taylor-geospatial/ftp-b3

FTP-PRUE+ (EfficientNet-B3) — Fields of the Planet

3m PlanetScope field-boundary segmentation model from Fields of the Planet (FTP), a PlanetScope companion to Fields of the World v2 (FTW). This is the paper's main headline model.

U-Net decoder over an EfficientNet-B3 encoder, trained on paired early- and peak-season 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/ftw-planet
• Paper: Fields of the Planet: Field Boundary Mapping Beyond 10m
• Architecture: smp.Unet(encoder_name="efficientnet-b3"), 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-b3.ckpt	PyTorch Lightning checkpoint	state_dict + hyper_parameters only — optimizer/scheduler state stripped
ftp-b3.onnx	ONNX (opset 17)	single-file, dynamic batch/H/W, needs onnxruntime (plain PyTorch cannot execute .onnx)
ftp-b3.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-b3").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-b3.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-b3.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-b3.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 + the repo's watershed post-processing (scripts/eval/postprocess_eval.py) recovers instance polygons.

Results

Macro-averaged over the 10 dense held-out FTW-v2 countries, true-polygon scoring at native 3m GSD (see paper Table 1): PQ 35.5, sub-0.5ha PQ 15.7, matched-boundary error 7.4m.

Citation

@misc{ftw-planet,
  author = {Corley, Isaac and Robinson, Caleb and Marcus, Jennifer and Kerner, Hannah},
  title  = {Fields of the Planet: Field Boundary Mapping Beyond 10m},
  year   = {2026},
  url    = {https://github.com/taylor-geospatial/ftw-planet}
}

License

Weights: CC-BY-NC-4.0. Trained on PlanetScope imagery (c) Planet Labs PBC, exported under the NICFI / research program, and FTW v2 polygons (CC-BY-4.0). Refer to those source terms for any redistribution of the underlying data.