idfarhan/ch4_emission_model

Estimating CH4 emissions over Africa using ML/DL based methods # CH4 Emission Model — Africa End-to-end machine-learning pipeline that estimates anthropogenic CH₄ emissions over Africa from satellite observations and reanalysis data, trained against EDGAR v2025 inventory and validated on an unseen year. - **Domain** : Africa, lon ∈ [−20°, 55°], lat ∈ [−40°, 40°] - **Grid** : 0.1° × 0.1° (NY = 800, NX = 750), aligned to EDGAR - **Train** : 2019–2023 (stratified-by-year 80/20 train/test split) - **Predict** : 2024 (full-grid, also used as unseen-year evaluation) - **Targets** : `emissions` (tonnes/cell/year) **or** `flux` (kg m⁻² s⁻¹) - **Models** : XGBoost, Random Forest, CNN, CNN-LSTM --- ## 1. Repository layout ``` ch4_emission_model/ ├── config.py # paths, grid, features, target, skip-zero & noise-floor flags ├── common.py # data loaders, splitter, metrics, CSV/NetCDF writers ├── prepare_data.py # build per-year NetCDF stacks (TROPOMI + ERA5 + LandScan + EDGAR) ├── xgboost_model.py # XGBoost + 9-stage sequential GridSearchCV tuning ├── random_forest_model.py # Random Forest + 5-stage sequential tuning ├── cnn_model.py # per-pixel CNN (9×9 patches) + 5-stage tuning ├── cnn_lstm_model.py # CNN-LSTM (T=3 yr window, 9×9 patches) + 5-stage tuning ├── slurm_xgboost_emissions.sh ├── slurm_random_forest_emissions.sh ├── slurm_cnn_emissions.sh ├── slurm_cnn_lstm_emissions.sh ├── data/ # training_data_ .nc (built by prepare_data.py) ├── models/ # saved models (xgboost_ .pkl, ..._ .pt) └── outputs/ ├── xgboost/ / # tuning_*.csv, best_params_*.json, train/test_pred_*.csv, metrics_*.csv ├── random_forest/ / ├── cnn/ / ├── cnn_lstm/ / └── predictions_2024_ _ .nc ``` ` ` is either `emissions` or `flux`. --- ## 2. Input datasets | Dataset | Variable used | Native resolution | Source path | |---|---|---|---| | TROPOMI WFMD (daily L2) | column-averaged dry-air XCH₄ (`xch4`, QA == 0) | ~0.1°, swath | `shared/farha …