Impact of atmospheric corrections on satellite imagery for corn yield prediction using machine learning

Satellite images, combined with machine learning algorithms, are widely used to monitor and estimate yield in various crops. Despite the extensive use of remote sensing images in agricultural monitoring, there remains a gap in the understanding of how atmospheric corrections can improve data accuracy. This study aims to analyze the influence of different atmospheric correction techniques - Dark Object Subtraction (DOS), Sentinel-2 Correction (Sen2Cor), Image Correction for Atmospheric (iCOR), and L1C data from Top of Atmosphere (TOA) radiation - on corn fields, and their impact on yield estimation in both irrigated and rainfed fields using machine learning algorithms (Random Forest, k-Nearest Neighbors, and Support Vector Machine). When atmospheric corrections were applied to each algorithm, no significant differences were found among them. However, the performance of the models differed for corn yield estimation. The SVM algorithm showed the lowest performance during the main crop season (R² = 0.36), while both RF and kNN yielded prediction results with an accuracy of over 55%, with RF providing the highest R² values and the lowest errors (RMSE = 0.3 t ha −1 ). In contrast, during the second crop season, the precision and accuracy of the estimates increased, particularly for SVM (R² = 0.7). RF remained the best-performing model, with R² values exceeding 0.80 and errors below 0.20 t ha −1 .