Change Detection and Description from Remotely Sensed Data
Opis
This service moves beyond traditional binary change detection to provide semantic change understanding from multi-modal remotely sensed data. It identifies not only where landscape change has occurred, but also classifies the type of change, estimates its magnitude, and generates human-readable descriptions with attribution of potential drivers. The service integrates explainable AI methods that provide spectral, temporal, and concept-based explanations for detected changes, ensuring transparency and interpretability for environmental policy and land-use planning applications. The service employs foundation-model-based change detection architectures operating on learned representations from fine-tuned EO models, supporting both bi-temporal and multi-temporal analysis. Multi-modal fusion combines optical data (spectral changes), SAR data (structural changes), and lidar data (volumetric changes). Explainable AI methods include spectral attribution maps (adapted Grad-CAM, Integrated Gradients), temporal attribution via transformer attention patterns, perturbation-based importance (SHAP over spectral bands and time steps), and domain-relevant concept probes. Vision-language model fine-tuning (following the ChangeCLIP paradigm) generates context-aware change narratives from bi-temporal image—text pairs.
Pričakovani rezultati
Users receive classified change maps with semantic labels (change type, magnitude), human-readable change descriptions with driver attribution, and explainability outputs (spectral/temporal attribution maps, concept-level explanations). The service supports end-to-end workflows from data cube access through model-based change detection to interpretable, policy-relevant change analysis.
Kartica storitve
| Kategorija | AI development and deployment |
| Domena | Earth Observation, Environmental Monitoring |
| Vertikala | Green Transition |
| Ciljni uporabniki | Environmental Monitoring Agencies, Spatial Planning Authorities, Forestry Services, Disaster Management Bodies, Infrastructure Operators, Research Institutions |
| Tip storitve | Model, Pipeline |
| Rezultat | Semantic change maps; change description reports; explainability visualisations (attribution maps, concept explanations); open-source XAI toolkit for EO foundation models; reusable XAI evaluation framework and protocols. |
| Zahteve | NVIDIA GPU node for model inference (available via SLAIF HPC infrastructure). Python environment. Input data can be sourced directly from the Slovenian EO Data Cube service or provided by the user. |
| Omejitve | Semantic change descriptions are initially developed for Slovenian landscape change types (deforestation, urbanisation, flood damage, bark beetle infestation, agricultural conversion, mining subsidence). Extension to additional change categories requires labelled training examples. Lidar-based volumetric change detection depends on availability of multi-temporal lidar campaigns. |
| Na voljo od | Q2/2027 |
| Licenca | Open (Apache 2.0 for code and models; CC BY 4.0 for datasets) |
| Slika | |
| Oznaka slike |