Student information
MSc thesis topic: Evaluating the Impact of Deep Foundation models on Forest Inventory
Current deep learning models are based on vision transformers (ViTs) that can ingest data at various spatial, spectral and temporal resolutions. When pre-trained on large datasets, these general-purpose deep learning models can outperform more specific deep architectures like convolutional or recurrent neural networks. Satellite data provides a wealth of data and, in particular, for forestry-related problems. Hence, a series of foundation models like TreeSatAI [1] and benchmarks like FoMo Bench [2] have been proposed.
The task of this thesis will be to familiarize yourself with these foundation models and benchmark and evaluate their effectiveness on multi-label tree classification for national forest inventories in one or multiple European countries.
This will require expertise and affinity to learn new concepts in deep learning and artificial intelligence and interest in forest-related applications. These novel deep learning methods will be compared to more conventional convolutional neural networks and classical random forest classification approaches. Potentially, location-specific information can be integrated through the pre-trained SatCLIP [3] location encoder model.
Relevance to research/projects at GRS or other groups
This thesis is supervised by Marc Rußwurm (WU-GRS) and Carmelo Bonannella (OpenGeoHub). MR will support the topic from a machine learning perspective, while CB provides data and expertise in the Forestry task.
Objectives
- Survey related work and literature with respect to foundation models and, in particular, the forestry application field.
- Construct and adapt a downstream task dataset for forest inventory so that current foundation models can be fine-tuned for forest inventory classification.
- In the optimal case, a new foundation model can be trained through access HPC (SURF/Anunna) and/or the Microsoft Planetary Computer.
Requirements
- required: Deep Learning (This Thesis requires a strong personal interest in Deep Learning)
- optional: Interest in Forestry
Expected reading list before starting the thesis research
- Ahlswede et al. (2022). TreeSatAI Benchmark Archive: A multi-sensor, multi-label dataset for tree species classification in remote sensing. Earth System Science Data Discussions, 2022, 1-22.
- Bountos et al. (2023). FoMo-Bench: a multi-modal, multi-scale and multi-task Forest Monitoring Benchmark for remote sensing foundation models. arXiv preprint arXiv:2312.10114.
- Klemmer, et al. & Rußwurm, M. (2023). SatCLIP: Global, general-purpose location embeddings with satellite imagery. arXiv preprint arXiv:2311.17179.
Theme(s): Modelling & visualisation, Integrated Land Monitoring