Using statistical learning algorithms to predict cover crop biomass and nitrogen content

Authors: MARCILLO, GUILLERMO - North Carolina State University; Mirsky, Steven; AURELIE, PONCET - North Carolina State University; REBERG-HORTON, S CHRIS - North Carolina State University; Timlin, Dennis; Schomberg, Harry; RAMOS, PAULA - North Carolina State University

Submitted to: Agronomy Journal
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/17/2020
Publication Date: 10/28/2020
Citation: Marcillo, G.S., Mirsky, S.B., Aurelie, P., Reberg-Horton, S., Timlin, D.J., Schomberg, H.H., Ramos, P. 2020. Using statistical learning algorithms to predict cover crop biomass and nitrogen content. Agronomy Journal. 112(6):4898-4913. https://doi.org/10.1002/agj2.20429.
DOI: https://doi.org/10.1002/agj2.20429

Interpretive Summary: Biomass and N content are prime indicators of the potential services that cover crops can bring to a cropping system. Optimum amounts of cover crop biomass, for example, can help growers to control weeds, reduce soil erosion, or cool-off soils to reduce water loss. Cover crop biomass and N content are highly variable in response to soil, weather, and management. Models that account for this variability are needed such that growers can obtain early-season estimates of cover crop performance that facilitate scheduling their operations. In this research, we used machine learning to optimize predictions of late-season biomass and N content of cereal rye cover crops in the North and South Eastern US. The field experiments were conducted to evaluate cereal rye biomass and N content in response to N fertilizer. Overall, cover crop biomass and N content predictions from the models were 80% and 69% accurate when compared to field ground truth. Furthermore, cover crop attributes sampled early in the season, such as C:N ratio or tiller counts, were shown to complement well the remote-sensed inputs, e.g., NDVI, and contributed to enhancing the biomass and N predictions of the models. The results of this work demonstrated that relatively inexpensive inputs, collected when a cover crop is “greening up”, can be effectively used to predict late-season biomass and N content of a cover crop in the Eastern United States. These results show that modern data-driven methods can be successfully applied to characterize cover crop performance and will guide future research given the momentum and funding pushed to design big-data pipelines for sustainable agriculture. This work will be used by researchers and can serve as input to decision support tools.

Technical Abstract: Cereal rye (Secale cereale sp.) is a cover crop species known to improve soil and water quality. Late season biomass production is information growers need to maximize cover crop benefits and schedule field operations. Statistical learning (SL), built upon statistical and computational algorithms that “learn” from data, may help to improve predictions of cover crop biomass as a function of initial soil inorganic nitrogen levels. Three models [LASSO (Least Absolute Shrinkage and Selection Operator)], Ridge, and Random Forest (RF) were trained and optimized on a three-year allometric and remote sensing dataset of cereal rye responses to N fertilization in the Mid-Atlantic North and Southeast US. Shoot biomass (mean= 9,800 kg ha-1) was accurately predicted with a RF model (RMSE = 2,039 kg ha-1). Targeting shoot N content (mean= 107 kg ha-1), on the other hand, LASSO made accurate and more stable predictions (RMSE = 34 kg. ha-1). Early-season information (cover crop C:N ratio, tiller counts, and ground-sensed NDVI) contributed to enhancing biomass and N content predictions. A final test on untrained data revealed that 92 and 73% of the predictions from either algorithm corresponded to ground-truthed biomass and shoot N content observed under different N regimes. Modern data-intensive approaches, such as statistical learning, show promise to characterize end-season performance of a cover crop and may contribute to better farm decision making.