From plots to commercial fields: scalable, transferable cotton morphology and productivity estimation using...

Product demand and climate variability are progressively increasing the need for real-time, scalable crop monitoring to support varietal selection and in-season input optimisation. However, producers still have limited information on the temporal and spatial variability of cotton health and performance beyond point-scale field surveying. In addition, given cotton's high phenotypic plasticity, near real-time derived metrics are essential to improve input efficiency and strengthen long-term sustainability of the cotton industry in Australia. Therefore, we proposed a functional integrated predictive sensing framework to estimate and predict cotton canopy morphological (i.e., height) and productivity traits (i.e., dry matter and lint yield) across large plots (12m × 6m). Scalability was validated by applying the proposed framework to estimate and map cotton yield across commercial fields. To do this, we explored the accuracy of high-resolution multispectral imagery from two platforms (unmanned aerial vehicle (UAV) and PlanetScope (PS)) collected across two 144-plot trials for two cotton seasons. These were designed with a large range in nitrogen rates (N), shading, and two growth-regulator doses, thus, creating variable environments. Sensing metrics were obtained from UAV imagery (1.3-1.6 cm pixel size) acquired once in 2022/23 and eight times in 2023/24, while PS composites (3 m pixel size) provided near-daily coverage in both seasons. Time-series gaps were imputed using Savitzky-Golay smoothing in thermal time (GDD), enabling extraction of growth dynamic metrics (GDMs) as single-date (SD; e.g., peak canopy) and multi-date (MD; e.g., daily average growth rate) metrics. After reducing collinearity and dimensionality, random forest (RF), support vector regression (SVR), and gaussian process regression (GPR) were trained and interpreted with SHAP, for feature contribution. UAV single-date models (SD_ UAV) achieved strong accuracy for height (R² = 0.77), biomass (R² = 0.73), and yield (R² = 0.81). Incorporating UAV time-series metrics (MD_UAV) improved the performance R² = 0.87, 0.86, and 0.85 for height, biomass and yield, respectively. Application of the derived models using high resolution satellite data (MD_PS) for different farming systems showed highly significant accuracy (R² = 0.67) to predict cotton yield at aggregated field scale. As such, enabling the detailed spatial prediction of cotton yield within a field. It is anticipated that the proposed functional sensing framework will improve the estimation of key cotton production traits, supporting field- and within-field decision-making, ultimately contributing to more resilient and sustainable cotton production in Australia.