Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment

Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment
Kamyar Nasiri, William Guimont-Martin, Damien LaRocque, Gabriel Jeanson, Hugo Bellemare-Vallières, Vincent Grondin, Philippe Bournival, Julie Lessard, Guillaume Drolet, Jean-Daniel Sylvain and Philippe Giguère
Paper: https://www.mdpi.com/1999-4907/16/4/616

This repo contains the source code and the datasets used in our paper Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment, published in the Classification of Forest Tree Species Using Remote Sensing Technologies: Latest Advances and Improvements special issue of the Forests MDPI journal.

Repository organization

The repository is composed of two main directories:

• lowAltitude_classification contains the code for the image classifier $C_{\text{DINOv2}}$
• lowAltitude_segmentation contains the code for the segmentation model $S_{\text{M2F}}$

Installation

To ease the installation of the dependencies and the training of the models, we provide two Dockerfiles, DockerfileClassif and DockerfileSeg, respectively, for the image classifier and the segmentation model. We provide make commands to build the containers.

With docker:

make cls-build # Image classifier
make seg-build # Segmentation model

With podman:

make cls-podbuild # Image classifier
make seg-podbuild # Segmentation model

WilDReF-Q (UAV imagery dataset)

Labeled Dataset

Dataset	Description	Link
Training set	71 patched UAV images and masks (1024x1024)	Download
Validation set	46 patched UAV images and masks (1024x1024)	Download
Test set	36 patched UAV images and masks (1024x1024)	Download

Unlabeled UAV Datasets

Dataset	Content	Description	Link
Original (Non-Patched)	Images	Raw UAV imagery without patching, over 11k images	Download
Patched (for $S_{\text{M2F}}$ Pre-training)	Images	Patched UAV images for segmentation model training, over 143k images	Download
Patched (for $S_{\text{M2F}}$ Pre-training)	Masks	Generated pseudo-labels for patched UAV images by the classifier $C_{\text{DINOv2}}$	Download

Pre-trained Models

Model	Link
Classification	Download
Segmentation (Pre-trained PT)	Download
Segmentation (Finetuned FT)	Download

Citation

If you use the code or data in an academic context, please cite the following work:

@article{Nasiri2025,
  title     = {{Using Citizen Science Data as Pre-Training for Semantic Segmentation of High-Resolution UAV Images for Natural Forests Post-Disturbance Assessment}},
  volume    = {16},
  issn      = {1999-4907},
  url       = {http://dx.doi.org/10.3390/f16040616},
  doi       = {10.3390/f16040616},
  number    = {4},
  journal   = {Forests},
  publisher = {MDPI AG},
  author    = {Nasiri,  Kamyar and Guimont-Martin,  William and LaRocque,  Damien and Jeanson,  Gabriel and Bellemare-Vallières,  Hugo and Grondin,  Vincent and Bournival,  Philippe and Lessard,  Julie and Drolet,  Guillaume and Sylvain,  Jean-Daniel and Giguère,  Philippe},
  year      = {2025},
  month     = mar,
  pages     = {616}
}