fenicXs/Wheat-Rust-Forecast-Ethiopia

# Wheat Rust Forecast Ethiopia Improving wheat rust disease prediction in Ethiopia using machine learning and explainable AI, extending Meyer et al. (2021). ## Results Binary classification (disease presence/absence) evaluated with **Leave-One-Year-Out Cross-Validation** (2010--2019), AUC-ROC: | Rust Type | Meyer et al. (Logistic Regression) | Our Best Model | Improvement | |-----------|-----------------------------------|----------------|-------------| | **Stem** | 0.772 | **0.833** (XGBoost) | +0.061 | | **Stripe** | 0.598 | **0.719** (Gradient Boosting) | +0.121 | | **Leaf** | 0.638 | **0.798** (Gradient Boosting) | +0.160 | ### AUC Comparison Across Models ### Per-Year AUC Trajectories | Stem Rust | Stripe Rust | Leaf Rust | |:---------:|:-----------:|:---------:| | | | | ### Spatial vs Temporal Cross-Validation | Stem Rust | Stripe Rust | Leaf Rust | |:---------:|:-----------:|:---------:| | | | | ### Key Findings - **Per-rust model selection** outperforms a single model: XGBoost for stem rust (climate-driven), Gradient Boosting for stripe (race-driven) and leaf (spatial-driven) rust - **Spatial cross-validation** (leave-one-block-out, K=8) confirms generalization with only 5--7% AUC degradation for stem/leaf rust and no degradation for stripe rust - **SHAP analysis** identifies latitude-altitude interactions, pathogen race pressure, and crop growth stage as the dominant predictive features - ERA5 climate features help stem rust (+0.011 AUC) but not stripe/leaf rust - MODIS NDVI provides minimal direct signal; engineered features (anomaly, change) rank low ## Explainability (SHAP) ### Feature Importance | Stem Rust | Stripe Rust | Leaf Rust | |:---------:|:-----------:|:---------:| | | | | ### SHAP Beeswarm Plots | Stem Rust | Stripe Rust | Leaf Rust | |:---------:|:-----------:|:---------:| | | | | ### Top Feature Dependence Plots | Stem: Lat x Alt | Stripe: Race Pressure | Leaf: Altitude | |:---------:|:--------- …