Urban Tree Species Classification using Sentinel-2 and PlanetScope

Overview

This repository contains the code and datasets used for the research paper "Urban trees species classification using Sentinel-2 and Planetscope satellite image time series", presented at JURSE 2025 in Tunis, Tunisia (May 4-7, 2025). The study focuses on classifying the 20 most representative urban tree species in Strasbourg using deep learning models applied to Satellite Image Time Series (SITS) from Sentinel-2 and PlanetScope.

Key Features

• Utilizes InceptionTime, H-InceptionTime, and LITE models for multivariate Time Series Classification (TSC).
• Implements sensor fusion to combine spectral and temporal features from Sentinel-2 and PlanetScope.
• Achieves 69% accuracy using the best-performing model (H-InceptionTime with sensor fusion).
• Codebase written in PyTorch.

Repository Structure

📂 Urban-Tree-Classification
├── main.py          # Main script to run classification
├── models.py        # Definition of deep learning models (InceptionTime, H-InceptionTime, LITE)
├── params.py        # Configuration and hyperparameters
├── test.py          # Evaluation and performance analysis
├── train.py         # Training pipeline
├── utils.py         # Utility functions for data preprocessing and visualization

Dataset

• Sentinel-2: 10 spectral bands, 10-20m resolution, 22 cloud-free images (2022).
• PlanetScope: 4-band imagery, 3.125m resolution, 53 cloud-free images (2022).
• Data preprocessing includes zonal statistics, buffer-based feature extraction, and co-registration.

Results

Model	Sentinel-2	PlanetScope	Fusion (S2 + PS)
InceptionTime	62.9% ± 0.4%	62.2% ± 1.3%	67.7% ± 1.4%
H-InceptionTime	63.0% ± 0.3%	62.4% ± 1.0%	69.1% ± 0.3%
LITE	45.7% ± 2.4%	48.7% ± 2.5%	56.1% ± 2.5%

Citation

If you use this code, please cite:

@inproceedings{Latil2025JURSE,
  author    = {Marie Latil, Romain Wenger, David Michéa, Germain Forestier, Anne Puissant},
  title     = {Urban trees species classification using Sentinel-2 and Planetscope satellite image time series},
  booktitle = {JURSE 2025},
  year      = {2025},
  address   = {Tunis, Tunisia}
}

Acknowledgments

• Thanks to Open Data Strasbourg for the urban tree inventory.
• Computational resources provided by Mesocentre Unistra.
• Funded by CNES-TOSCA project AIM-CEE and ANR project M2-BDA (ANR-24-CE23-1130).

License

This project is licensed under the MIT License.