taylor-geospatial/ftp-b7

FTP-PRUE+ (EfficientNet-B7) — Fields of the Planet

3m PlanetScope field-boundary segmentation model from Fields of the Planet (FTP), a PlanetScope companion to Fields of the World (FTW). Larger-backbone variant of the paper's main reported model (ftp-b3).

U-Net decoder over an EfficientNet-B7 encoder, trained on paired planting- and harvest-window PlanetScope surface-reflectance imagery (8 input channels: 2 seasonal windows x 4 bands) to predict a 3-class field mask (background / field / boundary) at 3m ground sample distance.

• Repo: taylor-geospatial/fields-of-the-planet
• Paper: Fields of the Planet: Field Boundary Mapping Beyond 10m
• Dataset: taylor-geospatial/ftw-planet
• Architecture: smp.Unet(encoder_name="efficientnet-b7"), in_channels=8, classes=3
• Training resolution: 512x512 crops, 3m/px
• Loss: log-cosh Dice, class weights [0.05, 0.2, 0.75], ignore_index=3
• Recipe: PRUE+ (geometry/noise augmentations, watershed post-processing, D4 TTA at eval)

Files

File	Format	Notes
config.json + model.safetensors	segmentation_models_pytorch Hub format	recommended for most users — safetensors (no pickle), self-describing architecture, needs only pip install segmentation-models-pytorch
ftp-b7.ckpt	PyTorch Lightning checkpoint	state_dict + hyper_parameters only — optimizer/scheduler state stripped
ftp-b7.onnx	ONNX (opset 17)	single-file, dynamic batch/H/W, needs onnxruntime (plain PyTorch cannot execute .onnx)
ftp-b7.pt2	torch.export ExportedProgram	standalone, dynamic batch / static 512x512, loads with torch.export.load — no ftw_planet install needed

Usage

segmentation_models_pytorch (recommended)

import segmentation_models_pytorch as smp

model = smp.from_pretrained("taylor-geospatial/ftp-b7").eval()
logits = model(image)  # image: (B, 8, H, W) float32, 2 seasonal windows x 4 bands

Lightning checkpoint (raw smp.Unet)

The checkpoint's state_dict is just an smp.Unet under a model. prefix. Unless you need the Lightning training wrapper, smp.from_pretrained above is simpler.

import torch
import segmentation_models_pytorch as smp

ckpt = torch.load("ftp-b7.ckpt", map_location="cpu")
hp = ckpt["hyper_parameters"]
model = smp.Unet(encoder_name=hp["backbone"], encoder_weights=None, in_channels=hp["in_channels"], classes=hp["num_classes"])
model.load_state_dict({k.removeprefix("model."): v for k, v in ckpt["state_dict"].items()})
model.eval()

logits = model(image)  # image: (B, 8, H, W) float, 2 seasonal windows x 4 bands

ONNX

import onnxruntime as ort

sess = ort.InferenceSession("ftp-b7.onnx", providers=["CPUExecutionProvider"])
logits = sess.run(None, {"image": image_np})[0]  # image_np: (B, 8, H, W) float32

torch.export (standalone, no ftw_planet needed)

import torch

exported = torch.export.load("ftp-b7.pt2")
model = exported.module()
logits = model(image)  # image: (B, 8, 512, 512) float32, any batch size

Output is 3-class logits (background / field / boundary). Argmax plus the repo's watershed post-processing (scripts/eval/postprocess_eval.py) recovers instance polygons.

Results

Polygon-level results macro-averaged over the ten dense-label held-out countries dominated by smallholder fields (paper Table 1). This checkpoint is the FTP-PRUE+ / EfficientNet-B7 row.

Method	Sensor	Backbone	PQ	SQ	RQ@.5	F1[.5:.95]	|ΔN|/N ↓	Bd. err mean (m) ↓	Bd. err p95 (m) ↓	Pixel IoU †	PQ small ‡	PQ med ‡	PQ large ‡
DelineateAnything *	PlanetScope	YOLO11x	9.5	73.3	12.7	7.0	0.75	13.7	37.8	51.1	1.7	7.1	16.3
DelineateAnything-S *	PlanetScope	YOLO11n	3.5	70.8	4.8	2.5	0.82	13.2	34.2	40.7	0.8	2.8	6.7
DelineateAnything v2 *	PlanetScope	YOLO11x	7.5	75.0	10.0	5.6	0.82	9.4	25.5	27.1	4.4	11.4	14.8
FTW-PRUE+	Sentinel-2	EfficientNet-B3	21.0	71.4	28.9	14.6	0.33	18.6	54.7	61.8	5.8	25.3	33.8
FTW-PRUE+	Sentinel-2	EfficientNet-B7	24.2	71.0	32.8	17.2	0.35	14.4	43.4	63.6	7.5	28.4	37.7
FTP-PRUE+	PlanetScope	EfficientNet-B3	35.5	75.7	46.2	27.1	0.33	7.4	22.8	68.8	15.7	39.2	52.0
FTP-PRUE+ (this model)	PlanetScope	EfficientNet-B7	35.4	74.4	46.1	27.0	0.30	7.4	22.8	74.2	15.6	40.6	50.9

Bold marks the best value per column. * Released models evaluated without training on FTW or FTP, each at its best swept inference resolution and confidence setting. † Pixel IoU is not comparable across sensors due to differences in resolution. ‡ PQ for small (<0.5 ha), medium (0.5-2 ha), and large (>2 ha) ground-truth fields.

The B3 variant scores higher PQ (35.5 vs 35.4) at about 5x fewer parameters, but B7 has better pixel IoU (74.2 vs 68.8) and medium-field PQ (40.6 vs 39.2).

Citation

@misc{corley2026fieldsplanetfieldboundary,
  title         = {Fields of the Planet: Field Boundary Mapping Beyond 10m},
  author        = {Isaac Corley and Caleb Robinson and Jennifer Marcus and Hannah Kerner},
  year          = {2026},
  eprint        = {2607.04449},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.04449}
}

License

Weights are released under CC-BY-NC-4.0 (non-commercial), subject to the licensing terms of the underlying data sources.

Trained on PlanetScope imagery © Planet Labs PBC, obtained directly from the Planet archive under a research license for academic and nonprofit use. Use and redistribution of Planet imagery remain subject to the applicable Planet license terms. See the Planet Licensing Information Center for additional information.

Trained on FTW field-boundary polygons (CC-BY-4.0); see fieldsoftheworld/ftw-baselines for source terms.