Counting of rice panicles using drone mounted RGB sensor and deep learning approaches
Abstract
Accurate estimation of rice panicle number is important for yield prediction and cultivar selection for genetic improvement. This study explored deep learning methods for rice panicle detection and counting using images acquired from drone mounted RGB sensors across 21 different rice genotypes with diverse morphological characteristics and yield attributes. Four image-based approaches i.e. GIS classification, Otsu thresholding, YOLOv8s and customized Roboflow model were used to detect and classify rice panicles. All four methods were finetuned to maximize precision and accuracy. Among all the models, Otsu thresholding achieved the highest overall accuracy (89.44%), RMSE of 178.06 and R^2 = 0.98, demonstrating its reliability and strong predictive ability in rice panicle detection and classification in field conditions. Similarly, the GIS classification, YOLOv8s and Roboflow customized model, performed with moderate predictive accuracy values of 81.96%, 80.48% and 80.87%. The rice varietal characteristics such as panicle structure and other plant attributes affected the models for panicle detection. UAV images coupled with deep learning models provides an effective, non-destructive tool for large-scale panicle estimation for high throughput phenotyping. Future studies should aim at model generalization by integration of multi-season and multi-location validation.