dienbien-coffee-yield

Coffee Yield Forecasting Model — Điện Biên, Vietnam

Predicts coffee yield (tons/hectare) for districts in Điện Biên province, NW Vietnam, from growing-season weather features (rainfall timing by phenological stage, frost/cold days, growing degree days, elevation).

Part of a broader yield forecasting & harvest planning pipeline (weather ingestion → feature engineering → model → harvest-window + risk alerts).

Model details

• Algorithm: Random Forest
• Format: skops (model.skops) — safe to load without pickle's arbitrary-code-execution risk
• Cross-validated MAE: 0.115 t/ha (5-fold CV)
• Cross-validated R²: 0.471
• Training samples: 80 (district-year observations)
• Trained: 2026-07-18

⚠️ Important: this model was trained on synthetic, domain-informed weather+yield data, not real measured yield records (see the parent repo's README for why, and how to retrain on real GSO/ICO data). Treat predictions as illustrative until retrained on real district-level yield history.

Input features

dry_spell_rain_mm
flowering_rain_mm
fruitdev_rain_mm
ripening_rain_mm
fruitdev_rain_cv
frost_days_sensitive
mean_annual_temp_c
growing_degree_days
mean_sunshine_ripening
elevation_m
annual_rain_mm

All features are derived from daily weather (tmax_c, tmin_c, tmean_c, precip_mm, et0_mm, sunshine_hours) aggregated over crop-specific phenological windows — see features.py in the parent repo for exact definitions (e.g. flowering_rain_mm = total rainfall during the flowering month(s), frost_days_sensitive = count of nights below the frost threshold during frost-sensitive months).

Usage

from huggingface_hub import hf_hub_download
import skops.io as sio
import pandas as pd

model_path = hf_hub_download(repo_id="imaflower/dienbien-coffee-yield", filename="model.skops")
model = sio.load(model_path, trusted=sio.get_untrusted_types(file=model_path))

X = pd.DataFrame([{
    "dry_spell_rain_mm": 30, "flowering_rain_mm": 60, "fruitdev_rain_mm": 450,
    "ripening_rain_mm": 120, "fruitdev_rain_cv": 1.8, "frost_days_sensitive": 1,
    "mean_annual_temp_c": 21.5, "growing_degree_days": 4200, "mean_sunshine_ripening": 6.2,
    "elevation_m": 900, "annual_rain_mm": 1600
}])
predicted_yield_t_ha = model.predict(X)[0]

Intended use & limitations

• Scoped to Điện Biên province district-level, seasonal (annual) yield forecasting — not validated for other provinces/climates without retraining.
• Does not account for management practices (fertilization, pest control, variety-specific agronomy beyond the multiplier applied in the parent pipeline's predict.py).
• Synthetic training data encodes plausible agronomic relationships but is not a substitute for real yield statistics.