Moving beyond book values: Using machine learning to improve the accuracy of manure nutrient concentration...

Area of Science:

• Agricultural Science
• Soil Science
• Data Science

Background:

• Animal manure is a valuable nutrient source, but its composition varies significantly between farms.
• Static average nutrient values are often unreliable for precise crop nutrient demand.
• Regional and farm-specific variations in manure nutrient content necessitate advanced analytical methods.

Purpose of the Study:

• To analyze regional variations in total nitrogen (N), phosphorus (P2O5), and potassium (K2O) in animal manure across the US.
• To develop and evaluate machine learning models for predicting total NPK concentrations in manure.
• To compare the predictive performance of different machine learning models and identify key predictor variables.

Main Methods:

• Utilized the ManureDB database for analyzing manure nutrient composition.
• Employed three machine learning models: Random Forest, Extreme Gradient Boosting, and a hybrid approach.
• Used predictor variables including ammonium-N (NH4-N), pH, moisture percentage, region, and animal source.

Main Results:

• Machine learning models demonstrated high predictability for total nitrogen (R² = 0.8-0.9).
• Moderate predictability was achieved for total phosphorus (R² = 0.6-0.7), and lower predictability for total potassium (R² = 0.4-0.7).
• Model performance varied between solid and liquid manure, with better N prediction in liquid and better K prediction in solid forms.

Conclusions:

• Machine learning models offer a more precise alternative to static nutrient averages for manure management.
• Key predictors for solid manure include animal source and NH4-N, while moisture is secondary.
• NH4-N and moisture percentage are primary predictors for liquid manure, highlighting the complexity of manure nutrient variability.