Weed interference is one of the main yield constraints in sugarcane (Saccharum spp.), andsite-specific management depends on identifying infested areas accurately and at low cost.This paper describes an end-to-end pipeline for weed segmentation in sugarcane from un-manned aerial vehicle (UAV) orthomosaics. Annotation is semi-automatic: operator-drawnbounding boxes prompt the Segment Anything Model 2 (SAM 2) to produce binary masks,which are then vectorised and exported as georeferenced GeoJSON. Two encoder–decodersegmentation models, built on VGG16 and MobileNetV2 encoders, were trained on 5,000tiles cropped from those orthomosaics under identical early stopping, checkpointing, and L2regularisation. Evaluated over approximately 6.9 million labelled pixels across four classes— sugarcane canopy (background), narrow-leaf weeds, broadleaf weeds, and castor bean(Ricinus communis) — the VGG16 model reached 83.06% pixel accuracy and 0.692 meanintersection over union (mIoU), against 73.87% and 0.549 for MobileNetV2. Per-class anal-ysis shows that the dominant error is not confusion between weed types, which accountsfor 1.9% of the pixels belonging to those two classes in the VGG16 model, but confusionbetween weed patches and the surrounding sugarcane canopy: broadleaf weeds lose 25.0%of their pixels to background, and 15.9% of background pixels are assigned to a weed class.Because the two models were trained on different colour representations, we report thecomparison as an operational baseline rather than a controlled architectural benchmarkand state explicitly what that confounding does and does not permit.