This repository contains the code and data for the upcoming ECCV26 paper, Counting Trees from Satellite Imagery with Noisy Supervision:

• TinyTrees — a multi-sensor tree counting benchmark with point-level annotations across three geographic regions and satellite sensors.
• TreeMatch — a training method for tree density estimation that leverages optimal transport to learn from both strong and weak annotations.

TinyTrees Dataset

TinyTrees provides georeferenced satellite imagery with per-tree point annotations across three regions, sensors, and resolutions:

Region	Sensor	GSD	Train (strong)	Train (weak)	Test	Total trees	Total area
Rwanda	PlanetScope	3.0 m	231 tiles / 309k trees	73 tiles / 3.4M trees	645 tiles / 237k trees	3.9M	283 km²
China	Gaofen-2	0.8 m	446 tiles / 55k trees	16,364 tiles / 7.7M trees	2,189 tiles / 70k trees	7.8M	2,344 km²
France	SPOT-6	1.5 m	492 tiles / 11k trees	CHM-derived (via Open-Canopy)	493 tiles / 11k trees	22k	0.7 km²

Each tile is a 5-band GeoTIFF (4 spectral bands + 1 binary validity mask). Point annotations are stored in a single GeoPackage per split with a tile column linking each point to its image.

SPOT weak labels are CHM-derived pseudolabels bundled in spot/train_weak/. The corresponding SPOT-6 imagery is not redistributed; to use it, download the Open-Canopy dataset and point spot_imagery_root at its canopy_height/ directory:

canopy_height/
├── 2021/spot/compressed_pansharpened_*.tif
├── 2022/spot/compressed_pansharpened_*.tif
└── 2023/spot/compressed_pansharpened_*.tif

Add to conf/local/local.yaml:

spot_imagery_root: /path/to/open-canopy/datasets/canopy_height

Download

The dataset is hosted on HuggingFace and mirrored at https://sid.erda.dk/cgi-sid/ls.py?share_id=ET2ix678WL. PlanetScope and Gaofen imagery are under the CC BY-NC 4.0 license: usage is reserved for research and education only.

tinytrees/
├── ps/              # PlanetScope (Rwanda)
│   ├── train_strong/   # *.tif + points.gpkg
│   ├── train_weak/
│   └── test/
├── gf/              # Gaofen-2 (China)
│   ├── train_strong/
│   ├── train_weak/
│   └── test/
└── spot/            # SPOT-6 (France)
    ├── train_strong/
    ├── train_weak/  # pseudolabels only (imagery from Open-Canopy)
    └── test/

Loading data

from data.ps import PlanetScopeStrong
from data.gf import GaofenStrong
from data.spot import SPOTStrong

# Each dataset returns (image, valid_mask, count_map)
# image: (C+1, H, W) float tensor — C normalized bands + 1 validity channel
# valid_mask: (1, H, W) binary tensor
# count_map: (1, H, W) sparse 0/1 tensor with tree locations

ds = PlanetScopeStrong(imsize=64, split="train_strong", root="/path/to/tinytrees/ps")
ds = GaofenStrong(imsize=64, split="train_strong", root="/path/to/tinytrees/gf")
ds = SPOTStrong(imsize=64, split="train_strong", root="/path/to/tinytrees/spot")

TreeMatch

TreeMatch is an optimal-transport-based training method for tree density estimation that supports mixed supervision from strong (expert) and weak (e.g. CHM-derived) point annotations. It uses unbalanced optimal transport to match predicted density maps to point annotations, with a slack mechanism to down-weight weak labels.

Pretrained models

Pretrained TreeMatch checkpoints for all three sensors are available on HuggingFace.

from hub_model import UNetR50
import torch

model = UNetR50.from_pretrained("dgominski/TinyTrees", subfolder="ps")   # Rwanda / PlanetScope
# model = UNetR50.from_pretrained("dgominski/TinyTrees", subfolder="gf")   # China / Gaofen-2
# model = UNetR50.from_pretrained("dgominski/TinyTrees", subfolder="spot") # France / SPOT-6
model.eval()

# Input: (B, 5, H, W) — RGBI + 1 binary validity mask (last channel)
x = torch.randn(1, 4, 128, 128)
valid = torch.ones(1, 1, 128, 128)            # 1 = valid pixel, 0 = nodata
density = model(torch.cat([x, valid], dim=1)) # (B, 1, H, W), trees/pixel

Training

Training is configured via Hydra. The main entry point is train.py:

# Train TreeMatch on PlanetScope with strong labels only
python train.py dataset=ps model=treematch train.strong_ratio=1.0 model.lr=8e-05

# Train with 80% strong + 20% weak labels
python train.py dataset=ps model=treematch train.strong_ratio=0.8 model.lr=8e-05

# Train density regression baseline
python train.py dataset=ps model=density_regressor

# Train DM-Count baseline
python train.py dataset=ps model=dm_count

Override dataset paths for your machine in conf/local/local.yaml:

# @package _global_
dataset:
  root: ${dataset_roots.${hydra:runtime.choices.dataset}}
dataset_weak:
  root: ${dataset_roots.${hydra:runtime.choices.dataset_weak}}
dataset_roots:
  ps: /path/to/tinytrees/ps
  gf: /path/to/tinytrees/gf
  spot: /path/to/tinytrees/spot
  ps_weak: /path/to/tinytrees/ps/train_weak
  gf_weak: /path/to/tinytrees/gf/train_weak
  spot_weak: /path/to/tinytrees/spot/train_weak

# SPOT weak labels use imagery from Open-Canopy (not bundled in TinyTrees)
spot_imagery_root: /path/to/open-canopy/canopy_height

Available models

Model	Config	Description
TreeMatch	model=treematch	Unbalanced OT density matching with slack for weak labels
Density Regression	model=density_regressor	Gaussian density map regression (MSE loss)
DM-Count	model=dm_count	Balanced OT with total variation loss
CenterNet	model=centernet	Keypoint detection with heatmap regression
P2PNet	model=p2p	Point-to-point matching

Available backbones

Backbone	Config	Architecture
ResNet-50 U-Net	backbone=resnet50	segmentation_models_pytorch U-Net with ResNet-50 encoder
Swin Transformer	backbone=swint	SwinV2-Small with FPN decoder
ViT	backbone=vit	Vision Transformer with FPN decoder

Rwanda-Tanzania application

We trained a PlanetScope+Sentinel-1+Sentinel-2 (20-band composite) model to count trees in Rwanda and Tanzania.

Reference

@inproceedings{gominski2026counting,
  title     = {Counting Trees from Satellite Imagery with Noisy Supervision},
  author    = {Gominski, Dimitri and Mugabowindekwe, Maurice and Xu, Qiue and Tong, Xiaowei and Brandt, Martin and Le, Hieu and Fensholt, Rasmus and Samaras, Dimitris and Landrieu, Loic},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}

,