Brazilian coffee yield forecasting by machine learning

Abstract

Brazil is the largest producer and exporter of coffee in the world, with climatic variability being the main factor responsible for the oscillations and crop frustrations. Predicted productivity forecasts work efficiently in decision making at all levels of the production chain, since they allow estimating the level of investment and strategies to be made in the crops. The agricultural system is an example of an extremely complex multivariate system, and Machine Learning (ML) is at the forefront of modeling complex systems around the world. ML associated with "big data" technologies and high-performance computing are creating a growing wave of agronomic applications in the last decades, clearly exemplified by the growing number of international scientific papers. MLs are already used by government agencies with success in some countries of the world for productivity forecasting. MLs are not based on physical and biological relationships between variables and, because of this, in the 90s these were considered, together with the lack of computational power, the main limitation of the use of these models. However, ML models are currently considered to fill theoretical gaps, the lack of specific data, especially in field conditions, and generalizes agricultural problems very efficiently, which makes it possible to make estimates and forecasts. Several studies are found in the literature on coffee productivity estimates. However, prediction problems (future) are mainly statistical, and the literature in the agronomic area in Brazil and in the world is scarce. The objective of this work is to forecast productivity in advance, using MLs, at least 6 months before the start of canephora and arabica coffee harvesting for different spatial scales: 1) National, 2) State and 3) Regional regions in Brazil). The predictions will be made from decennial climatic variables before harvest and referring to that productive cycle: air temperature, rainfall, global solar irradiance, wind speed, radiation balance and relative humidity from grid data of the NASA-POWER system .Data from water balance components will also be used, such as: reference and actual evapotranspiration, soil water storage, water deficiency, and surplus. All data used will be from 2002 to 2018 and a selection of the most important meteorological data for each region will be made by main component analysis. The MLs to be tested will be 'Multilayer Perceptron', 'Support Vector Machine' and Decision Trees', with various topologies, activation functions, learning rates and variables in the input layer. The first part of the work will be done with data from the coffee productivity field, which will be provided by the COOXUPÉ cooperative for the cerrado and south of Minas Gerais (20 locations), which will indicate the choice of topologies and types of MLs to be applied to other states. In the second stage of the work, official productivity data from PAM-SIDRA / IBGE will be used to train MLs in the country, state and coffee-growing regions. The MLs and topologies adjusted in the first step will initially be used in this second step but may require new parameterizations. It is hoped that this work will bring a robust method for early forecasting of coffee productivity for Brazil from different climatic conditions. (AU)