Supervisor:
Svetlana Illarionova, Head of Research Group;
Description:
AI algorithms are actively used to enhance the characteristics of spatial data, particularly satellite monitoring data. Key characteristics of satellite data include spatial resolution, spectral resolution, revisit frequency, and the presence of imaging artifacts (e.g., clouds, cloud shadows). Generative AI algorithms are employed for image super-resolution to improve the level of detail in satellite imagery. However, the synthesis of new data using AI algorithms may introduce distortions and model hallucinations — for instance, objects that were not present in the original images may appear. To develop robust spatial monitoring algorithms, it is necessary to detect such distortions and assess the reliability of the generated data.
Key tasks:
- Analysis of existing generative super-resolution methods for satellite imagery.
- Development and implementation of metrics for the quantitative evaluation of hallucinations (artifacts, false objects, spectral distortions).
- Conducting experiments on real multispectral data (Sentinel-2, Landsat) with controlled perturbations (clouds, shadows).
-Comparative analysis of the behavior of several generative models (GANs, diffusion models) to identify scenarios most prone to hallucinations.
Skills the intern will acquire:
- Proficiency in working with geospatial data and satellite imagery (rasterio, GDAL, xarray).
- Practical application of generative AI methods (training/fine-tuning super-resolution models in PyTorch/TensorFlow).
- Ability to design and implement fidelity metrics for generated data (including no-reference metrics).
- Skills in critical analysis of AI algorithm robustness to distortions.
- Experience in preparing scientific results and publishing in leading peer-reviewed journals and conferences.
Internship duration: 2 months
Internship start date: flexible
Requirements for the candidate:
Required:
- Python — strong proficiency
- PyTorch — experience writing custom modules (e.g., DataLoaders)
-Basic understanding of Generative AI (how GANs or basic diffusion approaches work)
- Fundamentals of digital image processing (OpenCV, filters, transformations)
Desirable
- Familiarity with modern libraries (diffusers, timm, albumentations)
- Experience working with satellite data
Advantageous
- Participation in competitions (Kaggle)
Monthly compensation: not available
Contact person: Illarionova Svetlana, s.illarionova@skoltech.ru
