Digitizing a Durian Orchard: 243 Photos into a Living Map

Most orchards are documented the same way: a phone full of photos, a few spreadsheets, and the grower's memory. That works until you need to answer a specific question — how many trees are actually fruiting this month, which corner of the block is lagging, what last season's yield really was. This is the story of turning one working durian orchard into a digital twin you can query, using nothing exotic: a phone, GPS, and a vision model.

The raw material

The survey started as 243 geotagged photos walked across the orchard — trunks, canopies, hanging fruit, flower clusters on the bark. Every photo carried GPS coordinates in its EXIF metadata, which is the single most important detail: without location on each frame, a pile of photos is just a pile of photos.

243 GPS-tagged photos → 72 trees → 113 fruit and 45 flower clusters counted.

From photos to trees

The first step is clustering. Photos taken within a few metres of each other, of the same trunk, almost always belong to the same tree — so location plus a little judgment collapses 243 frames down to 72 distinct, individually tracked trees. Each tree gets its own page: its coordinates, its photos over time, and its counts.

Counting fruit and flowers

A vision model looked at every photo and counted two things durian growers care about: mature or maturing fruit, and the cream-coloured flower clusters that grow straight off the trunk and branches (durians are cauliflorous). Across all 243 photos the raw tally was 158 fruit and 59 flower clusters.

But a raw tally over-counts: photograph the same laden tree three times and its fruit gets counted three times. So counts are deduplicated per tree — each tree keeps its best (highest) single-photo count rather than the sum. That brought the orchard total down to 113 fruit and 45 flower clusters across 19 fruiting and 15 flowering trees. That deduped number is the one that reflects what's actually on the trees.

What the twin makes possible

• A satellite map where every marker is a real, numbered tree you can open.
• A harvest baseline — a defensible count you can compare against next season instead of guessing.
• Per-tree history: this tree flowered in this window, set this much fruit, lagged the block.
• A data model that's ready for sensors and repeat surveys without redoing the groundwork.

What we'd tell the next grower

1. 1Turn on location tagging before you shoot. A photo without GPS can't be placed.
2. 2Photograph the trunk, not just the canopy — durian flowers and much of the fruit hang low and close to the bark.
3. 3Take a few frames per tree from different angles; the model counts better with more looks, and dedup handles the overlap.
4. 4Don't chase a perfect count. A consistent, repeatable method beats a one-off perfect number, because the value is in the comparison over time.

None of this required a drone, a crew, or specialist gear — just a disciplined phone survey and a pipeline to turn it into structured data. That's the whole idea behind a digital twin for a tropical orchard: the orchard already tells you everything, if you capture it in a form you can actually query.