How to Estimate Mango Yield Per Tree Before Harvest

Why growers estimate (and why the usual method fails)

A pre-harvest estimate answers three practical questions: how many pickers and ladders you need for a block, how many field bins and crates to stage, and whether you can forward-sell a quantity with a straight face. Research groups who build yield-forecast systems list exactly these uses: harvest logistics, transport and storage, and marketing decisions such as forward selling.

The industry's default method, documented in an Australian study on mango yield estimation, is visual counting of fruit on a small number of sample trees at the stone-hardening stage, roughly six weeks before harvest. The same study names the problem: fruit number per tree is highly variable across the orchard, so a handful of hand-picked sample trees produces an estimate that is often, in the authors' words, highly inaccurate. Pick the wrong five trees and your forecast is off by half before the first ladder goes up.

The fix is not a better sample. It is counting every tree, cheaply, from photos you already know how to take.

What the research says a photo count can achieve

Photo-based fruit counting is not a gadget promise; it has been measured against actual harvest counts:

• A comparison of mango fruit-load estimation methods by researchers at Central Queensland University, the University of Sydney and the University of New England reported machine-vision estimates of fruit per tree against harvest counts at R² = 0.88 with an RMSE of 22.5 fruit per tree, and a weighted average error of about 6% against packhouse counts for their day-imaging system.
• A 2024 study using YOLO-based detection on mango photos reported 96.7% accuracy but 77.4% recall, meaning the detector found about three of every four fruit it should have. Field counts from any camera undercount; the question is by how much, and whether you correct for it.
• A Senegal study of 209 mango trees across three cultivars ('Kent', 'Keitt' and 'Boucodiékhal') found cylinder canopy volume was the structure variable most correlated with yield (r = 0.76), and models combining structure with a fruit-load index reached R² above 0.77 with RMSE between 20% and 29% against measured production.

Read those numbers together and you get the honest picture: a per-tree photo count with a correction factor will typically land within 10-30% of the real harvest at tree level, and tighter at block level, where errors across trees partly cancel. A 2024 review of satellite-based mango prediction found exactly that pattern: individual tree predictions had errors around 26%, while block-level predictions improved to about 10%. For planning labor and bins, block-level accuracy is what matters anyway.

The method, step by step

This is the phone-only version. It works for blocks of a few dozen to a few thousand trees.

1. 1Pick your estimation windows. The first useful window is 3-4 weeks after fruit set, when fruit have survived drop and are large enough to see. The second is stone hardening, the traditional window. If you only do one pass, do the second; if you do both, the change between them is itself information (fruit drop between windows).
2. 2Walk every row with your phone camera. Take 2-4 photos per tree: one of the whole canopy from the row side, plus close-ups of dense areas. Geotagged photos matter here, because they let software cluster photos to individual trees automatically instead of you sorting a camera roll by memory.
3. 3Count fruit per photo, not per glance. Either have software detect and count (that is what CanopyTwin's pipeline does: detection boxes, a cross-check, and a per-tree roll-up), or count on the phone screen yourself. The point is that every fruit in every frame is counted once, not sampled by eye.
4. 4Roll photos up to a per-tree count. Where two photos of the same tree overlap, use the maximum count across the tree's photos, not the sum. Summing overlapping photos double-counts fruit that appear in both frames.
5. 5Estimate the visible fraction once per season. Every photo undercounts fruit hidden inside the canopy. Calibrate it: pick one or two representative trees, run your photo count on them, then hand-strip and actually count the fruit at harvest. If the photo method said 140 and the tree really carried 210, your visible fraction is about 0.67. Apply that factor to every tree in the block.
6. 6Convert counts to kilograms with your own sample. At the estimation date, pick 15-20 representative fruit across the block and weigh them. Average fruit weight at harvest is cultivar- and season-specific; your own sample beats any table. Estimated kg per tree = corrected fruit count × your sample mean weight (with a small allowance for fruit that will still be dropped or cull-graded).
7. 7Write the number on the tree record. Per-tree estimates that live in a notebook are gone by next season. Stored per tree, they become the baseline next year's estimate is compared against, and the difference between years is often more diagnostic than the absolute number.

A worked example: one hectare of 'Keitt'

Say you manage 55 'Keitt' trees on roughly a hectare. For reference, a Senegal study measured 'Keitt' trees averaging about 206 kg per tree and 'Kent' about 43.5 kg per tree under West African conditions, which tells you how wildly cultivar and management move this number; your own trees are the only reliable baseline.

Your survey produces these counts on three of the 55 trees (the rest follow the same pattern). Each tree has two overlapping side-of-row photos, so you take the maximum photo count per tree, not the sum. You multiply by your visible fraction of 0.70 from last season's calibration to get corrected fruit per tree. At stone hardening you picked 18 fruit and they averaged 410 g, so you plan on 410 g × corrected count, minus about 5% for drop and culls.

• Tree M07: max count 160 → 112 corrected fruit → about 43 kg marketable
• Tree M12: max count 74 → 52 corrected fruit → about 20 kg marketable
• Tree M15: max count 231 → 162 corrected fruit → about 63 kg marketable

Summing the whole block gives you a harvest plan: three pickers for four days, 60 field bins staged, and a range you can give a buyer (say 4.8-5.5 tonnes) with a stated confidence. The range matters: quoting a point estimate from a 20%-error method is how growers end up renegotiating in the field.

Five mistakes that wreck an estimate

• Sampling only convenient trees. The whole point of photo surveys is that no tree is skipped. If you still sample, sample randomly, not the easy rows.
• Summing overlapping photos. Double-counting inflates estimates by 20-40% on dense canopies. Use the per-tree maximum where photos overlap.
• Skipping the visible-fraction calibration. An uncorrected photo count is not an estimate, it is a lower bound. One season of hand-count calibration trees fixes this permanently.
• Estimating too early. Counting at full bloom or peak flowering tells you potential, not crop; fruit drop between set and harvest routinely removes a third or more of set fruit. Estimate after the main drop, then refine.
• Treating cultivars the same. The Australian machine-vision comparison trained its detector on one orchard and applied it across others because mango cultivars differ in fruit presentation. Your correction factors are cultivar-specific; keep them separate in your records.

From estimate to forecast

A single estimate is a snapshot. A forecast is a series: flowering intensity recorded weekly, fruit set counts after the drop, a stone-hardening count, and a final pre-harvest pass. A review of mango harvest-forecast systems designed exactly this loop: weekly flowering estimates and pre-harvest fruit-load estimates feed a model that predicts harvest timing and load per block per week, which is what you actually need to book labor and sell forward. Each season you run the loop, the correction factors and the variety baselines get better, and the estimate stops being a guess with a spreadsheet attached.

Timing cheat sheet: 3-4 weeks after fruit set, first photo pass and rough block estimate. Stone hardening (about 6 weeks before harvest), main photo pass, visible-fraction check against prior season. 2 weeks before harvest, refine with fruit sizing on a sample. At harvest, record actual picked weight per tree and compute your error. Next season starts smarter.

How CanopyTwin fits in

CanopyTwin does the bookkeeping part of this method. You walk the rows photographing trees with your phone; GPS clusters the geotagged photos into individually tracked trees on a satellite map, so nothing gets double-counted or skipped. The AI counts fruit in each survey photo and rolls counts up per tree, keeping a photo history on every tree's record, and CSV exports let you carry the numbers into whatever harvest planning you already do. The Orchard tier covers fruit and flower counts from survey photos and CSV exports, and every plan starts with a 30-day free trial, no credit card up front.