farmstomarket/agri-harvest-cameroon

ml powered yield prediction platform for cameroon agriculture combining soil, weather, and satellite data across 8 agroecological zones. # Agri-Harvest End-to-end yield prediction platform for Cameroon agriculture. Ingests soil, climate, satellite, and crop survey data through two ML pipelines (scikit-learn and LightGBM/XGBoost/PyTorch) to predict harvest yields across 8 agroecological zones and 27 crop types. ## Quick start ```bash git clone github.com cd agri-harvest-cameroon python -m venv .venv && source .venv/bin/activate pip install -e ".[ml,geo,climate,dev]" cp .env.example .env ``` Requires Python 3.12+. Train a model: ```python from models.v1.trainer import YieldModelTrainer trainer = YieldModelTrainer("data/features.parquet") comparison = trainer.run(["lightgbm"], optimize=True) ``` ## Dataset The training dataset (3M rows) is hosted on Hugging Face: **synthi-ai/cameroon-agricultural-data** | Property | Value | |---|---| | Rows | 3,000,000 | | Columns | 36 raw / 66 engineered | | Crops | 27 types across 7 groups (cereals, legumes, root & tubers, vegetables, tree crops, industrial, cash crops) | | Zones | 8 agroecological zones | | Period | 2018 -- 2024 | | Sources | Field measurements, weather stations, lab analyses, TerraClimate, CHIRPS | ```python from datasets import load_dataset ds = load_dataset("synthi-ai/cameroon-agricultural-data", split="train") df = ds.to_pandas() ``` Three Jupyter notebooks walk through the data pipeline: | Notebook | Purpose | |---|---| | `01_data_exploration.ipynb` | Exploratory data analysis | | `02_feature_engineering.ipynb` | Feature engineering (36 raw -> 66 features) | | `data_generation_notebook.ipynb` | Synthetic data generation | ## Models ### v0 -- scikit-learn (up to ~500K rows) Spatial train/test split on `agroecological_zone` via `GroupShuffleSplit` (no zone leaks across sets). `StandardScaler` on 40 continuous features, passthrough for 22 binary + 4 ordinal. | Model | Type | Hyperparameters | |---|---|---| | Stacking | Ensemble | RF + HGB base, Ridge meta-learner | | Hist Gra …