Master’s Thesis – Foundation Models for Remote Sensing (Hyperspectral Focus)

Overview

Foundation models have triggered a paradigm shift in computer vision, achieving strong generalization capabilities with little or no finetuning. These advances are largely driven by self-supervised learning algorithms, which enable the learning of useful representations on unlabeled data. These methods are now being applied to domain-specific problems in medical imaging, genomics, and remote sensing.

Remote sensing faces unique challenges, such as the diversity of sensors with varying spatial, spectral, and temporal characteristics. Common types of sensors include multispectral and hyperspectral sensors, synthetic aperture radar (SAR), and LiDAR, each providing distinct features.

Several efforts have focused on training foundation models for single sensor imagery and multi-sensor data. However, more efforts are needed to integrate more sensors and train truly general-purpose foundation models. For instance, the hyperspectral modality is often neglected due to the high cost of data acquisition, and large-scale pre-training of hyperspectral foundation models is still under-explored. Furthermore, the community lacks a systematic approach to evaluate these models, and the development of consistent and fair benchmarking practices is needed.

Thesis Objective

This thesis will investigate strategies for pre-training remote sensing foundation models, with a focus on the hyperspectral modality, and evaluation protocols to assess their performance under various scenarios.

Your Contribution

We seek a motivated Master's student to explore self-supervised learning techniques for remote sensing foundation models. The student will:

• Work on the integration of hyperspectral data into large-scale pre-training frameworks
• Contribute to the development of standardized evaluation methodologies
• Possibly perform comparative analysis across multiple sensor types to understand benefits and limitations

Project Scope

• Conduct a literature review on self-supervised learning and foundation models in remote sensing
• Develop a new pre-training framework for hyperspectral data, incorporating best practices from computer vision and geospatial AI
• Analyze and interpret results to assess the model’s performance, generalization capabilities, and limitations
• Contribute to the establishment of standardized evaluation protocols for remote sensing foundation models

✅ Your Qualifications

• Currently enrolled in a Master’s program in Geoscience, Earth Science, Remote Sensing, Computer Science, or a related field
• Strong programming skills in Python, with experience in deep learning frameworks (e.g., PyTorch, TensorFlow)
• Interdisciplinary expertise in at least two of the following areas: Geoscience, Remote Sensing, Machine Learning, Hyperspectral Imaging, Earth Observation
• Familiarity with self-supervised learning and foundation models is a plus
• Experience with geospatial software (e.g., QGIS, Google Earth Engine) is advantageous
• Strong motivation to work independently and contribute to a collaborative research environment

We Offer

• An opportunity to engage in cutting-edge research on remote sensing foundation models
• The possibility to publish research findings in a scientific journal or present at a conference (subject to approval)
• Access to computational resources and datasets for large-scale model training and evaluation

Impact

This thesis is an exciting opportunity to contribute to the development of next-generation AI-driven remote sensing technologies, with potential applications in environmental monitoring, climate resilience, and Earth observation analytics.