About N Scout
Detecting nitrogen deficiency in corn fields using Sentinel‑2 satellite imagery and the PRNT public research dataset.
This demo uses precomputed predictions on the 7 PRNT trial sites from 2016. Predictions include both training-set and test-set plots. Performance on training-set plots will appear inflated. The honest model evaluation metrics below are from the held-out test set only. This tool is a proof of concept for a hackathon and should not be used to guide real nitrogen management decisions.
How it works
N Scout detects probable nitrogen deficiency at the plot level by combining Sentinel-2 multispectral imagery with optional soil nitrate and weather records from the PRNT dataset. The core steps are:
1. Remote sensing features — 16 spectral features per V-stage: 5 raw reflectance bands (B02 blue, B03 green, B04 red, B05 red-edge, B08 NIR) and 11 indices (NDVI, GNDVI, NDRE, EVI2, CIrededge, NIRv, SAVI, OSAVI, TGI, MCARI, OCARI) extracted from Sentinel-2 (10 m) at growth stages V6 through V10 — 80 features total. Cloud gaps are filled using a nearest-in-time mosaic strategy.
2. Ground-truth labels — Nitrogen Nutrition Index (NNI) computed from PRNT plant tissue N and biomass measurements. Plots with NNI < 1.0 are labelled deficient (25.9 %).
3. Fusion features — 9 soil nitrate / ammonium columns (PPNT + PSNT sampling times) and 6 growing-season weather metrics (GDD, precipitation, solar radiation, heat days) aggregated over the planting to V10 window.
4. Classification — Logistic Regression trained with stage-dropout augmentation (robust to missing V-stages) and an F-beta (beta=2) threshold that weights recall over precision.
Model performance (test set)
Honest evaluation on held-out plots (treatments 1–8 only; 45 test plots, 25.9 % N-deficient).
Satellite only (RS model)
Satellite + Soil & Weather (fusion model)
Soil and weather context provides the dominant signal: adding those features improves AUC from 0.634 to 0.904 and lifts F1 from 0.514 to 0.870. Satellite imagery refines but does not replace the agronomic context.
Proximal sensing (leaf image classifier)
A MobileNetV3-Small model fine-tuned on a public maize leaf nutrient deficiency dataset classifies individual leaf images into 6 nutrient classes. This complements the satellite view by confirming plot-level deficiency at the individual plant level.
Overall test accuracy: 98.5% on 882 held-out images (7,056 train / 882 val / 882 test). Model: MobileNetV3-Small via timm, fine-tuned from ImageNet weights with Optuna HPO.
Data
All field data are from the publicly available PRNT dataset (Public-industry Research Network for N rate response studies in corn, 2014–2016), published on Dryad. The dataset covers 49 site-years across 8 U.S. Midwest states. This demo uses the 7 trials from 2016 that had reliable NNI ground truth and usable Sentinel-2 coverage.
Satellite imagery: Sentinel-2 Level-1C (TOA) via Google Earth Engine, COPERNICUS/S2_HARMONIZED, cloud-filtered at 30 % scene level + QA60 pixel mask.
Trial sites used
33 — IA Crawford34 — IA Story37 — IN Loam40 — MN Waseca41 — MO Bradford42 — MO Loess43 — MO TrothTechnical stack
ML pipeline: Python · scikit-learn · XGBoost · LightGBM · Google Earth Engine
Proximal model: PyTorch · timm · MobileNetV3 · Optuna
Backend: Flask 3 + Gunicorn · Python 3.14 · uv
Frontend: Next.js (static export) · Leaflet · TypeScript
Deployment: Docker multi-stage build to Render
Team
Built at Precision & Digital Agriculture Hackathon by Gustavo, Leonardo, Natalia, Pedro from the Ciampitti Lab, Purdue University.