Temporal, Multi-Modal Foundation Models from Remotely Sensed Information
Overview
This service provides access to state-of-the-art Earth Observation foundation models that have been systematically benchmarked and adapted to Slovenian environmental conditions through parameter-efficient fine-tuning. Users can leverage pre-trained models for downstream tasks such as land cover classification, crop type mapping, forest composition estimation, and urban area delineation without the need for large labelled training datasets. The service bridges the gap between globally pre-trained foundation models and local environmental monitoring needs, enabling accurate predictions even in data-scarce settings. The service benchmarks and adapts leading EO foundation models including TerraMind, Prithvi-EO-2.0, Clay, DOFA, and CROMA using data from the Slovenian EO Data Cube. Parameter-efficient fine-tuning strategies (LoRA, adapters, visual prompt tuning) are applied to adapt globally pre-trained models to local conditions while minimising computational cost. Multi-modal fusion architectures (early, mid-level, and late fusion) combine optical, SAR, and lidar modalities. Temporal encoding components capture phenological cycles and seasonal patterns from dense Sentinel time series.
Outcome
Users receive fine-tuned foundation models optimised for Slovenian landscapes, along with curated benchmark datasets with standardised training/validation/test splits published with DOIs. The service supports zero-shot, few-shot, and full fine-tuning workflows, enabling users to adapt models to their specific tasks with minimal labelled data.
Service Card
| Category | AI development and deployment |
| Domain | Earth Observation, Foundation Models, Environmental Monitoring |
| Vertical | Green Transition |
| Target users | Environmental Researchers, Geospatial Analytics Companies, Precision Agriculture Providers, Urban Planning Agencies, SMEs Developing EO-Based Products |
| Service type | Model, Pipeline |
| Outputs | Fine-tuned EO foundation model weights; benchmark datasets; adaptation scripts and pipelines; performance evaluation reports. |
| Requirements | NVIDIA GPU node for model fine-tuning and inference (available via SLAIF HPC infrastructure). Python environment with PyTorch and Hugging Face libraries. |
| Limitations | Initial model adaptation focuses on Slovenian landscapes (Alpine, karst, Pannonian, Mediterranean); transferability to other regions requires additional fine-tuning. Performance depends on availability of labelled data for downstream tasks. |
| Availability | Q1/2027 |
| Licensing | Open (model weights: Apache 2.0; benchmark datasets: CC BY 4.0) |
| Visual content | |
| Visual caption |