Machine learning models for dry matter and biomass estimates on cattle grazing systems

Abstract

Monitoring pasture conditions contributes to the animals' decision-making process, avoiding supplementation losses, and improving cattle performance. Environmental parameters and herd characteristics can influence pasture quantity (biomass and dry matter) and controlling these parameters is a challenge nowadays. Some research work aim to monitor pasture quantity parameters from environmental and spectral data. Our research did not find studies considering the impact of the animals together with spectral data and environmental parameters to estimate pasture quantity parameters. This work aims to design machine models to estimate biomass and dry matter from Brachiaria brizantha cv. Marandu pastures, analyzing the effects of the environment and animals' features on the models. We collected pasture samples and analyzed them in a laboratory to determine the biomass and dry matter from 12 paddocks over two years in the spring and summer seasons. In addition, we acquired and associated spectral and environmental data and herd characteristics (gender and the number of animals), from the Nellore breed, with biomass and dry matter to design the models. Variable selection methods such as Pearson's correlation analysis and Recursive Feature Elimination (RFE) were applied to determine which parameters significantly affected the response variables. Non-linear and linear prediction models were designed and compared. The results indicated that the number of animals, gender, and vegetation indices have meaningful effects on the response variables. Furthermore, the non-linear models achieved expressive results predicting biomass and dry matter. The eXtreme Gradient Boosting - XGB Regressor (Adjusted R-Squared - R2 = 0.761 and Root Mean Squared Error - RMSE = 1914.80 kg/ha) achieved the best results in estimating biomass. The Support Vector Regressor (SVR) technique (Adjusted R2 = 0.773 and RMSE = 822.65 kg/ha) achieved the best results on estimating dry matter. To the best of our knowledge, this study is the first to consider cattle parameters, environmental and spectral data for estimating biomass and dry matter in grazing systems.