Left: Soil categories from ASCAT data. Right: Accuracy of baseline predictions for mixed soil areas.
- Detect subsurface scattering in Metop-ASCAT soil moisture retrievals using microwave backscatter, slope, and curvature features.
- Targets arid and semi-arid regions where surface-dry soils produce non-monotonic backscatter responses.
- Binary classification distinguishes surface-dominated (wet edge) vs. subsurface-dominated (dry edge) signals to improve downstream soil moisture accuracy.
- Data curation: Assemble ASCAT 12.5 km backscatter (gpi_ascat), ERA5-Land and GLDAS soil moisture, plus topographic slope and curvature derivatives.
- Feature engineering: Build temporal summaries (rolling means, daily differences) and combine with vegetation/precipitation indicators for experiments.
- Per-pixel modelling: Train a random forest classifier for each ASCAT grid cell; capture feature importances and probability outputs.
- Evaluation: Track accuracy, precision, and recall per pixel; visualise spatial performance and class balance issues.
- Baseline features (backscatter + slope + curvature) yielded the most stable results, often >80% accuracy across the Iberian Peninsula test region.
- Extended temporal features (15/30/60-day rolling means, first/second differences) provided marginal gains while increasing training cost.
- Class imbalance remains a limitation; pixels with sparse wet events require careful interpretation despite high headline accuracy.
notebooks/: Jupyter notebooks covering data ingestion, feature engineering, and model training iterations.docs/report.pdf: Final write-up detailing methodology, experiments, and discussion.docs/proposal.pdf: Original project proposal and research questions.docs/poster.pdf: Poster summarising results and visualisations.