zoom-BT/zindi-chad-crop-yield

Winning solution (1st place, private LB) for a Zindi challenge predicting crop yields in Chad from satellite climate/vegetation data. 5-model tree ensemble + Ridge stacker, sequential 2024→2025 inference, zero external data. # Zindi Chad: Winning Solution (Zoom387) **Challenge:** Prédiction IA des rendements agricoles par télédétection et séries temporelles **Team / username:** Zoom387 **Best submission:** `submission_v16_best3.csv` **Submission ID:** `e3vJYTzM` **Submitted:** 15 August 2026, 07:40 **Public LB:** wMAPE ≈ **0.19848** (rank 3) **Private LB:** **#1** This package reproduces the exact submitted file and documents the full training pipeline. --- ## 1. How to set up folders Unpack this archive so the folder structure is: ``` Zoom387_Zindi_Chad_Winning_Solution/ ├── README.md ← this file (setup + run order) ├── SOLUTION.md ← full documentation (Zindi doc article) ├── CODE_REVIEW.md ← checklist vs Zindi code-review article ├── requirements.txt ├── generate_submission.py ← STEP 1: exact reproduction (< 5 s) ├── 00_data_insights.ipynb ← EDA / insights for the host ├── 01_generate_submission.ipynb ← same as STEP 1, notebook form ├── 02_train_from_scratch.ipynb ← STEP 2 (optional): full retrain ├── submission_v16_best3.csv ← file submitted on Zindi (reference) ├── train.csv ← Zindi original data ├── test.csv ├── SampleSubmission.csv ├── generate_submission.py ├── models/ │ ├── cv_metrics_v6.json ← Optuna hyperparameters │ └── v12_seeds/ │ ├── seed_7.csv ← seed predictions from winning run │ ├── seed_456.csv │ └── seed_789.csv ├── src/ │ ├── features.py │ ├── metrics.py │ ├── train_v3.py │ └── train_v9.py └── output/ ← created at runtime └── submission_v16_best3.csv ``` **Data used:** only the original Zindi files `train.csv`, `test.csv`, `SampleSubmission.csv`. **No external data** (no FAO, NASA POWER, FAOSTAT, DPSA, etc.). **Output:** `output/submission_v16_best3.csv` (identical to the submitted file). --- ## 2. Environment - **OS tested:** Windows …