MODELOS DE APRENDIZADO DE MÁQUINA PARA PREDIÇÃO DA PROTEÍNA BRUTA DE PASTOS DE CAPIM-TAMANI: DESENVOLVIMENTO, AVALIAÇÃO E INTERPRETABILIDADE COM A TÉCNICA SHAP

Abstract

ABSTRACT MONTEIRO, G.O.A. Machine learning models for predicting crude protein in Tamani grass pastures: development, evaluation, and interpretability with SHAP. 2025. Thesis (Ph.D.) – Faculty of Veterinary Medicine and Animal Science, Federal University of Mato Grosso do Sul, Campo Grande, MS, 2025.

Knowing forage quality is essential to meet animal nutritional requirements and optimize production. Traditional methods for assessing pasture nutritive value are destructive, costly, and time-consuming. Machine learning (ML) models have been increasingly applied to predictive tasks in agriculture. However, understanding how a model makes its predictions is just as important as the prediction itself. This study aimed to: (i) evaluate the applicability of machine learning models using tabular data, including climatic variables, light interception (LI), nitrogen (N) rates, grazing intervals (GI), and pre- (PREheight) and post-grazing (POSTheight) pasture heights to predict crude protein (CP) content in Tamani grass leaves; and (ii) use the SHAP technique to interpret how these variables influence the XGBoost model. A robust dataset was used, collected from an intensively managed Tamani grass pasture in the Brazilian Cerrado biome from October 2020 to April 2022. The dataset consisted of 90 instances, with 80% used for training and validation, and 20% for testing, applying 5-fold cross-validation with 100 iterations. Hyperparameters were tuned using grid search on the training set. The tested models included Linear Regression (LR), Multilayer Perceptron (MLP), Decision Trees (DT), Random Forest (RF), and XGBoost. The Shapley Additive Explanations (SHAP) technique was employed to understand the strength and direction of each variable’s contribution to the model's prediction. Performance metrics included Pearson correlation coefficient (r), coefficient of determination (R²), and mean absolute error (MAE). The MLP (r = 0.75, R² = 43.94%, MAE = 1.56), RF (r = 0.78, R² = 49.23%, MAE = 1.60), and XGBoost (r = 0.78, R² = 56.90%, MAE = 1.45) models showed the best prediction results (p < 0.001). The most important predictors of CP content were GI, followed by N rates, PREheight, and POSTheight. XGBoost outperformed the other models tested (p < 0.001). The impact of GI depended on interactions with other variables. Higher N rates had a positive effect on CP content in Tamani grass leaves. Rainfall between 100 and 180 mm positively influenced CP levels, while values outside this range had a negative impact. The application of 240 kg/ha/year of N reduced the influence of environmental conditions on CP content. In contrast, at 80 kg/ha/year of N, CP content was more strongly influenced by rainfall. The use of tabular data—such as N rates, GI, PREheight, POSTheight, LI, and climatic variables—proved to be a viable alternative for predicting CP using ML. The results underscore the importance of integrated management strategies, including pasture height targets, grazing management, nitrogen fertilization, and grazing intervals, as key tools for increasing crude protein in Tamani grass leaves.

Keywords: algorithms, machine learning, pasture management, Panicum maximum, precision livestock farming, Shapley Additive Explanations.