Abstract
Leaf nitrogen concentration (LNC) is a critical indicator for evaluating crop health and optimizing nitrogen management in sustainable agriculture. While multispectral and hyperspectral sensing techniques enable precise LNC estimation, their high cost and technical complexity often hinder practical application. This study assesses RGB imaging as a cost-effective and accessible alternative for estimating rice LNC across leaf, canopy, and plot scales. Field experiments conducted at two sites during the 2018-2019 reproductive stages acquired RGB images at three spatial resolutions. For canopy and plot images, rice vegetation was isolated using green minus red (GMR) band indices and thresholding. Stepwise multiple linear regression (SMLR) models incorporating 13 color indices were developed. Results demonstrated that leaf-scale models achieved superior accuracy (R(2) = 0.84-0.87, RMSE = 0.16-0.25%), validating RGB imaging's potential for high-precision diagnostics. At the canopy scale, vegetation segmentation enhanced model performance (an average R(2) increase of 3% compared to those from unsegmented images), confirming the necessity of background removal. Plot-scale analysis revealed that UAV flight altitude minimally affected model accuracy within the range tested, with 100 m yielding comparable performance (R(2) = 0.61-0.65) to other altitudes. Cross-site validation indicated promising generalizability at the leaf scale, while canopy and plot scale models exhibited greater sensitivity to environmental variations. This research establishes RGB imaging as a scalable tool for rice nitrogen monitoring, demonstrating that segmentation improves accuracy at larger spatial scales. These findings provide practical insights for implementing precision nitrogen management in smallholder farming systems, supporting ecological sustainability through reduced fertilizer overuse.