A practical model and online tool to predict cover crop nitrogen content and carbon-to-nitrogen ratio via remote sensing

Purpose:

Estimations of cover crop (CC), nitrogen (N) content and carbon-to-nitrogen (C:N) ratio are crucial for guiding N fertilizer recommendations for subsequent cash crops, yet prediction tools that can be easily adopted by practitioners are lacking. This study aims to (1) develop models to predict CC N content and C:N ratio using Normalized Difference Vegetation Index (NDVI) and Normalized Difference Red Edge Index (NDRE) from Uncrewed Aerial Vehicles (UAV) and satellite imagery, combined with growing degree days (GDD, base 5 °C) and CC species; and (2) integrate the models developed into an online application. 

Methods:

Over 260 CC biomass samples were collected from 10 Pennsylvania counties in the Chesapeake Bay Watershed representing diverse species (triticale, cereal rye, wheat, brassicas, legumes, and mixtures). Four models were evaluated: (1) UAV NDVI + GDD + species, (2) UAV NDRE + GDD + species, (3) satellite NDVI + GDD + species, and (4) satellite NDRE + GDD + species.

Results:

Random forest models (R² > 0.74 for N content and R² > 0.86 for C:N ratio) generally outperformed generalized additive models (R² > 0.70 for N content and R² > 0.85 for C:N ratio) based on 10-fold cross validation. GDD and vegetation indices were the strongest predictors for N content, while GDD and species group influenced C:N ratio the most. In the independent site-years validation, UAV-NDRE performed best for predicting CC N content (R² = 0.82), and both UAV-based models (R² = 0.52) outperformed the satellite models for C:N ratio prediction. 

Conclusion:

Random forest models accurately predicted CC N content and C:N ratio and were integrated into an online application to support precision agriculture practices.