📋 KEY INSIGHTS
- Precision agriculture uses data science to maximise crop yield, minimise input costs (fertiliser, water, pesticides), and reduce environmental impact by treating fields as heterogeneous landscapes rather than uniform areas — applying the right inputs at the right location and time based on observed variability. The five core data sources for precision agriculture are: satellite/multispectral imagery (Sentinel-2, PlanetScope, Landsat for vegetation indices and soil moisture), weather data (temperature, precipitation, solar radiation at hourly resolution from ERA5 reanalysis or on-farm weather stations), soil data (physical and chemical properties from grid sampling, digitalised soil survey maps, or proximal soil sensors), yield monitor data (GPS-tagged yield measurements from combine harvesters at sub-metre resolution), and equipment telematics (machinery position, speed, fuel consumption, field operation timing).
- Vegetation indices derived from multispectral satellite imagery are the most widely used remote sensing features for crop monitoring. NDVI (Normalised Difference Vegetation Index) = (NIR − Red) / (NIR + Red) measures canopy greenness and correlates with leaf area index, chlorophyll content, and biomass — high NDVI indicates dense, healthy vegetation. NDRE (Normalised Difference Red Edge) uses the red-edge band (680-730nm) instead of red and is more sensitive to chlorophyll content at high canopy density where NDVI saturates. NDWI (Normalised Difference Water Index) uses SWIR and NIR bands to estimate canopy water content and soil moisture. Time series of vegetation indices (NDVI trajectories through the growing season) are the strongest features for crop yield prediction: a field with early-season high NDVI, mid-season stable NDVI, and late-season gradual senescence typically outperforms one with irregular NDVI patterns that suggest stress events.
- Crop yield prediction models predict end-of-season yield before harvest from in-season satellite imagery, weather data, and soil properties. The dominant approaches: (1) empirical models (LightGBM or XGBoost trained on historical yield monitor data with NDVI time series, weather features, and soil properties as inputs — straightforward, interpretable, and competitive with more complex approaches); (2) process-based crop models (DSSAT, APSIM — simulate plant physiology from first principles, requiring detailed parameterisation but providing mechanistic insight); and (3) hybrid models (data-driven models calibrated against process-based model outputs, combining the extrapolation ability of mechanistic models with the flexibility of ML). The connection to our time series deep learning guide is direct: TFT applied to in-season NDVI and weather time series achieves state-of-the-art yield prediction accuracy at county and field levels, outperforming simpler models on the USDA NASS soybean and corn yield benchmarks.
- Soil health modelling uses spatial ML to map soil properties (organic carbon, pH, nitrogen, phosphorus, texture) across fields from remotely sensed proxies and sparse soil sample points. Digital Soil Mapping (DSM) uses the SCORPAN framework: soil properties are functions of climate (C), organisms (O), relief (topography, R), parent material (P), age (A), and spatial position (N) — machine learning models learn these relationships from soil sample points with high spatial coverage of SCORPAN covariates (satellite imagery, DEMs, geology maps). Gradient boosting models with SHAP interpretation are the standard DSM approach on global benchmarks (SoilGrids, LUCAS soil survey). Kriging (Gaussian Process interpolation with a spatial covariance structure) handles the spatial autocorrelation in soil properties: a soil measurement at one point provides information about nearby unmeasured points, and kriging quantifies this uncertainty with prediction intervals. The connection to our Gaussian Processes guide is direct: kriging is the geostatistical application of Gaussian Processes, and the GP framework unifies kriging with machine learning in the neural process and deep kriging approaches used for high-resolution soil mapping.
- Irrigation optimisation uses data-driven models of crop water stress and RL-based controllers to minimise irrigation water use while maintaining yield. The plant water stress index (computed from thermal infrared imagery — stressed plants have higher canopy temperature because they close stomata to reduce transpiration) triggers irrigation decisions. RL-based irrigation scheduling (Kelleners et al., 2020; Jha et al., 2022) trains a policy to decide irrigation timing and volume by modelling the soil-water-plant system as a Markov Decision Process: the state is soil moisture, crop growth stage, weather forecast, and accumulated heat units; the action is irrigation volume; and the reward is a combination of yield, water use efficiency, and energy cost. The connection to our RL for real-world applications guide is direct: irrigation scheduling is one of the cleanest real-world RL applications — the state is fully observable from soil sensors, the action space is small (irrigation volume per zone), and the reward is well-defined (maximise water use efficiency).
