cp500/ftw-field-centroids-30m

Fields of the World, repacked at 30 m — centroids, masks, multispectral

A beginner-friendly repack of four country subsets of Fields of The World (FTW), downsampled to the 30 m/px scale of Landsat, for training field-centroid U-Nets and point-promptable segmentation (PicoSAM / SAM-family) pipelines.

6,379 chips · ~121,000 labelled field instances · 4 countries (Belgium, Denmark, Luxembourg, Rwanda — chosen for permissive source licenses).

What is in each .npz (one per country)

array	shape	dtype	meaning
rgb	(N, 85, 85, 3)	uint8	display RGB from Sentinel-2 window A (B04,B03,B02), 30 m/px
ms	(N, 85, 85, 8)	uint16	full multispectral: windows A and B × (B04 red, B03 green, B02 blue, B08 NIR), raw S2 L2A DN
heatmaps	(N, 85, 85)	float16	Gaussian splats at every field centroid (sigma ~ field size) — the U-Net regression target
centroids	(N, M, 3)	float32	per field: (x, y, radius_px); padded with −1
instance_masks	(N, 85, 85)	uint16	field instance IDs (0 = background) — train instance/boundary models
semantic3	(N, 85, 85)	uint8	0 background, 1 field interior, 2 field boundary
chip_ids	(N,)	str	original FTW chip id, for joining back to the 10 m source

Quick start (newbie path)

import numpy as np
d = np.load("ftw30m_luxembourg.npz")
x = d["rgb"] / 255.0                 # (N,85,85,3) inputs
y = d["heatmaps"].astype("float32")  # (N,85,85) centroid targets
# train any small U-Net to regress y from x; peaks of the prediction are
# field centroids; feed each peak to PicoSAM/SAM as a point prompt.

A complete trainer is included: starter_train.py (--channels rgb | rgb+ndvi | all8), and picosam_prompt_demo.py shows peak → point-prompt → mask.

Use the invisible light

B08 (NIR) is where vegetation actually shouts. Turn the non-visible bands into feature channels instead of discarding them:

ms = d["ms"].astype("float32")
a_red, a_nir = ms[..., 0], ms[..., 3]
ndvi_a = (a_nir - a_red) / (a_nir + a_red + 1e-6)      # vegetation vigor, date A
ndvi_b = (ms[..., 7] - ms[..., 4]) / (ms[..., 7] + ms[..., 4] + 1e-6)
change = ndvi_b - ndvi_a                               # phenology between dates
x = np.concatenate([d["rgb"]/255.0, ndvi_a[..., None], change[..., None]], -1)

--channels all8 in the starter uses every band of both dates.

Provenance & licenses

Derived from the FTW benchmark (Kerner Lab et al.). Source polygon licenses: Denmark CC0-1.0 · Luxembourg CC0-1.0 · Belgium "no restrictions on public access" · Rwanda CC-BY-4.0. Sentinel-2 imagery © ESA Copernicus (free use with attribution). This repack: CC-BY-4.0. Please cite the FTW paper and https://fieldsofthe.world/ . Repacked by Team Picasso for the Places-Not-Pixels project (OpenCV AI Competition 2026).