A new sensing framework using drone and satellite imagery can predict cotton height, biomass and lint yield with high accuracy, offering growers a scalable way to monitor crop performance in near real time. By combining growth-curve metrics with machine learning, the approach moves beyond point-scale field surveying to deliver field-wide and within-field estimates, supporting more precise decisions on nitrogen, irrigation and growth regulators.
Cotton is notoriously difficult to monitor. Its indeterminate growth habit means vegetative and reproductive phases overlap, and environmental stresses such as heat or insect pressure can trigger fruit shedding, shifting resources to vegetative growth and reducing yield. Current monitoring relies on manual sampling at two to three locations per field, which is labour intensive and fails to capture spatial variability. Because of these challenges, there is a need for non-destructive, scalable tools that can track cotton canopy dynamics and productivity across entire fields and seasons.
A study led by Francesca Devoto and colleagues at The University of Queensland (UQ), with collaborators from Cotton Seed Distributors Ltd. (CSD), Wageningen University, CSIRO and the Terrestrial Ecosystem Research Network (TERN), was published (DOI: 10.1016/j.plaphe.2026.100220) in Plant Phenomics on 8 May 2026. The research tested whether growth dynamic metrics (GDMs) derived from multispectral unmanned aerial vehicle (UAV) and PlanetScope (PS) satellite imagery could predict cotton morphological and productivity traits, and whether those models could scale from experimental plots to commercial fields.
The team collected data across two cotton seasons (2022/23 and 2023/24) from 144-plot trials at UQ's Gatton Campus, designed with varying nitrogen rates, shading and growth-regulator doses. UAV imagery was acquired once in the first season and eight times in the second, while PS composites provided near-daily coverage. Time-series gaps were filled using Savitzky–Golay smoothing in thermal time (growing degree days, GDD), enabling extraction of single-date (SD) and multi-date (MD) metrics such as peak canopy, daily growth rate and area under the curve. After dimensionality reduction, random forest (RF), support vector regression (SVR) and Gaussian process regression (GPR) models were trained and interpreted with SHAP. UAV single-date models achieved R² = 0.77 for height, 0.73 for biomass and 0.81 for yield. Incorporating time-series metrics improved performance to R² = 0.87, 0.86 and 0.85 respectively. The best model was then transferred to eight commercial fields (30–60 ha) provided by CSD, where it predicted yield at field scale with R² = 0.67.
The authors said the key advance lies in fitting growth curves in thermal time rather than calendar days, which aligns phenological stages across seasons and environments. They explained that early-season growth-rate metrics capture crop vigour and stress responsiveness, while GDD-based area under the curve acts as a proxy for cumulative photosynthetic capacity. The framework enabled trait prediction as early as 38 days after sowing, with about 75% of variance explained, and accuracy stabilised from peak canopy onward. They added that combining drone and satellite data offers a practical pathway for industry deployment, though satellite-based yield estimates were less precise than drone-based ones due to coarser resolution.
The framework could support variable-rate application of nitrogen, irrigation and growth regulators, with early-season indicators informing fertiliser scheduling and post-flowering metrics guiding growth-regulator decisions. It may also give breeders scalable data for selecting stress-resilient genotypes. The authors noted that broader application across diverse production environments will be needed to characterise yield variability more fully, and that advances in Earth observation and AI-driven processing could streamline deployment at farm and regional scales.
