Scientific ML: Pasture biomass regression

The agent is given top-view pasture photographs with measured dry-biomass targets, and builds a model predicting five biomass components per image. Training, validation strategy, and architecture are its own; the environment is offline except for a submission gateway.

Grazing and land-management decisions run on biomass estimates, and the direct way to measure biomass is to cut, dry, and weigh the grass. That is slow, destructive, and doesn't scale past a handful of plots. If a model can read dry-matter composition off a photograph, the measurement becomes free.

Predictions are scored by a globally weighted R squared over all image-target pairs. Submissions go through a sealed leaderboard: a capped number of scored submissions per day returns a public score only, a hidden split decides the final result, and the reward is the private score of the best-public submission, normalized against a measured floor.

The graded runs converged on one recipe: ResNet-50 and small ConvNeXt backbones, ensembled. Between them they tried and discarded the more inventive moves against local validation: pseudo-labeling the test images, transductive self-training, mass-scaled crop augmentation. Post-hoc recalibration fell the same way.

What separated them was the late budget. The strongest run noticed its blend's gains were really coming from models retrained on the full training set, validation folds included. It spent its last hour sweeping variants of that one change. The weakest run was equally disciplined: it even measured that seed-to-seed noise outweighed its configuration differences. But it started from a weaker baseline and never closed the gap. It spent the same hour reshuffling ensemble membership for movements smaller than the noise it had just measured.