Spectral Sensing for Forage Nutritive Value Determination of Cool Season, Grass Pastures During the Grazing Season

Abstract

Management surveys suggest that few cow-calf producers in the Southeastern U.S. submit forage samples for laboratory analysis due to time and labor constraints. Although tools like near infrared reflectance spectroscopy (NIRS) have helped reduce costs associated with nutritive value determination in stored feeds, their performance for pasture analysis has been limited. Our objective was to explore the efficacy of spectral sensing in predicting the dry matter (DM), acid detergent fiber (ADF), neutral detergent fiber (NDF), and crude protein (CP) of fresh forages during the growing season. Weekly from May through October, two random samples were collected from each of 12 fields. Spectral readings were taken above canopy level in-field and again in-lab, followed by bench chemistry analyses of DM, ADF, NDF, and CP. Chemistry results and spectral readings were aligned by field, sample, and date. The 18 individual light spectra and lidar-measured distance were used as features in a random forest regression fit to predict each nutrient and separate models were developed for in-field and in-lab spectral readings. Data were randomly split for hyperparameter tuning (15%), model training (55%), and independent evaluation (30%). The root mean squared prediction error (RMSPE), calculated on the independent evaluation data, was used to explore the viability of this system to predict forage nutritive value. The in-field and in-lab models performed similarly for each forage nutritive value. To evaluate the prediction capability of the system under various atmospheric conditions, cloud cover was added as a feature in each in-field regression. The RMSPE of DM, ADF, NDF, and CP with cloud cover were 21.8%, 9.88%, 10.1%, and 21.9%, respectively. These models were also evaluated on new, unseen data from nine subplots and used to explore the implications of the prediction errors. The NASEM (2018) Beef Cattle Nutrient Requirements model was used to simulate diet nutritional adequacy using forage nutritive value estimated from the spectral sensor compared with forage nutritive value measured by bench chemistry. These forage nutritive value estimation methods resulted in a 4.48% and 3.03% difference in metabolizable energy (ME) and metabolizable protein (MP) allowable gain, respectively. Considerable future data collection and model refinement efforts are necessary to determine the value of the spectral sensing system in supporting low-cost, in-field nutritive value monitoring.