

MAPRISE draws on a lot of public open data, but we don't think simply laying it side by side creates value on its own.

FIT certification data only tells you location down to the neighborhood level, and OpenStreetMap outlines lack coverage. Each is hard to use effectively on its own. Only by adding AI estimation as a third layer, and overlaying all three, does a coherent picture finally emerge: what's here, roughly how large it is, and since when.

We think this solar panel detection work is a good example of that kind of extra processing step.

Above is the same location viewed simultaneously across four panels — map, aerial photo, terrain, and hillshade. You can't tell where on the slope the panels actually sit from the map alone.

For now, it's Kyushu only

To be honest, this feature is currently only available for the Kyushu 7 prefectures.

The reason is simple: GPU compute cost. Loading aerial imagery tile by tile and running each one through a deep-learning model adds up to a substantial amount of computation. Processing all of Kyushu alone meant handling several million tiles.

Using spot instances and reworking the model architecture cut costs significantly, but going nationwide is still an order of magnitude more.

Going forward, we'd like to gradually expand coverage starting from western Japan — taking our time and verifying accuracy as we go.

Disclaimer

The "panel coverage (AI estimate)" introduced in this post is an estimate derived by applying machine learning to aerial photography from the Geospatial Information Authority of Japan; it is not a formal survey. The outlines are not comprehensive, and the underlying aerial photography may be several years old depending on the region. Facilities installed in recent years may be absent from both the photography and the outlines. Please also check the FIT-certified point layer to confirm whether a facility exists.

Sources

• Renewable Energy Electronic Application website (FIT certification data): https://www.fit-portal.go.jp/
• Geospatial Information Authority of Japan tiles (digital base map orthophoto imagery, standard map, hillshade map): https://maps.gsi.go.jp/development/ichiran.html
• OpenStreetMap contributors (ODbL): https://www.openstreetmap.org/

Questions or feedback are always welcome at info@maprise.jp.