Geospatial environmental intelligence from Sentinel-2 satellite imagery.
NatureVision is a reproducible geospatial AI platform for analyzing vegetation indices, land-cover patterns, and environmental change from real Sentinel-2 observations.
The project combines a production-oriented analysis pipeline with a controlled research framework investigating how remote-sensing machine-learning models generalize across geographic and temporal conditions.
The central research question is:
How do spatial and temporal validation strategies affect the measured generalization performance of lightweight machine-learning models for Sentinel-2 land-cover classification?
The research framework evaluates:
- Random-pixel vs spatially separated validation
- Lightweight model architectures
- Spectral feature ablation
- Temporal transfer between observation periods
- Prediction confidence and calibration
- Feature importance
- Class imbalance
- Cloud/data-quality robustness
- NDVI change-threshold sensitivity
- Evidence-grounded environmental reporting
All experiments are reproducible and record their configuration, random seeds, dataset provenance, runtime, Git commit, and dataset digest.
The experiments use real Sentinel-2 Level-2A observations and ESA WorldCover 2021 v200 reference labels.
The current research cache contains 313,580 pixels across two observation periods.
Three source groups that could not be retrieved are recorded as missing rather than being replaced with synthetic data.
A central experiment compares random pixel-level validation with geographically separated spatial holdout validation.
The current results show a substantial difference:
| Protocol | Accuracy | Macro-F1 |
|---|---|---|
| Random pixel | ~0.84 | ~0.84 |
| Spatial holdout | ~0.53 | ~0.51 |
This gap is the primary result currently being investigated. It demonstrates why evaluation methodology is important when measuring generalization of remote-sensing models.
The research does not assume that any hypothesis is correct; results are generated directly from the experiments.
The framework compares:
- Random Forest
- Histogram Gradient Boosting
- Linear SVM
- MLP
and evaluates feature configurations ranging from raw spectral bands to derived vegetation and spectral indices.
The current feature-ablation results show only small differences between feature configurations, with macro-F1 remaining close to the baseline across the tested feature sets.
Temporal experiments additionally evaluate transfer between the 2021 and 2024 observation periods.
The production pipeline performs:
Geographic region
↓
Sentinel-2 observations
↓
Cloud / quality masking
↓
Spectral processing
↓
NDVI and environmental indices
↓
Temporal change detection
↓
Land-cover classification
↓
Evidence package
↓
Grounded environmental interpretation