Answers are grounded in this paper and cite their sources.
Precise crop yield estimation with remote sensing ensures global food security. Although light use efficiency (LUE) models and machine learning (ML) approaches are widely employed, they face notable limitations: LUE models rely on uncertain manual parameter calibration, and ML models often use indirect remote sensing features and require large training datasets, limiting their generalizability and interpretability. To address these limitations, we developed four novel growth kinetics indexes (GKIs) that bridge LUE and ML models for rice yield estimation. Based on these indexes, a hybrid framework termed LUE-GKML was established, consisting of three components: (1) GPP simulation using an LUE model; (2) derivation of GKIs, including absolute growth momentum (AGM), relative growth momentum (RGM), absolute growth kinetic energy (AGE), and relative growth kinetic energy (RGE); (3) a synthetic minority over-sampling technique (SMOTE)-based augmentation strategy to alleviate data scarcity. Validated in Yujiang County, Jiangxi Province, China, LUE-GKML outperformed both standalone LUE and ML models, increasing R 2 by at least 0.33 and reducing RMSE by more than 177.14 kg/ha. The GKIs were proven to be more closely related to yield than traditional parameters like EVI and LAI. Moreover, SMOTE-based augmentation substantially improved generalization performance, with optimal results typically achieved at multiplication ratios of two to five. Shapley additive explanations (SHAP) further revealed that AGM during the seedling stage, relative growth rate (RGR) during the heading–filling stage, and both GPP and AGM during the maturity stage were key to accurate yield estimation. This work highlights the potential of domain knowledge to guide model coupling, offering a scalable solution for crop yield mapping.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: