Understanding rice yield gaps with crop modeling and machine learning in a long-term continuous cropping experiment

Yamaguchi, Tomoaki; Radanielson, Ando Mariot; Angeles, Olivyn; Katsura, Keisuke; Saito, Kazuki

Understanding rice yield gaps with crop modeling and machine learning in a long-term continuous cropping experiment

Autores

Yamaguchi, Tomoaki

Radanielson, Ando Mariot

Angeles, Olivyn

Katsura, Keisuke

Saito, Kazuki

Editor

Elsevier

Descripción

"Context: Long-term experiments provide opportunities to develop strategies for sustaining rice productivity, improving resource use efficiency, and adapting to abiotic and biotic stresses. Yet yield gap approaches have rarely been applied to such datasets to disentangle the drivers of long-term productivity variation. Objective: We analyzed five decades (1971–2017) of rice yields from the Long-Term Continuous Cropping Experiment (LTCCE) at IRRI to quantify yield gaps, identify their causes across the dry, early wet, and late wet seasons, and model yields to simulate yield gaps under scenarios in which specific constraints were removed. Methods: Potential yields were estimated using the process-based model ORYZA v3, and yield gaps were quantified based on the potential yield and actual yields. Machine learning (ML) with Random Forest and SHAP-based interpretation was used to identify the determinants of yield gap variation in each season. Additionally, MLbased yield predictions were compared under scenarios with and without disease pressure and varietal aging to quantify their respective contributions to yield gaps. Results and conclusions: Actual yields under high nitrogen fertilizer treatments averaged 65% of the simulated potential yield in the dry season, but only 52% in the early wet season, and 47% in the late wet season. Seasonal constraints varied: tungro disease widened the dry-season gap, a greater cumulative number of cropping cycles reduced the early wet-season gap, and varietal aging increased the late wet-season gap. The gap between ORYZAsimulated potential yields and ML-predicted actual yields narrowed when biotic stresses and varietal turnover were accounted for, with predicted yields reaching 63–69% of potential. Even under these improved conditions, yields remained below the commonly cited 80% benchmark, suggesting that this global threshold is challenging in tropical wet seasons. Significance: These findings highlight the complementary value of ML and crop simulation models for diagnosing yield gaps and underscore the need for seasonally tailored interventions, particularly more frequent varietal replacement and improved disease management."