Game analysis of future rice yield changes in China based on explainable machine-learning and planting date optimization

Abstract

Context

Global warming's escalating severity necessitates sophisticated approaches for predicting rice yield.

Research Question

Combining crop models with data-driven techniques, such as machine learning, can more effectively grasp the complex interplay of variables influencing crop growth. It remains a significant challenge to balance accuracy and interpretability in such hybrid models.

Methods

The research integrated the Decision Support System for Agrotechnology Transfer (DSSAT) with statistical and machine learning models respectively, to assess rice yield changes in China under four future Shared Socio-economic Pathway (SSP). SSPs are scenarios that integrate socioeconomic trends with greenhouse gas emissions and radiative forcing pathways, which affect the phenology and yield of rice. The Shapley Additive Explanation (SHAP) method was employed to interpret the model, effectively determining the interplay among variables influenced rice yields. Mitigated the negative impacts of climate change on rice yield through the planting date optimization.

Results

Projections indicate significant rice yield losses in China without CO₂, worsening with increased radiative forcing (p < 0.001). Considering rising CO₂, single-season rice yields are projected to increase by 0.1–3.6 %, early rice by 4.6–9.5 %, while late rice yields are still decrease by 2.3–8.8 %. The rising CO₂ can offset yield losses for single and early rice but not for late rice. The hybrid approach which combined the Random Forest (RF) with the DSSAT performed best in predicting rice yield. Studies showed that rising temperatures caused rice yield losses in China, yet we found that Growing Degree Days (GDD) exerted a more negative impact (p < 0.001). In high-precipitation regions, deep soil moisture is more influential than shallow soil moisture, whereas the reverse was true in drier areas (p < 0.001). Advancing planting dates for early and single rice and delaying for late rice can increase yields (p < 0.001). Adjusting to optimal planting dates, single-season rice yields increased by 3.3–6.3 %, early rice increased by 9.7–18.3 %, while late rice still decreased by 1.0–4.7 %.

Conclusions

Without considering the impact of CO₂, significant rice yield losses in China are projected. Even with the fertilization effect of CO₂, rice yields remain negatively impacted by climate change. However, implementing appropriate measures, such as optimizing planting dates, can help Chinese rice production benefit under changing climate.

Implications

This study offers insights into balancing accuracy and interpretability in hybrid models and provides guidance for local policymakers to address future climate change.