HIDYM: A high-resolution gross primary productivity and dynamic harvest index based crop yield mapper

Reliable prediction of field-level crop yield over large regions is a prerequisite for informed decision-making in precision crop management. One of common Earth observation approaches is to predict crop yield through the estimation of gross primary productivity (GPP) and a fixed crop-specific harvest index (HI), but few studies have considered the spatio-temporal dynamics of HI. Although some studies have used two-leaf light use efficiency (TL-LUE) models to reduce GPP estimation uncertainties by distinguishing sunlit and shaded leaves, it remains unclear about the physical mechanism underlying the incorporation of environmental regulations into TL-LUE. This study proposed a high-resolution GPP and dynamic HI based yield mapper (HIDYM), which incorporated the generation of 10-m resolution GPP product via a modified TL-LUE (mTL-LUE) model and the estimation of dynamic HI from Sentinel-2 imagery. The mTL-LUE was developed to account for the effect of environmental factors on GPP. Dynamic HI was estimated per pixel and per year by combining the phenological difference ratio and tasseled cap transformation of Sentinel-2 imagery at three critical stages of crop growth. The results demonstrated that HIDYM could capture the spatial and interannual variations of field-level rice and winter wheat yields. The improvement of HIDYM over the fixed HI strategy was more pronounced for rice (R2: 0.64–0.72 vs 0.34–0.48 for 2019–2022) than for winter wheat (R2: 0.72 vs 0.66 for 2021–2022 and 0.71 vs 0.57 for 2022–2023). The proposed methodology has great potential for the routine prediction of crop yields over large-scale croplands, especially in smallholder farming systems. •A HIDYM methodology is proposed for rice and winter wheat yields prediction.•The mTL-LUE model outperformed TL-LUE in GPP estimation against flux site GPP data.•Dynamic harvest index (HId) is estimated per pixel for multiple years and two crops.•Integration of HR-GPP and HId leads to improved prediction and model transferability.•10-m crop yield maps can capture variability effectively across smallholder fields.