Accurately predicting spring wheat (Triticum aestivum L .) yield over the growing season is essential for optimizing management practices, improving economic returns, and supporting sustainable production. Predicting potential yield in spring wheat is challenging due to complex and nonlinear interactions among environmental, physiological, and management factors, including temperature, soil moisture, and nitrogen dynamics. While modern remote sensing solutions combined with conventional machine learning (ML) models have advanced yield prediction by integrating multispectral and temporal information, they often fail to capture the complex genotype × environment interactions that drive yield variability. Most of the traditional ML models, random forest (RF) or support vector machine (SVM), rely on limited feature transformation and shallow nonlinear mapping, which restrict their ability to generalize across heterogeneous field conditions. Consequently, their predictive performance tends to degrade when exposed to spatially or temporally diverse data. To overcome these challenges, a deep learning (DL)-based feed-forward neural network (FFNN) was developed in this study. The FFNN architecture enables successive nonlinear transformations across multiple hidden layers, allowing it to learn smooth, high-order relationships among spectral, environmental, and yield features. This approach provides a more flexible and robust framework for modeling continuous and complex yield dynamics compared to conventional ML techniques. The FFNN was benchmarked against two widely used ML models, random forest (RF), and support vector machine (SVM), to evaluate differences in nonlinear feature representation and predictive performance. Multispectral unmanned aerial system (UAS) imagery was collected across three critical phenological stages of spring wheat within the Data Intensive Farm Management (DIFM) framework on a commercial farm. Twenty-three vegetation indices (VIs) and five spectral bands were extracted as predictors, while yield monitor data were spatially interpolated using empirical Bayesian Kriging (EBK), inverse distance weighting (IDW), and ordinary Kriging (KG). A comprehensive evaluation showed that KG generated the most accurate yield surface, effectively capturing within-field heterogeneity. Although site-specific, the FFNN achieved superior accuracy (R² = 0.69) and demonstrated enhanced capacity to model nonlinear spectral-yield relationships compared to RF and SVM. These findings highlight the potential of integrating geostatistical interpolation with DL models for interpretable, scalable, and high-resolution yield prediction in spring wheat production systems.
Villamil-Mahecha et al. (Sun,) studied this question.