What we were building
TUAI Ventures' crop-vision pipeline flags healthy growth, early blight and harvest-readiness from field camera footage. It shipped first for onion smallholders under the seed.lab cohort, and the plan was to extend the same model to a pineapple pilot the following quarter.
The assumption
We assumed the classifier's learned features — edge detection, colour clustering, canopy segmentation — were general enough that pointing the same pipeline at a new crop would mostly need a fine-tuning pass on a modest new dataset, not a from-scratch retrain.
What actually happened
The first field trial against real pineapple rows told a different story. Precision on "harvest-ready" dropped from 91% on onions to somewhere in the high 50s on pineapple — barely better than guessing across three maturity classes. Pineapple crowns occlude the fruit differently than onion foliage occludes bulbs, and the colour signal for ripeness sits in a completely different part of the spectrum at the growth stage our cameras capture.
We tracked every retraining attempt as its own row so we could compare accuracy across crop type and model version without losing the history:
class ClassificationRun(models.Model):
crop_type = models.CharField(max_length=32) # 'onion' | 'pineapple'
model_version = models.CharField(max_length=16) # e.g. 'v3.2-pineapple'
accuracy = models.FloatField() # validation accuracy, 0-1
dataset_size = models.PositiveIntegerField()
created_at = models.DateTimeField(auto_now_add=True)
class Meta:
ordering = ["-created_at"]Once we retrained on a pineapple-only labelled set (roughly 1,800 field images across three growth stages), accuracy climbed back to 88% — comparable to the original onion model, just with its own weights.
Where this leaves us
We're keeping the onion and pineapple models as separate deployable versions rather than chasing one "universal" crop classifier, and any future crop addition to TUAI Ventures' pipeline now gets budgeted as a fresh labelling-and-training cycle by default, not an afterthought fine-tune.