ftw-q

Agricultural field boundary segmentation from one year of quarterly Sentinel-2 cloudless mosaics. A U-Net with an EfficientNet-B3 encoder, trained on FTW-Q.

Files

File	Description
ftw-q-unet-efb3.pt	PyTorch state dict
ftw-q-unet-efb3.onnx	ONNX (opset 18, fp32), dynamic batch

Input and output

• Input: (batch, 16, 512, 512) float32. Channels are [B04, B03, B02, B08] for each of Q1, Q2, Q3, Q4, in that order (channel 4 * q + b), from the Copernicus Sentinel-2 Global Mosaics (int16, reflectance × 10000).
• Preprocessing: divide the raw values by 3000, then upsample the 10 m pixels 4× with bilinear interpolation to 2.5 m. A 512 × 512 input covers 128 × 128 pixels at 10 m.
• Output: logits, (batch, 3, 512, 512) at 2.5 m. Classes: 0 background, 1 field, 2 field boundary.

Usage

import numpy as np
import onnxruntime as ort
import torch
import torch.nn.functional as F

x = torch.from_numpy(stack.astype(np.float32) / 3000)[None]  # stack: (16, 128, 128) at 10 m
x = F.interpolate(x, scale_factor=4, mode="bilinear", align_corners=False)

sess = ort.InferenceSession("ftw-q-unet-efb3.onnx")
logits = sess.run(None, {"input": x.numpy()})[0]

PyTorch, with segmentation-models-pytorch 0.5.0:

import segmentation_models_pytorch as smp
import torch

model = smp.Unet(encoder_name="efficientnet-b3", encoder_weights=None, in_channels=16, classes=3)
model.load_state_dict(torch.load("ftw-q-unet-efb3.pt", map_location="cpu"))
model.eval()

Training

• Data: the FTW-Q train split; the checkpoint with the best validation polygon F1 is kept.
• Initialization: ImageNet encoder weights.
• Optimization: 35 epochs, batch size 16, AdamW, learning rate 1e-3 with cosine decay, cross-entropy + Dice loss, bf16 mixed precision.
• Augmentation: random 512 × 512 crops; horizontal and vertical flips, rotation within ±90° and sharpness (each p = 0.5); a per-sample normalization divisor drawn from U(1500, 4500) in place of 3000; per-band gamma U(0.8, 1.2) (p = 0.3) and per-band affine a·x + b with a ~ U(0.9, 1.1), b ~ U(−0.1, 0.1) (p = 0.5); with p = 0.4, one quarter is replaced by a copy of another.

Evaluation

FTW-Q test split, 7,817 chips. Field polygons are extracted from the argmax field class, polygons under 1,000 m² are dropped, and each is matched to a label polygon at IoU > 0.5. Scores are averaged over countries, excluding Kenya and Portugal.

Object F1	Precision	Recall
0.526	0.601	0.475

License

CC-BY-4.0.