A PyTorch Multilayer Perceptron (MLP) that forecasts crop yield (kg/ha) one year ahead for 165 countries and 102 crop types, trained on multi-source climate, soil, and land-cover data spanning 2010–2022. Achieves R² = 0.9452 and Pearson r = 0.9681 on unseen validation years (2020–2021).
| Metric | Value |
|---|---|
| Validation R² | 0.9452 |
| Validation Pearson r | 0.9681 |
| Validation RMSE | 3926.49 kg/ha |
| Validation MSE | 1.54 × 10⁷ |
| Training MSE | 9.62 × 10⁶ |
| Training samples | 52,370 |
| Validation samples | 10,368 |
| Prediction samples (2022) | 5,190 |
| Countries | 165 |
| Crop types | 102 |
| Input features (post-encoding) | 43 |
| Epochs | 120 |
- Temporal train/val split — no data leakage; model trained on 2010–2019, validated on 2020–2021, deployed on 2022
- Exponential decay weighting — recent years contribute more to weight updates via
w = e^(−λ · year_diff), λ = 0.1 - Multi-source feature engineering — 30+ raw climate variables collapsed into meaningful annual summaries (mean, min, max, sum)
- Haversine-based geo-matching — satellite grid points assigned to countries using great-circle distance with centroid radius logic
- Log-space target —
log(1 + yield)applied before training, back-transformed before evaluation to handle skewed kg/ha values - BatchNorm over Dropout — empirically found batch normalisation sufficient to prevent overfitting without dropout
The model converges rapidly in the first 20 epochs. Both train and validation R² exceed 0.94 and plateau after epoch ~80, confirming stable generalisation with no overfitting.
R² Score over Epochs
Loss (MSE) over Epochs
The network is a fully-connected feed-forward MLP:
Input → [Linear(d → 128) → BatchNorm → ReLU]
→ [Linear(128 → 64) → BatchNorm → ReLU]
→ Linear(64 → 1)
→ Predicted Yield (kg/ha)
| Component | Detail |
|---|---|
| Input layer | d-dimensional feature vector (numeric + one-hot) |
| Hidden layer 1 | Linear → BatchNorm(128) → ReLU |
| Hidden layer 2 | Linear → BatchNorm(64) → ReLU |
| Output layer | Linear → scalar (log-yield) |
| Optimizer | Adam, lr = 0.001 |
| Loss function | Weighted MSE |
| Batch size | 64 |
| Epochs | 120 |
Why no Dropout? After extensive experimentation, BatchNorm alone prevented overfitting. Dropout introduced unnecessary variance with no measurable regularisation gain on this dataset.
A snapshot of model predictions against true yield values for the 2021 validation year. Predictions are back-transformed from log-space to the original kg/ha scale before comparison.
| Country | Predicted (kg/ha) | Actual (kg/ha) | Residual | Crop |
|---|---|---|---|---|
| Algeria | 1997.48 | 1731.30 | -266.18 | Triticale |
| Algeria | 1829.48 | 1717.00 | +112.48 | Treenuts, Total |
| Angola | 642.35 | 643.50 | +1.16 | Groundnuts, excl. shelled |
| Angola | 4714.58 | 4749.80 | +35.22 | Other fruits, n.e.c. |
| Angola | 744.17 | 710.00 | -34.17 | Sunflower seed |
The predictions_2022.csv file in this repo contains model predictions for the 2022 forecast year alongside true values for all 5,190 country–crop pairs.
The model uses multi-source earth observation data merged at the country–year level:
| Feature Group | Variables | Aggregation |
|---|---|---|
| Soil Moisture | 0–10 cm, 10–40 cm, 40–100 cm, 100–200 cm | Annual mean |
| Soil Temperature | 4 depths × (mean, min, max) = 12 columns | Annual mean/min/max |
| Evaporation (ESoil) | ESoil_tavg | Annual mean |
| Transpiration (TVeg) | TVeg_tavg | Annual mean |
| Canopy Water | CanopInt | Annual mean |
| Terrestrial Water Storage | TWS anomaly | Annual mean |
| Precipitation – Rain | Rainf_tavg | Annual total (mm) |
| Precipitation – Snow | Snowf_tavg | Annual total (mm) |
| Land Cover | 17 land-use class percentages | Per-country mean |
| Feature | Detail |
|---|---|
| Country | One-hot encoded (165 countries) |
| Crop Item | One-hot encoded (102 crop types) |
| Year | Numeric (2010–2022) |
Yield (kg/ha) for year T+1 — features from year T are paired with the yield from year T+1 by decrementing the yield year by 1 before merging.
