Work

Winter wheat plant density determination: Robust predictions across varied agronomic conditions using multiscale RGB imaging

Year: 2025

Type: article

Abstract: • High-throughput RGB imaging enables wheat density estimation across two years. • Ridge models excel over PCR in handling Vegetation Index (VI) multicollinearity. • RGB derived VIs are reliable predictors when vegetation coverage is abundant. • Best prediction (17.91% MAPE) achieved with the lowest resolution (1.18 cm pixel-1). • Robust density estimation across Haun's stages 4.4 – 8.4 regardless of resolution. Cereal plant density is a crucial agronomic factor affecting resource management and yield. This study automated wheat density estimation using multiscale imaging from ground and Unmanned Aerial Vehicles (UAV) at 15, 30, and 50m Conducted over two agronomic seasons (2022 and 2023) with different water profiles, it analyzed three wheat genotypes (cv. Bologna, Hondia, and Marcopolo) sown at five densities ranging from 35 to 560 seeds m -2 . Images collected through RGB sensors across Haun's developmental stages 2.6 – 12.2 provided data for calculating 15 Vegetation Indexes (VIs), which, along with their Principal Components (PCs), were used as inputs for Ridge and Principal Component Regression (PCR) models. Training was conducted on the 2022 datasets using 4-fold, 10-repeated cross-validation to determine the most predictive growth stages, with Haun stages 5.3 to 7.3 yielding the best results, irrespective of resolution. Testing on 2023 datasets showed that Ridge models consistently outperformed PCR, especially for medium to high-density ranges (140–560 seeds m -2 ), though they underperformed at lower densities, leading to their exclusion from the testing data. The top-performing Ridge model, trained on Haun stages 7.1–7.3 at 50 m (1.18 cm pixel-1), achieved Mean Absolute Percentage Error (MAPE) 17.91% – 28.54% (0.9 – 0.68 R 2 ) values across various test sets, with stable performance throughout resolutions and stages (4.4 – 4.8). These findings show robust prediction capabilities across a broader developmental range and from the lowest resolution recorded, especially when vegetation coverage is abundant. The study highlights the practicality of high-throughput RGB imaging for scalable, flexible and affordable plant density estimation.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2025.100921

Authors: Jara Jauregui-Besó, Adrian Gracia-Romero, Constanza S. Carrera, Marta da Silva Lopes, José Luis Araus, Shawn Carlisle Kefauver

Institutions: Agrotecnio - Centre for Food and Agriculture Research, Universitat de Barcelona, Institute of Agrifood Research and Technology, Universitat de Lleida

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $1,770