Impacto da correção atmosférica em imagens de satélite na previsão de produtividade utilizando aprendizado de máquina

Abstract

The cultivation of corn (Zea mays L.) in Brazil is crucial for food security, and precision agriculture, which uses remote sensing, can optimize agricultural management. However, the use of orbital images presents several challenges, including the availability of images with the absence of clouds and the interference of the atmosphere in the radiation captured by the sensors. Atmospheric correction methods have been developed by several authors, however the impact of the methods on continuous monitoring and the impact on advanced data processing techniques is still poorly studied. This study evaluates three atmospheric correction methods, Dark Object Subtraction (DOS), Image Correction for Atmospheric Effects (iCOR), and Sentinel-2 Correction (Sen2Cor), as well as images with L1C processing that are data referring to radiation at the Top of the Atmosphere (TOA) to improve the quality of free orbital imagery data from the Sentinel-2 satellite in the characterization of the corn cycle and yield prediction using vegetation indices. The analysis showed that the iCOR correction is the most effective method in characterizing the corn cycle when using NDVI and GNDVI. For yield prediction, the DOS and Sen2Cor corrections were the most effective in all the algorithms tested, Random Forest (RF), k-Nearest Neighbors (kNN), and Support Vector Machine (SVM), with RF presenting the best performance, especially in the off-season, with an average R2 of 0.82. The proper choice of atmospheric correction method and vegetation index is essential to increase accuracy in monitoring and forecasting productivity, contributing to more efficient and sustainable agricultural management.