Nine Coupled Irrigation-Agronomic Treatments for Water-Saving Rice Production on Albic Soil: An Interpretable...

Area of Science:

• Agricultural Science
• Agronomy
• Soil Science

Background:

• High-latitude japonica rice systems face challenges in maintaining productivity due to freshwater scarcity and low temperatures.
• Developing integrated irrigation and agronomic strategies is essential for enhancing yield stability and water use efficiency (WUE) in cold-region rice production.

Purpose of the Study:

• To investigate the effects of coupled irrigation regimes and auxiliary practices on rice yield and WUE under cold-region conditions.
• To develop and apply an interpretable machine-learning framework for identifying key yield predictors and outlier treatments in multifactorial field experiments.

Main Methods:

• A two-year field experiment was conducted with nine treatments combining controlled irrigation (CI) at varying thresholds (60-80% θfc) with film mulching (FM) or soil amendment (SA).
• An interpretable machine-learning diagnostic framework using elastic net (EN) and random forest (RF) was employed to analyze yield predictors and identify outliers.
• A four-level evidence convergence framework integrating model consistency, cross-model consensus, independent rhizosphere data, and permutation testing was utilized.

Main Results:

• The soil amendment combined with 75% θfc CI (SACI) treatment increased grain yield by 12.3% and reduced water input by 17.0%, achieving a 35.3% higher WUE (1.801 kg m⁻³) compared to conventional flooding.
• Root volume at the milk stage was identified as the most significant yield predictor, supported by rhizosphere evidence.
• The film mulching combined with continuous flooding (FMCG) treatment was identified as a yield-response outlier, linked to hot-anoxic rhizospheric inhibition.

Conclusions:

• The SACI strategy offers an evidence-based co-management approach for enhancing rice productivity and water use efficiency in cold-region agriculture.
• The developed machine-learning diagnostic paradigm provides a generalizable and reliability-quantified methodological template for complex field experiment analysis.