Suporte à Decisão em Pastagens: Análise Espaço-temporal e Aprendizado de Máquina para Predição da Disponibilidade de Forragem no Contexto de Smart Farming

Abstract

In extensive herding, it is important to know the amount of food available in the pasture so that the ideal amount of animals that occupy that area can be estimated. In Empresa Brasileira de Pesquisa Agropecuária - Pecuária Sul, the forage mass available in certain experimental areas is estimated by a direct method, where cages are used to isolate a particular area of the pasture and, every month, the vegetation of this area is cut, dried and then weighed. Thus, the dry mass of the collected sample is obtained, which is used to estimate the pasture accumulation rate for the subsequent month. This work proposes a computational model based on Long Short-Term Memory Artificial Neural Networks with the aim of improving the estimation of pasture availability currently done at Empresa Brasileira de Pesquisa Agropecuária - Pecuária Sul, based on data from pasture availability, collected through the direct method, and meteorological. The methodology is based on exploratory and experimental interdisciplinary research, and its development, carried out in the form of a case study, pointed out some problems in the pasture availability sampling process that had been carried out. In order to circumvent the problems identified, an architecture was defined where the pasture availability and meteorological data are persisted in a geographic database, after an Extract, Transform, and Load process. After being loaded into the geographic database, the data goes through a pre-processing that prepares them for the training process of the Long Short-Term Memory Artificial Neural Network developed. The results show that the proposed methodology is capable of estimating with significant accuracy the availability of instantaneous pasture and the subsequent month, since the average error verified is of the same magnitude of the error associated to the sampling method used.