Forecasting Spring Wheat Maturity from UAV-Based Multispectral Imagery Using Machine and Deep Learning Models

Prabahar Ravichandran, Keshav D. Singh, Harpinder S. Randhawa, Shubham Subrot Panigrahi

Accurate forecasting of crop maturity supports efficient harvest planning and accelerates selection decisions in breeding programs. In spring wheat, maturity is typically assessed through manual scoring late in the season, which limits its usefulness for timely harvest management and early selection decisions in breeding programs. This study evaluated uncrewed aerial vehicle (UAV)–based multispectral imagery for forecasting maturity in spring wheat grown at Lethbridge, Alberta (AB), Canada, during the 2024 and 2025 growing seasons. Thirty cultivars were monitored using seven-band UAV multispectral imagery during grain filling, enabling derivation of core vegetation and senescence-related indices from radiometrically calibrated orthomosaics. Strong correlations (|r|>0.85) were observed between vegetation indices and days remaining to maturity (DRTM), motivating baseline regression models and subsequent evaluation of eleven machine-learning and deep-learning approaches. Among these, support vector regression (SVR) and multi-layer perceptron (MLP) achieved the highest predictive accuracy (R2=0.95–0.96; mean absolute error (MAE) ≈1.25 days). Deep learning models achieved performance comparable to machine-learning approaches; however, incorporating spatial information through convolutional neural networks did not improve prediction accuracy. Feature-attribution analysis identified the red, red-edge (RE), and near-infrared (NIR) spectral bands as key predictors, enabling non-destructive, early, and scalable UAV-based maturity forecasting.