Rice yield prediction in unseen years at field level with high-resolution gross primary productivity derived from Sentinel-2 imagery

Accurate field-level rice yield prediction for an unseen year is valuable for optimizing precision farming practices and strengthening national food security frameworks. Although many studies have used vegetation indices or gross primary productivity (GPP) to predict crop yield, few have systematically evaluated their differences in predictive performance and stability using time series satellite imagery across the entire growing season. Simultaneously, little research has focused on field-level prediction for unseen years over large regions. To address these issues, we conducted an in-depth comparison between the Sentinel-2-derived normalized difference red edge index (NDRE) and high-resolution GPP generated via a modified two-leaf light use efficiency model in their correlations with rice yield. The optimal time window for yield prediction was identified using original and harmonic fitted GPP data at 10-day intervals. Additionally, cross-year GPP correction (CGC) was proposed as an efficient approach for model transfer to unseen years and compared with that of the adversarial discriminative domain adaptation (ADDA), an emerging data-driven domain transfer learning algorithm. Specifically, these methods were assessed with an extensive field-level rice yield dataset from eastern and northeastern China spanning 2019–2022.

We found that GPP outperformed NDRE in predicting rice yield (individual monthly: ∆r² = 0.04–0.29, cumulative monthly: ∆r² = 0.22–0.41), with greater stability and reliability. Furthermore, the harmonic fitted GPP could improve the yield prediction accuracy. Additionally, the CGC method improved interannual prediction accuracy (R² = 0.55–0.73) for the two regions, showing better predictive performance than the ADDA model (R² = 0.54–0.62). The proposed method relied only on a limited amount of ground-truth yield samples and exhibited robust performance in years characterized by pronounced interannual yield variability (2019) or extreme weather conditions (2022). This research has great potential for implementing rice yield prediction over large regions with publicly available imagery and limited ground-truth yield data, particularly for smallholder farming systems in the context of precision crop management and food security assessment.