- Pest and disease detection uses computer vision on drone and smartphone imagery to identify crop diseases (leaf rust, blight, powdery mildew) and pest infestations at the individual plant level. CNN-based classifiers fine-tuned on the PlantVillage dataset (54,000 labelled plant disease images) achieve >95% accuracy on 38 plant disease classes. Field deployment challenges: class imbalance (healthy plants vastly outnumber diseased ones), domain shift (lab images vs. field images with different lighting, soil background, and growth stages), and the need for real-time inference on edge devices (tablets, drones with limited compute). The connection to our model compression guide is direct: plant disease detection on edge agricultural devices requires model compression (quantisation to INT8, pruning to reduce FLOPS) to fit CNN inference within the memory and latency constraints of field-deployed hardware.
Agriculture feeds eight billion people and accounts for roughly 70% of global freshwater use, 11% of greenhouse gas emissions, and 40% of land area — making it one of the highest-impact domains for data science application. Precision agriculture — using data to tailor farming inputs to spatial and temporal variability within fields — has the potential to simultaneously increase yields, reduce input costs, cut emissions, and improve soil health by applying the right amount of fertiliser, water, and pesticide at the right location and time, rather than applying uniform rates across heterogeneous fields. This guide covers the full agricultural data science toolkit: satellite and drone remote sensing for vegetation monitoring, crop yield prediction from in-season imagery and weather data, soil health mapping with digital soil mapping techniques, irrigation optimisation with RL, pest and disease detection with computer vision, and market price forecasting for agricultural commodities.
Remote Sensing, Vegetation Indices and Crop Yield Prediction
Sentinel-2 (European Space Agency) provides free multispectral imagery at 10-20m spatial resolution with 5-day revisit time — a transformative data source for agricultural monitoring at the field level. The 13 Sentinel-2 bands cover visible (blue, green, red), red-edge (three bands at 705nm, 740nm, 783nm), near-infrared, and short-wave infrared wavelengths, enabling a rich set of vegetation indices beyond NDVI. The red-edge bands are particularly valuable for agriculture: NDRE is more sensitive to chlorophyll content and nitrogen status than NDVI at the canopy densities typical of mature crops, and nitrogen status monitoring with NDRE enables variable-rate nitrogen application that reduces fertiliser cost and nitrous oxide emissions. Cloud cover is the main challenge for Sentinel-2 agricultural applications in tropical and humid temperate regions: frequent cloud cover creates gaps in time series that must be filled by temporal interpolation (Savitzky-Golay smoothing, Gaussian Process interpolation) or by fusing Sentinel-2 optical imagery with Sentinel-1 SAR (Synthetic Aperture Radar) imagery, which penetrates clouds and provides complementary information about crop structure and soil moisture. Crop yield prediction at the county level uses historical USDA NASS yield data as training labels combined with NDVI, EVI (Enhanced Vegetation Index), and weather time series as features. The LightGBM/XGBoost approach: aggregate per-pixel vegetation index time series to county-level statistics (mean, standard deviation, percentiles at key phenological stages — planting, peak greenness, senescence), combine with cumulative growing-degree days, precipitation, and drought indices (PDSI, SPI), and train a gradient boosting model on 30+ years of historical county-level observations. County-level corn and soybean yield prediction accuracy with this approach is typically R² = 0.90-0.95 and RMSE = 5-8% of mean yield, competitive with USDA official estimates. Field-level yield prediction (mapping sub-field yield variability) uses GPS-tagged combine yield monitor data as labels combined with high-resolution (3-5m) PlanetScope or drone imagery features. The connection to our spatial data science guide is direct: field-level yield prediction is a spatial regression problem with strong spatial autocorrelation (adjacent pixels in the same field have similar yield due to shared soil properties and topography), requiring the spatial feature engineering and spatial cross-validation approaches described there to avoid spatial data leakage.
