Predição da variabilidade espacial da produtividade agrícola com modelos ocultos de Markov

Abstract

The work developed in this Master’s Thesis is characterized as exploratory research using a case study based on data collected from one of Embrapa Pecuária Sul production areas, and problem-related literature review. The work is justified by the need to try to understand and predict land productivity over different times and seasons. The goal is to predict what might happen in a crop, using a hidden Markov model for probabilistic inference on historical data. The data were organized in state sequences, where each state represents a productivity result (the model hidden part) or data regarding conditions gathered from meteorological, soil, water balance, and other data (the model visible part). Model implementation was done using R software libraries. A comparison was made between models with real and simulated data. The results point to the need for a larger set of productivity data so that the model results are reliable. The model was adequate to predict yield throughout the crop, but the estimation of variability within a given area is more sensitive to input data availability and discretization.