Spectral image processing and deep learning for rice crop identification

Abstract

Accurate monitoring of agricultural cycles is essential to ensure food security and optimize water resource management in cereal-producing regions. This research addresses the challenge of automated vegetation cover classification through the design and validation of an advanced rice crop identification model based on deep learning and the processing of multispectral satellite imagery. The methodology employed a convolutional neural network architecture, U-Net, optimized for the semantic segmentation of geospatial data obtained from the Sentinel-2 mission. Input tensors composed of four strategic spectral bands were processed, enabling the network to generate highly accurate binary masks associated exclusively with rice production areas. The study incorporated a comparative phase in which several hyperparameter configurations were evaluated to determine the architecture with the highest convergence. Experimental results showed outstanding model performance during the training phase, achieving an accuracy of 92.3 %. Furthermore, the robustness of the segmentation was validated using the Intersection over Union (IoU) metric with 86 % and a Dice coefficient of 89 %, values that exceed conventional thresholds for the identification of complex crops. It is concluded that the integration of U-Net architectures with openly accessible multispectral data constitutes a viable and scalable technical tool for precision agricultural remote sensing. These findings support proposing the developed model as an efficient solution for large-scale crop inventory, providing support for governmental decision-making and crop yield prediction in variable climatic environments.