Predição da produtividade de cana-de-açúcar a partir de parâmetros biométricos e imagens orbitais do sensor MSI/sentinel-2

Abstract

During the entire sugarcane harvest, crop monitoring is essential to monitor the results of the management adopted, identify problems that can cause low yields, and establish logistical planning according to the production of each cultivated area. To meet this demand, field techniques are employed to measure the characteristics of the crop and estimate the productivity of the plots in advance. However, most practices currently employed estimate only average productivity, disregarding its spatial variability. In this sense, the present work had as objective to generate a model of multiple linear regression that allows to estimate the productivity of the sugar cane five months before the harvest, from remote sensing data and biometric data obtained in the field. For the elaboration of this work, the biometric data used were the average stem diameter (DMC), the number of industrializable tillers (NPI), average height of the stems (EMC) and Ton of Sugarcane per Hectares (TCH); orbital images from the Sentinel 2 satellite's MSI sensor on the date of the biometrics; and the productivity data obtained by the productivity monitor of the John Deere CH670 harvester. 16 scenarios were created with different combinations of biometric data and vegetation indices, and a scenario based exclusively on remote sensing data. To obtain the multiple linear regression models, the WEKA software was used. Through the correlation and occurrence analysis, the EMC and NPI biometric parameters and the Modified Chlorophyll Absorption Ratio Index (MCARI1) and Canopy Chlorophyll Content Index (CCCI) indices, showed a higher correlation with productivity. In cross-validation, the scenario using the DMC, NPI, EMC and CCCI parameters, presented the lowest RMSE, with a value of 3.62 ton.ha-1 and an adjusted R² of 0.72. Due to the incompatibility of scale between the biometric data and the SR data, only the models that do not use biometric data in their variables were applied spatially. The best scenario with SR data alone obtained an RMSE of 8.15 ton.ha-1 in the validation data.