adaniele/Agriculture-DelAny-v2-Plane

DelAny v2 — Aerial / VHR

This repository packages the exact Delineate Anything v2 checkpoint used for the Tuscany GEOscopio 20 cm orthophoto experiments.

Important provenance

This is not a new aerial fine-tune. It is an unmodified copy of the official MykolaL/DelineateAnything DelineateAnythingv2.pt checkpoint, repackaged with a reproducible deployment configuration for this project. Credit, citations and AGPL-3.0 obligations of the upstream project apply.

Use

• Task: agricultural field instance segmentation / boundary delineation.
• Input: 3-channel optical image, supplied to Ultralytics as uint8 BGR.
• Output: one mask and confidence for every predicted field; convert masks to polygons using the source GeoTIFF affine transform and CRS.
• It does not predict crop type.

from ultralytics import YOLO

model = YOLO("model.pt")
result = model.predict(image_bgr_uint8, imgsz=1024, conf=0.15,
                       retina_masks=True, device=0)[0]

Starting settings used in the Tuscany VHR test: imgsz=1024, conf=0.15, then remove polygons smaller than 50 m² in a metric CRS. These are starting values, not universal defaults.

Files

• model.pt: 119 MB Ultralytics segmentation checkpoint.
• weights_manifest.json: SHA-256 and exact intended configuration.
• INTEGRATION.md: backend/API and georeferencing guidance.

SHA-256 of model.pt: 46700b8a279b07922953a11adaeb5e658d9a2384b6334c8e0a3090886218915a

Limitations

The model produces candidate physical fields, not cadastral truth. Tile large rasters, deduplicate overlap outputs, validate topology, and calibrate confidence/area thresholds locally before operational use.