Fire blight in apple orchards is detected by human scouts walking between rows, looking at dormant trees for the characteristic discoloration of infected branches. The disease spreads fast — a late detection costs an entire block. But scouting is slow, expensive, and limited by the observer's visual field angle from the ground.
The autonomous system in this work does something humans can't: it decides where to look based on what it's uncertain about. A semantic 3D map is built in real time from flash-illuminated stereo RGB cameras, with instance-level segmentation labeling each visible branch as healthy, symptomatic, or uncertain. The robot's viewpoint planner then prioritizes moving to positions that observe low-confidence symptomatic regions, gathering additional observations exactly where the classification is least certain.
The semantic approach achieves an F1 score of 0.91 in laboratory trials, outperforming geometric baselines that plan viewpoints for coverage rather than for uncertainty reduction. The difference matters: a geometric planner sees more of the tree but spends time on branches that were already classified. The semantic planner sees less total surface area but resolves more ambiguity per observation.
The practical distinction is spatial resolution. Manual scouting operates at the orchard-block level — is this block infected? The robot operates at the individual-tree level — which specific branch, at what coordinates? The resolution enables targeted intervention rather than block-wide treatment.
Looking where you're uncertain is more efficient than looking everywhere, even if it means you see less.