Preprocessing strategies for plant disease instance segmentation: annotation efficiency, context signals, and...

Preprocessing choices can strongly affect how well segmentation models localize disease lesions in field imagery, but no systematic instance segmentation study has yet been reported on PlantSeg, the largest publicly available in-the-wild plant disease segmentation dataset (11,458 images; 115 disease categories; 34 host plant species). We evaluated six preprocessing strategies, ranging from annotation-free baselines to expert polygon-guided approaches, each trained for approximately 11.5 h on a single NVIDIA A100 GPU, and assessed them with pixel-level mean intersection over union (mIoU) and instance-level bounding-box and mask average precision (AP). Annotation-guided cropping gave the strongest results: bounding box-guided preprocessing reached 34.12% mIoU and 13.34% mask mAP50, compared with 27.93% and 10.07% for the unprocessed baseline. Context-aware mask augmentation consistently degraded performance, and pixel-level and instance-level metrics ranked the strategies differently. Because the annotation-guided conditions derive their crop windows from ground-truth lesion locations on every split, these gains represent an oracle upper bound on what disease-aware spatial guidance can contribute, not an end-to-end deployable pipeline. The degradation under context-aware augmentation supports the hypothesis that the color and texture contrast at the healthy-to-necrotic boundary carries diagnostic signal the model actively exploits. The divergent metric rankings arise because semantic overlap and instance matching capture complementary properties, which argues for dual-metric reporting in plant disease segmentation. To our knowledge, this is the first systematic instance segmentation study on PlantSeg, and all code and per-instance predictions are publicly released.