Where the reports come from
Twice a day, a scheduled job collects recent public photos of each plant from
iNaturalist, worldwide. Only photos whose Creative
Commons licence allows showing them are used. For world-famous viewing spots with few
photos, it also searches the web in the spot's own language (through SerpApi), for example
嵐山 紅葉 見頃
.
Who decides the stage
Gemma, an open-weight AI model, runs on the job's own free cloud machine. It looks at each photo and answers: is this really the plant, and is it green, turning, at peak or past peak (or, for cherry blossoms: buds, first bloom, full bloom, falling petals)? It also reads web snippets in any language. Photos are never sent to a third-party AI service.
Reports are grouped into areas about 5 km across. Fresh, confident reports count more. With fewer than two reports, a place shows Not enough reports; Peak Season never guesses.
Testing before anything goes live
A plant only appears after its prompt agrees with iNaturalist's community annotations ("Green Leaves", "Colored Leaves", "Flowers") on at least 20 photos, at 70% or better.
- Maples: first prompt 62% (failed); improved prompt 81%, and 81% again on 26 older photos it had never seen.
- Cherry blossoms: 92%. Caveat: none of the test photos showed buds only, so "buds" versus "bloom" has not really been tested.
Known limits
- The free machine is slow (about 45 seconds per photo), so each run checks about 110 photos; coverage grows run by run.
- Several photos from the same person count as several reports. Often that's one tree photographed a few times.
- Most photos come from North America and Europe; famous spots in Asia rely more on web reports.
- AI can be wrong. Every card links to the original photos so you can check.
Photos, data and privacy
Photos are never copied or stored here. Each one is shown from its original host with its photographer's credit and licence, and links to the original post. Your location stays in your browser; the site has no accounts, no cookies and no trackers.
Who made it
Built by klee1611 for the DEV Hacktoberfest Open-Source AI Challenge (October 2026). The code, prompts and validation results are open source on GitHub (MIT licence).