Abstract:Aiming to address the core limitations of existing rice yield estimation methods, particularly the neglect of intra-species variability and the coupled influence of planting density and single-panicle weight, which compromise accuracy, a novel approach was proposed integrating an improved YOLO v8s-obb model with the NPRP-A method. The YOLO v8s-obb architecture was enhanced by incorporating the C2f_DCNv4 module, GSConv, EPSANet, and DAT attention mechanism to strengthen multi-scale detection of rice panicles. To ensure estimation reliability, ground-truth yield data were collected through field harvesting of sample plots at maturity. Gaussian kernel density estimation and the NPRP-A-based single-panicle weight modeling were further introduced to establish a nonlinear mapping between planting density and panicle weight, capturing their interactive effects. Experimental validation was conducted across three 1m2 plots. Results showed prediction errors consistently below 5.3%, with the lowest error at 2.2%, significantly outperforming traditional methods. This demonstrated the method’s high accuracy and robustness in real-world conditions. The proposed framework not only delivered a reliable technical solution for precise and efficient rice yield estimation but also advanced crop phenotyping and yield analysis in smart agriculture. By explicitly accounting for individual plant variation and density-yield interactions, the approach bridged a critical gap in current remote sensing-based estimation practices. Its design supported scalable deployment and offered practical value for precision farming applications, highlighting strong potential for broader adoption in agricultural monitoring systems.
