Benedikt Blumenstiel
Geospatial AI, IBM Research | Based in Zurich
About me
Hello everyone! I'm Benedikt — an open-minded researcher passionate about machine learning, innovation, and sustainability. I work in the AI for Climate Impact Team at the IBM Research Zurich lab. We build multi-modal foundation models for Earth observation and work on climate-related applications using satellite data.
Looking forward to connecting with you!
Selected projects
TerraMind
TerraMind is the first multimodal any-to-any generative foundation model for Earth Observation jointly developed by IBM and ESA. TerraMind uses a dual-scale transformer-based encoder-decoder architecture, simultaneously processing pixel-level and token-level data. The model was pre-trained on 500B tokens from 9M spatio-temporally aligned multimodal samples from the TerraMesh dataset and is available on HuggingFace.
TerraTorch
TerraTorch is an open-source toolkit for fine-tuning geospatial foundation models. Built on PyTorch Lightning and TorchGeo, it provides a modular framework for working with state-of-the-art models like Prithvi, TerraMind, and Clay. With ready-to-use tasks for segmentation, classification, and regression, TerraTorch makes it easy to develop and deploy geospatial AI applications—from research prototypes to enterprise solutions.
Prithvi EO 2.0
Prithvi-EO-2.0 is the second generation EO foundation model jointly developed by IBM and NASA. Prithvi-EO-2.0 is based on the ViT architecture, pretrained using a masked autoencoder (MAE) approach with two modifications: 3D positional embeddings for multi-temporal input and additional location and temporal embeddings to encode metadata. The models are pre-trained on 4.2M time series samples and released on HuggingFace.
MESS Benchmark
Holistic benchmarks allow a robust quantitative comparison of model architectures. However, such evaluation practice covering a wide range of real-world applications is lacking for multi-modal (i.e., text-to-image) zero-shot semantic segmentation tasks. Thus, our motivation to compose the Multi-domain Evaluation of Semantic Segmentation (MESS) benchmark, which consists of 22 datasets for varying application domains. More details can be found on our website.
Résumé
Publications
* Equal contribution
Work Experience
Research Software Engineer, AI for Climate Impact, IBM Research
02/2024 - present
- Built TerraMind, the leading EO foundation model on PANGAEA benchmark, using multimodal token prediction
- Enabled large-scale EO pre-training and improved wildfire and flooding fine-tuning by curating multimodal datasets (TerraMesh, ImpactMesh)
- Developed MS-CLIP, a multispectral vision-language EO model, with 7 pp. accuracy gains compared to RGB-only models using MLLM-generated captions
- Applied EO embeddings from Prithvi to remote sensing image retrieval, outperforming general computer vision models by 6 pp. mAP
- Increased external adoption to 500k+ model downloads by open-sourcing the TerraMind and Prithvi models, integrating them in a easy-to-use manner into the TerraTorch package, and promoting them through talks and webinars
Intern, AI for Climate Impact, IBM Research
08/2023 - 02/2024
- Curated a 4.2M time series dataset for the Prithvi 2.0 pre-training, resulting in +8 pp. downstream task performance compared to the previous version, by using diverse sampling strategy based on land-cover and bioregions
- Developed data pipelines for automated data fetching of missing inputs and faster data loading in the IBM GEO Studio, a low-code GeoAI webapp
Junior Researcher, Applied AI in Services Lab, KIT
05/2020 - 04/2022
- Published insights on human–AI collaboration by studying interactive few-shot learning in computer vision
- Trained few-shot object detection models for first-of-their-kind AEC applications using TensorFlow
Design Thinking project, KIT & HUK Coburg
10/2019 - 06/2020
- Used design thinking to develop user-centered, digital innovations and prototypes for retirement saving
Intern, R&D Strategy, AUDI AG
07/2017 - 02/2018
- Supported the project manager in defining a new technology-focused R&D strategy
- Worked in the project management office of an R&D transformation project
Education
M.Sc. Information Systems, Karlsruhe Institute of Technology
10/2019 - 04/2023
- Focus on machine learning and service innovation
- Thesis on multi-domain adaptation of zero-shot semantic segmentation
Exchange, Norwegian University of Science and Technology
01/2021 - 06/2021
- Courses in computer vision, natural language processing and statistics
B.Sc. Industrial Engineering, Karlsruhe Institute of Technology
10/2015 - 09/2019
- Focus on digital services and business strategy
- Thesis on large-scale topic modeling of scientific papers using NLP