ASABE Technical Library - Abstract
Member and Access Notice
Computer Vision-Based Early Red Rice Detection System Development for Mechanized Row Rice Farming
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2026 ASABE Annual International Meeting 2600848.(doi:10.13031/aim.202600848)Authors: Md Nurul Azmir, Cengiz Koparan, Nilda Roma Roma, Benjamin Runkle, Dongyi Wang
Keywords: Agricultural Automation, Computer vision, Multispectral imaging, Rice, Weed Management
Abstract.
Red rice (Oryza sativa f. spontanea) is one of the most problematic weeds in rice production systems due to its close genetic and morphological similarity to cultivated rice, which severely limits the effectiveness of manual scouting and selective herbicide application. Red rice competes aggressively for nutrients, water, and light, reduces grain yield and quality, and leads to varietal contamination that can persist in the seed bank for multiple growing seasons. Early-stage detection of red rice is critical for precision weed management, because control options become increasingly limited due to wet field conditions as well as crops and weed canopies overlap at later stages. Limited red rice control options have created a strong need for autonomous sensing and vision-based detection systems capable of identifying red rice during early phenological stages, when targeted intervention is most effective and yield losses could be minimized. The objective of this research was to evaluate spectral differences between red rice and cultivated rice using multispectral imagery in red (660 nm), green (550 nm), and near-infrared (NIR, 850 nm) bands to facilitate automated computer vision-based red rice detection. Images were collected inside a dark room from 1 m above the canopy across 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28 and 30 days after emergence (DAE), corresponding to the critical pre-flooding period during which red rice must be identified and controlled so that earlier field flooding can be implemented to suppress weeds before conventional management options become limited. A Python-based image-processing workflow was developed to segment foliage using binary thresholding method and extract mean Digital Number (DN) values for each spectral channel. Green band DN values were consistently higher for red rice than for cultivated rice across all observation days (6–30 DAE), the separation was most visually distinct and reliable at 22, 24, and 28 DAE. A paired-samples t test confirmed that the difference between the red rice and cultivated rice green DN number was statistically significant, t(64) = −5.39, p < .001. DN values in NIR band from red rice were statistically significant than DN values from rice at 10, 18, 22, and 24 DAE, indicating DN values in NIR band between 18,20,22 DAE as the most significant phenological window for spectral discrimination. DN values in the green and NIR bands are effective indicators for differentiating red rice from cultivated rice and provide a reliable signal input for automated red rice detection systems for mechanized row rice farming in Arkansas, USA.
(Download PDF) (Export to EndNotes)