Estimation of rice crop production using Sentinel-2 multispectral images through a ResNet18

Abstract

Global food security depends on accurate agricultural monitoring systems, with rice being one of the most critical crops. However, estimating its production faces significant challenges in regions where detailed parcel-level information is scarce. This study develops and validates a reproducible methodology to estimate rice production in Rio Grande do Sul, Brazil, by integrating multispectral imagery from the Sentinel-2 mission and deep convolutional neural networks (CNNs). The implemented automated workflow encompasses the download of satellite products, radiometric preprocessing, and mosaic segmentation corresponding to two key phenological stages: reproductive and maturation. For the analysis, a ResNet18 architecture was adapted by modifying its input layer to process 10 spectral channels, and the model was trained using 2,304 patches of 64 × 64 pixels. Training labels were generated proportionally to official yield data from IBGE and CONAB, allowing the model to learn the relationship between spectral response and actual productivity at the municipal level. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²), achieving R² values close to 0.91. These results reflect significantly higher accuracy compared with traditional methods based exclusively on vegetation indices such as NDVI. It is concluded that the proposed methodology is highly reproducible and scalable, facilitating its application to other geographical regions and crop types. The contributions of this research represent a concrete advance toward the democratization of precision agriculture through the use of open-source software and freely available satellite data.