Nine Coupled Irrigation–Agronomic Treatments for Water-Saving Rice Production on Albic Soil: An Interpretable Machine-Learning Diagnosis

Sustaining rice productivity under the dual constraints of freshwater scarcity and low-temperature stress represents a pressing challenge for high-latitude japonica rice systems worldwide. There is an urgent need to develop coupled irrigation–agronomic management strategies that jointly safeguard yield stability and water use efficiency (WUE) in cold-region rice production. In this study, a two-year field experiment was conducted in 2024–2025 on albic soil (Albic Luvisols, WRB; θfc 38.2% v/v, pH 5.80, clayey texture with poor permeability and a propensity for subsurface waterlogging) in the Sanjiang Plain, Heilongjiang Province, China (47°15′ N, 133°28′ E), with nine coupled “irrigation regime × auxiliary practice” treatments, comprising conventional continuous flooding, four-level controlled irrigation (CI) at lower thresholds of 60%, 70%, 75%, and 80% θfc, and their combinations with film mulching (FM) or a humic-acid-based soil amendment (SA). An interpretable machine-learning diagnostic framework was developed, with elastic net (EN) as the primary analytical model and random forest (RF) as a nonlinear control, to simultaneously identify core yield predictors and outlier treatments. The principal findings were: (i) The soil-amendment-coupled 75% θfc CI treatment (SACI) increased grain yield by 12.3% and reduced water input by 17.0% relative to conventional continuous flooding, with WUE reaching 1.801 kg m−3, a 35.3% gain over the control (p < 0.05); these improvements were consistent across both individual years (year × treatment interaction: p = 0.601; inter-year rank correlation ρ = 0.967). Lowering the CI threshold below 75% θfc significantly reduced grain yield through diminished effective-panicle retention. (ii) Multi-method consensus analysis (Kendall’s W = 0.871, p < 0.01) identified root volume at the milk stage as the most strongly and consistently associated statistical predictor of yield formation, with convergent mechanistic support from independent rhizosphere evidence (Eh, TTC reductive activity). Definitive causal validation awaits isotope-tracing experiments. (iii) The film-mulching × continuous-flooding treatment (FMCG) was diagnosed as a yield-response outlier (permutation test p = 0.003), three in situ rhizosphere measurements (redox potential, root TTC-reducing activity, and rhizosphere temperature) supported the proposed mechanism of hot–anoxic rhizospheric inhibition. Methodologically, this study develops a four-level evidence convergence framework that integrates intra-model self-consistency, cross-model (EN vs. RF) consensus, independent rhizosphere evidence, and distribution-free permutation testing, with Jackknife+ conformal prediction and companion Monte Carlo simulations (1000 replicates) used to quantify the reliability boundaries under small-sample conditions (n = 27). These findings provide an evidence-based irrigation–soil co-management strategy for cold-region rice production in Northeast China, and the proposed diagnostic paradigm offers a generalizable, reliability-quantified methodological template for interpretable small-sample modeling in multifactorial coupled field experiments.