An explainable causal machine learning framework integrating multi-source sensor data for early warning of crop yield and quality: A case study of rice under low-light stress

Summary

This study addresses the interpretability challenge in precision agriculture by proposing an analytical framework combining SHAP explainability analysis with automated causal mediation scanning applied to rice grown under low-light stress. Using multi-source sensor fusion (meteorological, physiological, and hyperspectral), the framework decoupled causal pathways underlying yield and quality co-evolution, revealing that abnormal protein increases stem from nitrogen concentration effects under restricted carbon assimilation, and that vapour pressure deficit serves as a key micro-meteorological proxy for grain quality. The resulting early-warning model achieved R² > 0.85 for yield prediction one month before harvest, whilst hyperspectral data alone outperformed multi-source fusion for predicting internal grain physicochemical traits.

Regional applicability

The methodology is geography-agnostic and transferable to United Kingdom rice production where it occurs (limited commercial scale), or more broadly to other temperate cereals experiencing low-light stress from adverse weather or shading. The explainable machine learning framework and causal inference approach are applicable to UK precision agriculture programmes, though validation would be needed under local environmental and cultivar conditions.

Key measures

Yield prediction accuracy (R²), grain protein content, grain chalkiness, vegetation indices (enhanced vegetation index, photochemical reflectance index, red-edge slope), vapour pressure deficit, cumulative environmental stress, plant physiological recovery capacity

Outcomes reported

The study developed an interpretable machine learning framework integrating multi-source sensor data (micro-meteorological, physiological, hyperspectral) to predict rice yield and grain quality under low-light stress conditions. Early-warning models achieved high-accuracy yield prediction (R² > 0.85) one month before harvest, whilst spectral-only models effectively predicted internal physicochemical quality using canopy features.

Topic tags