This thesis aims to evaluate three state-of-the-art foundation models for the task of semantic segmentation,
specifically targeting the segmentation of glaciers calving fronts in SAR1 imagery. Foundation
models are recognized as general-purpose, task-agnostic models that are pre-trained on extensive
datasets, allowing them to be adapted for specific tasks with minimal additional training[1][2][3]. This
research will explore the efficacy of these models when applied to SAR data, which presents unique
challenges due to its complex imaging characteristics. The models selected for this analysis are based
on their performance metrics, methodologies, and the datasets used. To assess the suitability of learned
features for our CaFFe2[4] dataset, the models will be compared quantitatively and qualitatively with
each other and shall be implemented in pytorch. This involves fine-tuning the decoders for calving
front delineation tasks versus only fine-tuning the classifier head within backend frozen features.
• Foundation model 1, DINOv2 [1]: DINOv2, developed by Meta AI, represents a significant
advancement in self-supervised learning for computer vision applications. This model employs
a transformer-based architecture and utilizes a teacher-student training paradigm to facilitate
learning general-purpose visual features without needing labeled data. A critical aspect of
DINOv2 is its emphasis on scaling both the model and the dataset. Unlike previous foundation
models, DINOv2 maintains strict control over data quality and diversity, essential for producing
effective visual representations. For evaluation purposes, we focus on the CaFFe dataset and
assess at least one reported model trained on the ADE20K[5] and Pascal VOC 2012[6] datasets.
• Foundation model 2, Prithvi [2]: Prithvi, developed by IBM and NASA, represents a pioneering
foundational model specifically tailored for geospatial data. This model has been tested
across a variety of Earth observation tasks. It uses Masked Autoencoder technique with a Vision
Transformer architecture. Prithvi leverages multispectral satellite imagery from the Harmonized
Landsat Sentinel-2 (HLS) dataset, which offers high-resolution data suitable for diverse ecological
analyses. The model incorporates statistical factors such as precipitation and temperature,
minimizing bias towards specific landscapes and reducing redundancy across different regions
and time periods. For evaluation, this study will utilize the CaFFe dataset and assess at least one
of the three pre-trained models focused on flood mapping[7], wildfire scar mapping[8], and crop
segmentation[9].
• Foundation model 3, SMLFR [3]: SMLFR3 model is a generative convolutional neural network
designed for analyzing remote sensing data. Like Prithvi, SMLFR uses Masked AutoEncoder
technique, but it is built on a convolutional architecture called ConvNeXt[10], which is an
updated version of traditional ConvNets inspired by transformers and competes well with
transformers regarding accuracy and scalability. In addition, it improves feature representation
during training by applying high-frequency filtering to images. The SMLFR model is trained on
a geographical dataset collected from various sensors, including Sentinel-2, Gaofen, Landsat,
and QuickBird, and contains images from different continents and environments. This study will
evaluate the model on the CaFFe dataset using at least one of the two pre-trained models trained
on the Potsdam2[11] and LoveDA[12] datasets.
References
[1] Maxime Oquab et al. “DINOv2: Learning Robust Visual Features without Supervision”. In:
(2024). arXiv: 2304.07193 [cs.CV]. URL: https://arxiv.org/abs/2304.
07193.
[2] Johannes Jakubik et al. “Foundation Models for Generalist Geospatial Artificial Intelligence”.
In: (2023). arXiv: 2310.18660 [cs.CV]. URL: https://arxiv.org/abs/2310.
18660.
[3] Zhe Dong, Yanfeng Gu, and Tianzhu Liu. “Generative ConvNet Foundation Model With Sparse
Modeling and Low-Frequency Reconstruction for Remote Sensing Image Interpretation”. In:
IEEE Transactions on Geoscience and Remote Sensing 62 (2024), pp. 1–16. DOI: 10.1109/
TGRS.2023.3348479.
[4] N. Gourmelon et al. “Calving fronts and where to find them: a benchmark dataset and methodology
for automatic glacier calving front extraction from synthetic aperture radar imagery”. In:
Earth System Science Data 14.9 (2022), pp. 4287–4313. DOI: 10.5194/essd-14-4287-
2022. URL: https://essd.copernicus.org/articles/14/4287/2022/.
[5] Bolei Zhou et al. “Scene Parsing through ADE20K Dataset”. In: (2017), pp. 5122–5130. DOI:
10.1109/CVPR.2017.544.
[6] Mark Everingham et al. “The pascal visual object classes (VOC) challenge”. en. In: Int. J.
Comput. Vis. 88.2 (June 2010), pp. 303–338.
[7] Derrick Bonafilia et al. “Sen1Floods11: a georeferenced dataset to train and test deep learning
flood algorithms for Sentinel-1”. In: 2020 IEEE/CVF Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW). 2020, pp. 835–845. DOI: 10.1109/CVPRW50498.
2020.00113.
[8] NASA IBM.Wildfire Scar Mapping Dataset. URL: https://huggingface.co/datasets/
ibm-nasa-geospatial/hls%20burn%20scars.
[9] NASA IBM. Multi-Temporal Crop Segmentation. URL: https://huggingface.co/
datasets/ibm-nasa-geospatial/multi-temporal-crop-classification.
[10] Zhuang Liu et al. A ConvNet for the 2020s. 2022. arXiv: 2201.03545 [cs.CV]. URL:
https://arxiv.org/abs/2201.03545.
[11] BSF Swissphoto. 2D Semantic Labeling Contest – Potsdam. URL: http://web.archive.
org / web / 20080207010024 / http : / / www . 808multimedia . com / winnt /
kernel.htm.
[12] Junjue Wang et al. LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive
Semantic Segmentation. 2022. arXiv: 2110.08733 [cs.CV]. URL: https://arxiv.
org/abs/2110.08733.