Crop Yield Prediction with ML & Satellite Data

Traditional yield forecasting relies on expert estimates and historical averages, often resulting in 20–40% MAPE. We replace this with ML models using satellite NDVI and meteorological data, achieving 8–12% MAPE 4–6 weeks before harvest. Our approach outperforms expert estimates by 2–3 times. NASA Earth Observatory confirms that combining NDVI with growing degree days (GDD) improves accuracy by 30%. We've delivered 30+ projects for agriholdings and banks. Request a demo for your farm and get a consultation on implementation.

How ML Outperforms Traditional Methods

Traditional methods (multi-year averages, expert opinions) yield 20–40% error. ML models capture dynamics of vegetation indices (NDVI, EVI, LAI) and accumulated temperatures (GDD). Result: MAPE 8–12% within 4–6 weeks before harvest. The model automatically identifies phenophases and adjusts forecast under stress (drought, frost). ML gives 2–3× higher accuracy, reducing storage and logistics costs by up to 20%.

What Data Is Needed for Accurate Prediction

Satellite Data (Remote Sensing):

• Sentinel-2 (ESA): 10–20 m resolution, 5-day revisit, free
• Landsat 8/9 (NASA/USGS): 30 m, free
• PlanetScope: 3 m resolution, daily, commercial

Vegetation Indices:

• NDVI (Normalized Difference Vegetation Index): (NIR - Red) / (NIR + Red) — density and health
• EVI (Enhanced Vegetation Index): improved, resistant to atmospheric effects
• LAI (Leaf Area Index): leaf area, linked to biomass

Meteorological Data:

• Air temperature (min/max/avg): accumulated GDD
• Precipitation: decadal, monthly, seasonal totals
• Solar radiation
• Soil moisture (from Sentinel-1 SAR or agrometeorological stations)

Soil Data:

• SoilGrids (ISRIC): global soil type map, 250 m resolution
• SoilMap: national soil quality maps

Model Architecture

We use three approaches depending on data volume and task.

Feature-based ML

# For each field: aggregated features per season
field_features = {
    'ndvi_peak': max(ndvi_time_series),
    'ndvi_integral': sum(ndvi_time_series),  # seasonal biomass
    'gdd_accumulated': sum(max(0, temp_avg - base_temp)),
    'precipitation_total': sum(precipitation),
    'drought_days': count(spi < -1),  # SPI: Standardized Precipitation Index
    'soil_type_encoded': one_hot(soil_type),
    'field_size_ha': field_area,
    'variety_encoded': crop_variety_embedding
}
model = LightGBM.train(field_features, yield_targets)

Time Series Deep Learning

NDVI time series + weather per season → LSTM → yield. Advantage: uses growth dynamics, not only final aggregates.

Process-based Hybrid

Simulation crop model (DSSAT, APSIM) + ML correction. Physical model provides structure, ML learns residuals from real data. LightGBM runs 1.5× faster than LSTM on small datasets.

Why Phenology Tracking Is Critical

Growth stages (phenology) directly affect final yield. Automatic phenophase detection from NDVI dynamics and GDD allows the model to react to stress. Delay in phenophase due to drought or frost reduces the forecast. Without phenology, accuracy drops by 15%.

How to Automatically Detect Phenophases

We build the NDVI curve over the season and locate characteristic points: start of season, peak, senescence. Simultaneously we calculate accumulated GDD. If the NDVI peak lags behind the norm GDD, the model signals stress and reduces the yield forecast. This method increases accuracy by 15% compared to a model without phenology.

Spatial Aggregation

Field → Farm → District → Region:

• Field-level forecast: for agronomists (10–30 ha)
• Farm level: for financial planning
• District/region level: for government monitoring

Geostatistical approach: Kriging interpolation for spatially continuous yield map. Allows estimating fields without historical data.

Practical Application

Agriholding:

• Storage capacity planning
• Forward sales contracts
• Operational alerts for fields with lagging NDVI

Bank Lending:

• Collateral valuation of standing crop
• Default risk assessment under predicted drought

Integration with Ag Platforms:

• Agrosignal, GIS Mercury: Russian field management platforms
• Trimble Ag Software, John Deere Operations Center: global
• Export via API to ERP (SAP/1C:Agro)

Savings for an average agriholding are estimated at 5–10 million rubles per season from optimized logistics and storage. Contact us for a preliminary assessment and consultation on AI prediction implementation.

What's Included in the Work

1. Data audit — check availability and quality of satellite imagery, weather data, historical yields.
2. Model prototype — build a baseline on one crop (2–3 weeks).
3. Production pipeline development — automated data collection, training, validation, deployment.
4. Integration — API, data dashboard, connection to 1C or SAP.
5. Team training — transfer know-how to your agronomists and analysts.
6. Support — maintenance, model retraining each season, accuracy guarantee.

Timelines and Guarantees

Timelines: basic NDVI-based model for one crop/region — 5–7 weeks. Multi-crop system with phenology tracking and API — 3–4 months. Cost is individual after audit.

Guarantee: we guarantee MAPE no higher than 15% on the pilot. If accuracy is lower, we refine it free of charge. Request a demo version for your agriholding and get a consultation on implementation.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.