| Application | Data Source | Model | Output | Benchmark/Metric |
|---|---|---|---|---|
| County yield prediction | Sentinel-2 NDVI + weather | LightGBM, TFT | Yield (t/ha) per county | R² >0.90 on USDA NASS |
| Soil organic carbon mapping | Sentinel-2 + DEM + soil samples | Gradient boosting + kriging | SOC map at 30m | RMSE <5 g/kg on SoilGrids |
| Crop type mapping | Sentinel-2 time series | Random Forest, LSTM | Crop type raster | OA >90% on EuroCrops |
| Disease detection | Drone/smartphone RGB | EfficientNet fine-tuned | Disease class + severity | >95% accuracy on PlantVillage |
| Irrigation scheduling | Soil sensors + weather forecast | SAC RL / rule-based | Irrigation volume per zone | WUE improvement 15-30% |
Soil Mapping, Irrigation Optimisation and Commodity Price Forecasting
Digital Soil Mapping (DSM) generates continuous spatial maps of soil properties from point observations (soil samples) and exhaustive spatial covariates (satellite imagery, digital elevation models, geology maps, land cover). The machine learning approach: treat each soil sample as an observation (soil property value at a known location), extract SCORPAN covariate values at the sample location, train a gradient boosting model to predict soil property from covariates, then apply the model to predict the entire map area. The critical challenge is spatial autocorrelation in model residuals: if the model’s residuals are spatially correlated (nearby samples have similar unexplained variation), standard cross-validation overestimates prediction accuracy. Spatial cross-validation — leaving out all observations within a minimum distance of each test point — provides unbiased accuracy estimates for spatially autocorrelated data. Kriging the residuals (after fitting the ML model) further improves accuracy by exploiting spatial structure in the unexplained variation. The connection to our Gaussian Processes guide is direct: DSM kriging uses a GP with a spatial covariance function (exponential, Matérn, or spherical) calibrated to the empirical semivariogram of soil samples, providing both predictions and uncertainty maps at unsampled locations. Commodity price forecasting for agricultural markets (corn, soybean, wheat futures) combines crop production forecasts (from yield models), supply-demand balance models, and macro-financial factors (dollar index, energy prices, shipping costs) to forecast price levels and volatility at 1-3 month horizons. The modelling approach: tree-based models (gradient boosting) for point forecasts and GARCH models for volatility forecasting, with the production forecast as the most important feature at the peak of growing season and macro-financial factors dominating in the off-season. The connection to our data science for finance guide is direct: agricultural commodity futures are traded financial instruments, and the same quantitative finance techniques (GARCH volatility modelling, momentum and carry factor models) used for equity and FX markets apply to agricultural commodity trading. Food supply chain data science: connecting field-level production forecasts to supply chain optimisation (inventory, logistics, processing capacity) is a growing area that bridges our supply chain data science guide with the agricultural applications described here — estimating harvest timing from crop growth stage models enables logistics companies to pre-position trucks and processing capacity before harvest.
✦ SUMMARIZE THIS ARTICLE WITH AI
Spatial data science (satellite imagery processing, spatial cross-validation, spatial autocorrelation) is in our Spatial Data Science guide. Time series deep learning (TFT for in-season yield prediction) is in our Time Series Deep Learning guide. Gaussian Processes (kriging for soil mapping, GP interpolation for NDVI time series gap filling) are in our Gaussian Processes guide. RL for real-world applications (irrigation scheduling as an MDP) is in our RL guide. Model compression (EfficientNet quantisation for edge disease detection) is in our Model Compression guide. Data science for supply chain (harvest logistics optimisation) is in our Supply Chain guide. Data science for climate (agriculture-climate feedbacks) is in our Climate Data Science guide.



