Multisource Time Series Classification Models for Urban Tree Species

Authors: Victor Thiebaut¹, Ali Ismail-Fawaz¹, Cyril Meyer¹, Romain Wenger², Anne Puissant³ and Germain Forestier^1,4

¹ IRIMAS, Université de Haute-Alsace, France
² CESBIO-CNES/CNRS/INRAE/IRD, University of Toulouse, France
³ LIVE, University of Strasbourg, France
⁴ DSAI, Monash University, Australia

This repository is the source code of the article titled "Multisource Time Series Classification Models for Urban Tree Species", accepted at the 11th AALTD Workshop at ECML/PKDD 2026. In this article, we present multiple architectures that use different classifier to solve multisource time series classification problem.

Abstract

Multisource time series classification is a critical computational challenge spanning diverse applications, from clinical health monitoring and human activity recognition to Earth observation. However, effectively integrating asynchronous, heterogeneous data streams remains architecturally complex due to varying sampling rates, source-specific noise, and distinct spatial-temporal resolutions. This paper presents a comprehensive comparative analysis of fusion strategies for a multisource time series dataset of Satellite Image Time Series (SITS) representing urban tree species. We systematically assess the capabilities of existing models through diverse feature extraction techniques, which we adapt in this work to handle multisource data. To do this, we employ a broad spectrum of deep and non-deep learning methods, specifically LITE, TempCNN, MultiRocket, K-Nearest Neighbors across different distance metrics, Quant, and Catch22. Crucially, the classifiers coupled with these methods are systematically selected and tailored to match the specific nature of the intermediate representations generated by each preceding extraction block. By exploring these various architectural combinations and fusion concepts, this work highlights the specific contributions, strengths, and limitations of each method.

Data

This archive uses the dataset presentented in the article "Urban tree species benchmark dataset for time series classification", accepted in the Data in Brief Journal, the dataset is available here.

To use the code data has to be in the main folder in a /data/ folder. It has to contain /data/train/, /data/test/ and /data/val/. Each folder has to contain /ps/, /s2/and /y/ folders. Moreover, each serie has to be .npy file and for each source and label the .npy filename has to correspond.

Docker

This repository supports the usage of docker. In order to create the docker image using the [dockerfile](This repository supports the usage of docker. In order to create the docker image using the dockerfile, simply run the following command (assuming you have docker installed and nvidia cuda container as well):

docker build --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) -t multisource-tsc-image .

After the image has been successfully built, you can create the docker container using the following command:

docker run --gpus all -it --name multisource-tsc-container -v "$(pwd):/home/myuser/code" --user $(id -u):$(id -g) dmultisource-tsc-image bash

The code will be stored under the directory /home/myuser/code/ inside the docker container. This will allow you to use GPU acceleration. ), simply run the following command (assuming you have docker installed and nvidia cuda container as well):

docker build --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) -t multisource-tsc-image .

After the image has been successfully built, you can create the docker container using the following command:

docker run --gpus all -it --name deep-rehab-pile-container -v "$(pwd):/home/myuser/code" --user $(id -u):$(id -g) multisource-tsc-image bash

The code will be stored under the directory /home/myuser/code/ inside the docker container. This will allow you to use GPU acceleration.

This project supports python>=3.11 only.

Running a model

Check the config file and update with the model you would like to use and with which configuration you want.

Then you can run with python3 the main.py file.

If you want to get your full results with deep learning models you should run either tests_DL.py for Decision Level fusion or tests_FL.py for Feature Level fusion.

Acknowledgments

This work was supported by the ANR M2-BDA project (ANR24-CE23-1130) of the French Agence Nationale de la Recherche. The authors would like to acknowledge the High Performance Computing Center of the University of Strasbourg for supporting this work by providing scientific support and access to computing resources. Part of the computing resources were funded by the Equipex Equip@Meso project (Programme Investissements d’Avenir ) and the CPER Alsacalcul/Big Data.