Timely and accurate mapping of wheat cultivation is vital for food security, agricultural policy, and precision farming. Despite the widespread use of Sentinel-2 imagery and machine learning for crop mapping, existing studies often rely on single-feature sets or coarse spatial resolutions and provide limited evaluation in smallholder and fragmented agricultural systems. This study presents a robust machine learning-based framework for wheat classification using Sentinel-2 imagery and the Extreme Gradient Boosting (XGBoost) algorithm. Three input feature models were designed to assess classification performance: Model A used only spectral bands, Model B relied solely on vegetation indices (NDVI, GNDVI, NDRE, MSAVI), and Model C combined both feature types. The models were trained and evaluated over Vehari District, Pakistan, during the 2023–2024 wheat season using high-quality field samples. Model C achieved the highest accuracy (OA = 94.31%, Kappa = 0.886), outperforming Models A and B, due to the synergistic effect of spectral and index features. Visual inspection and spatial overlays with wheat/non-wheat ground truth polygons further confirmed the superior precision of Model C, especially in heterogeneous field conditions. Spectral analysis and phenological profiling revealed that red-edge and NIR bands, along with indices like NDRE and GNDVI, significantly enhanced separability. The study highlights the operational value of combining reflectance and vegetation indices to improve classification accuracy in smallholder wheat systems. This research provides a robust, scalable, and interpretable framework for crop monitoring, showing that feature fusion enhances classification accuracy and spatial precision. The methodology directly supports the advancement of the Sustainable Development Goals (SDGs) by enabling data-driven agricultural management (SDG 2: Zero Hunger) and promoting sustainable land-use practices.
Haseeb et al. (Sun,) studied this question.