Affordable, practical, and robust biomass estimation in rice with minimal equipment using deep learning‐based image segmentation

Abstract

Plant biomass measurements are critical for crop improvement and breeding programs. Such measurements rely on destructive and labor‐intensive methods that limit continuous measurements of biomass over plant growth stages and time. This study aimed to develop an inexpensive, nondestructive method for estimating aboveground biomass in rice ( Oryza sativa L.) plants to help researchers monitor biomass accumulation. A deep learning approach using convolutional neural networks (CNNs) was applied to side‐view RGB (red, green, blue) digital color images for estimating shoot biomass. The study was conducted using a diverse population of rice genotypes grown under greenhouse conditions, and side‐view RGB images were captured using a standard digital camera with a uniform background and a reference object for scale calibration, requiring no specialized equipment. The R package “imageseg” was utilized to perform CNN‐based image segmentation, estimating leaf area as a proxy for biomass. The results showed high consistency between leaf area estimations from two different images of each plant and a strong positive correlation ( R ²  = 0.763) between estimated leaf area and measured dry biomass. The method demonstrated robustness under ambient lighting conditions and without precise control of plant orientation. In conclusion, this inexpensive and practical CNN‐based nondestructive imaging approach provides an accurate and affordable tool for biomass estimation in rice, which could be valuable for phenotyping workflows and studying growth dynamics in various crops under real‐world conditions without the need for specialized cameras or equipment.