Enhancing crop yield predictions through spectral-temporal integration across multiple agricultural fields

Global food security demands accurate crop yield prediction systems that can adapt to spatiotemporal variability across diverse agricultural landscapes. Current satellite-based machine learning (ML) approaches face critical limitations when applied across multiple fields: dependency on specific growth stage timing, vulnerability to cloud interference, and inability to capture pixel-level temporal heterogeneity in crop development. These constraints reduce model generalizability and prediction accuracy, particularly when maximum growth periods vary spatially within and between fields. To address these fundamental challenges, we developed a spectral-temporal integration strategy that leverages the extreme values (maximum and minimum) of vegetation indices throughout the growing season rather than fixed-date observations. We evaluated this approach using 24 Sentinel-2 images across the maize growing season and yield data from three fields with distinct environmental characteristics: field A (20.6 ha), B (32.5 ha), and C (20.9 ha). The proposed MaxMin strategy consistently outperformed conventional date-specific modeling approaches. Root means square error (RMSE) decreased to 16.5, 26.5, and 24.6 bu.ac −1 (1 bu ≈ 62.77 lb ≈ 27.2 kg) for fields A, B and C, respectively, compared to the best single-date predictions of 26.3, 18.1 and 28.4 bu.ac −1 . Correlation coefficients improved by 0.01–0.03, reaching r values of 0.86, 0.93, and 0.86. The normalized RMSE reductions translated to 4–6% improvements in mean absolute error across fields. Feature importance analysis revealed that green normalized difference vegetation index (GNDVI) and green chlorophyll index (GCI) dominated predictions across all sites, with importance scores ranging from 0.545 to 0.940. Critically, the MaxMin strategy maintained high accuracy without requiring precise identification of peak growth dates, addressing a fundamental challenge in operational yield forecasting systems. This temporal flexibility enables robust predictions across fields with varying soil properties, climatic conditions, and management practices. Our findings demonstrate that integrating seasonal spectral extremes provides a practical and scalable solution for crop yield prediction systems with strong cross-field generalizability, which are essential for precision agriculture and food security planning.