Causal Deep Learning for Cross-Country Generalization in Multimodal Remote Sensing

This project aims to develop an automated and robust satellite-based pipeline to map hedgerows, which are import for carbon storage, biodiversity and agricultural landscapes. Even though a large amount of satellite data is available, current methods often rely on manual labeling and are sensitive to changes in region or sensors. During this internship, a machine learning model will be trained on annotated data in France and evaluated on manually labeled data in Canada to study cross-country transfer. Causal machine learning methods will be added to the model to reduce sensitivity to spurious correlations and improve generalization. The expected final result is a semi-automated scalable and reliable pipeline for detecting hedgerows across region while reducing labeling effort using foundation models.

Faculty Supervisor:

Mathias Lécuyer

Student:

Partner:

CentraleSupélec

Discipline:

Computer science

Sector:

Education

University:

The University of British Columbia

Program: