---

name: scienceclaw-benchmark-for30 description: "Use when segmenting top-down field or UAV RGB images of crops (sugar beet, PhenoBench-style) into soil / crop / weed semantics plus crop-plant and crop-leaf instances (hierarchical panoptic segmentation), from a few labelled images, delivering per-image integer label maps scored by PQ+. Corresponds to FoR30 of the companion ScienceClaw-Eval benchmark." metadata: { "openclaw": { "emoji": "📊" } }

Hierarchical panoptic segmentation of field images (PhenoBench)

Task. Input: uint8 RGB images (n, H, W, 3), usually plus a few labelled images (semantics, plant and leaf instances, optional visibility maps). Deliverable: a dict of integer arrays (n, H, W) at the input size: semantics (0 soil, 1 crop, 2 weed; labels 3/4 count as 1/2), plant_instances, leaf_instances (ids >= 1, 0 = none).

Quality. PQ+ in percent, higher is better: mean of IoU soil, IoU weed, PQ of crop plants and PQ of crop leaves (heavily occluded instances are filtered). scilib.phenoseg.pq_plus(pred, gt) scores labelled images locally.

Library (check scienceclaw_tools(operation=show|weights)):

• CPU: scilib.phenoseg.fit_predict(train_images, train_semantics, targets): LightGBM pixel classifier plus watershed instances (parts: fit_pixel_classifier, predict_probs, oof_probs, panoptic_from_probs, growth_scale). Variants: phenoseg_deep.fit_predict_deep (DINOv2, assets dinov2_base / dinov2_large); phenoseg_sam.fit_sam_selector + predict_sam_panoptic (SAM 2.1, asset sam2_hf, GPU).
• Pretrained, guard with available(): phenoseg_m2f.predict_panoptic (Mask2Former, asset phenobench_m2f); phenobench_weyler.predict_panoptic (asset phenobench_weyler; crop plants and leaves only, no weeds). phenoseg_hapt.predict_panoptic (asset hapt) was trained on cauliflower: a cross-domain diagnostic only.
• Operators: field_panoptic_segmentation_pixel_lgbm, field_panoptic_segmentation_m2f, field_panoptic_score_pq_plus.

Routes. Baseline: excess-green threshold with connected components. Then pixel classifier, DINOv2 / SAM variants, and pretrained Mask2Former when staged. Compare on held-out labelled images with pq_plus.

Rules.

• Validate with folds blocked by capture or field (oof_probs(..., groups=...)), never by random pixels; never tune on the images being predicted.
• The released PhenoBench checkpoints were trained on PhenoBench itself (Mask2Former: official train partition, 1,407 images): images from it are not held-out evidence. Disclose the overlap.
• Instance sizes depend on growth stage and resolution; take them from labelled data (model.plant_size, growth_scale).
• Keep shapes, integer dtypes and label values exactly as above.