adaniele/Agriculture-DelAny-v2-Sentinel

DelAny v2 — Toscana Sentinel-2 head v1

This repository packages the Delineate Anything v2 checkpoint adapted for the Tuscany Sentinel-2 boundary-delineation pipeline.

What was trained

Starting from the official Delineate Anything v2 checkpoint, only the YOLO segmentation head (top-level layer 23) was trained: 8,323,187 parameters out of 62,051,411. Backbone layers 0--22 remained frozen.

Training used the geographic Toscana EuroCrops training split. Validation was used to monitor/select the checkpoint; the held-out test split was never used for training. The model has one field class (the internal Ultralytics class name is item; downstream applications should expose it as field).

Required preprocessing

This is not a general raw Sentinel-2 multi-band checkpoint. Build the same input used at training/inference:

1. Select Sentinel-2 2021 acquisitions from June through October.
2. Build a cloud-reduced temporal median RGB composite from B04/B03/B02.
3. Apply robust 1st--99th percentile normalization to uint8.
4. Pass the result to Ultralytics in BGR channel order.

from ultralytics import YOLO

model = YOLO("model.pt")
result = model.predict(s2_seasonal_bgr_uint8, imgsz=512, conf=0.15,
                       retina_masks=True, device=0)[0]

Starting Tuscany settings: imgsz=512, conf=0.15, then remove polygons smaller than 2,500 m² in a metric CRS.

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: 3490d4afef6cf255bcd5bccb261b394e0526d18b92a55be13ba3e28ef612f07f

Scope and license

The model delineates field boundaries only; it does not classify vineyards, olives, or other crop types. It derives from Delineate Anything v2 and carries the upstream AGPL-3.0 license and attribution requirements.