A production-grade preprocessing pipeline was built entirely from scratch:
| Variable | Strategy |
|---|---|
| TWS month 2 & 3 (~14,946 rows) | Weighted linear interpolation: ⅔ × Jan + ⅓ × Apr |
| Soil Moisture 100–200cm month 8 | Average of month 7 (July) and month 9 (September) |
| Country lat/long nulls (7 rows) | Dropped — insufficient coverage data |
| Variable | From | To | Method |
|---|---|---|---|
| Precipitation | kg/m²/s | mm/month | × seconds/month ÷ 1000 |
| Soil Temperature | Kelvin | °Celsius | − 273.15 |
| Yield | raw kg/ha | log-scale | log(1 + x) |
Each satellite grid point (lat, lon) was assigned to a country using the Haversine formula to compute great-circle distance to each country centroid. Points falling within a country's radius boundary were assigned to the closest centroid. Edge cases were resolved by choosing the country whose border was nearest.
Monthly time-series columns were collapsed into annual summaries:
- Mean — used for soil moisture, soil temperature, ESoil, TVeg, canopy water, TWS
- Sum — used for precipitation (total annual rainfall/snowfall matters more than rate)
- Min + Max — added for soil temperature to capture seasonal extremes (overwintering stress, heat stress)
Numeric features → StandardScaler (zero mean, unit variance)
Categorical → OneHotEncoder (ignore unseen categories)
Combined via → ColumnTransformer (sklearn Pipeline)
To emphasise more recent agricultural patterns (e.g., climate shifts post-2015):
weight = exp(−λ × (target_year − sample_year)), λ = 0.1
Samples from 2019 receive weight ≈ 1.0; samples from 2010 receive weight ≈ 0.41. These weights are passed directly to the loss function per batch.
git clone https://github.com/alinjfz/crop-yield-forecasting-mlp-pytorch.git
cd crop-yield-forecasting-mlp-pytorchpip install torch scikit-learn pandas numpy matplotlib tqdmPlace the raw dataset files (soil, weather, land cover, yield CSVs) in a data/raw/ directory. The notebook documents each expected filename at the top of each preprocessing cell.
jupyter notebook crop_yield_forecasting.ipynbThe notebook is structured sequentially:
- Preprocessing — imputation, unit conversion, merging, geo-assignment
- Feature engineering — annual aggregation, decay weights, train/val split
- Model definition & training — MLP with BatchNorm, Adam, weighted MSE
- Evaluation — R², Pearson r, RMSE on validation set (2020–2021)
- 2022 Forecasting — generates
predictions_2022.csv
crop-yield-forecasting-mlp-pytorch/
├── crop_yield_forecasting.ipynb # Full pipeline: preprocessing → training → prediction
├── predictions_2022.csv # Model output: Country, Item, True_Value, Predicted_Value
├── report.pdf # Technical report with full methodology
├── assets/
│ ├── model_radar.png # Performance radar chart
│ ├── r2_training_curve.png # R² score over 120 training epochs
│ ├── loss_over_epochs.png # MSE loss over 120 training epochs
│ ├── network_architecture.png # MLP layer diagram
│ └── report_title.png # Report cover page
└── LICENSE
| Tool | Role |
|---|---|
| PyTorch | MLP model definition, custom training loop, weighted loss |
| scikit-learn | ColumnTransformer, StandardScaler, OneHotEncoder, SimpleImputer |
| pandas | Data loading, merging, groupby aggregation |
| NumPy | Haversine distance computation, array operations |
| Matplotlib | Training curve visualisation |
| tqdm | Epoch progress monitoring |
- 165 countries × 102 crops — the model must generalise across radically different climates and crop types simultaneously
- No random split — temporal splitting (train ≤ 2019, val 2020–2021) forces genuine forecasting with no future leakage
- Multi-source heterogeneous data — merging satellite climate grids, yield statistics, and land-cover maps required custom geo-matching logic
- Sparse targets — not every country grows every crop, creating a non-uniform label space
- Scale variance — yields range from tens to hundreds of thousands of kg/ha depending on crop type, making log-transformation critical
Ali Najafzadeh MSc Artificial Intelligence — University of Sussex GitHub · LinkedIn
This project is licensed under the MIT License. See LICENSE for details.