Satellite imagery and machine learning for sugarcane yield estimation in regions of São Paulo State

Abstract

The objective of this project is to develop sugarcane yield prediction systems in regions of São Paulo state, based on the main drivers of crop production (soil, climate and sugarcane variety) and data from remote sensing using machine learning techniques. Remote sensing and agrometeorological data will be used to compare two methods of sugarcane yield estimation in regional scale: i) agrometeorological-spectral empirical model using machine learning algorithms, ii) a coupling model with remote sensing data and crop growth models. For this, we will use Sentinel-2 satellite imagery (spectral variables: bands and vegetation indices) and agronomic and climatic information. Agronomic data will be obtained in sugarcane production plots in SP regions referenced with variety, soil type, harvest data, production environment, etc. Moreover, climatic data such as precipitation, temperature and radiation will also be included as models' variables. The empirical models will be created in regional scale and based on agrometeorological and spectral variables of each study area, using machine learning algorithms. To improve the yield estimation at regional scale, crop growth indicators derived from remote sensing data (e.g., leaf area index - LAI) will be used, coupling spectral information in models of crop growth and the results will be compared with the empirical models. The creation of sugarcane yield models using satellite data enables the spatiotemporal crop monitoring and, it is relevant for the decision makers involved in the sugar-energy sector, such as producers, companies, mills, and banks, supporting the production of biomass and biofuels as well as helping in public policies related to climate change. (AU)