Decoding soil health constraints in regional agroecosystems: Machine learning reveals microbial enzymatic thresholds and drivers

Soil health assessment in heterogeneous agroecosystems presents significant challenges in integrating biological mechanisms and diagnosing multifactorial constraints. This study applied interpretable machine learning approaches to establish scientifically derived thresholds for constraint typology and provide spatially explicit diagnostics in farmlands of Ningbo, China. Eight soil health index (SHI) formulations were evaluated, incorporating combinations of total (TDS) versus minimum datasets (MDS) with linear and nonlinear scoring and integration methods. A random forest model was employed to identify key drivers of SHI, with SHAP (SHapley Additive exPlanations) values used to determine critical constraint thresholds. Structural equation modeling and microbial history strategy analyses were subsequently applied to quantify interactive mechanisms affecting soil health in sites with multiple constraints. Results demonstrated that SHI derived from the TDS using nonlinear weighted integration (TD-N-WQI) exhibited superior sensitivity to soil organic carbon (SOC) content (Pearson r = 0.840, p < 0.001) and effectively discriminated against different land-use types (ANOVA F = 50.5, p < 0.001). Microbial enzyme activities emerged as the predominant regulators of soil health, with N-Acetyl-beta-D-glucosaminidase (NAG, 28.2 %), beta-glucosidase (BG, 22.0 %), and xylanase (XYL, 13.9 %) collectively accounting for 64.1 % of feature importance, substantially exceeding contributions from SOC (9.40 %) and total nitrogen (7.80 %). Critical constraint thresholds were as follows: bulk density 1.13 g/cm(3), SOC 16.9 g/kg, electrical conductivity 0.160 mS/cm, pH 4.50 and 6.24, microbial biomass carbon 508 mg/kg, and enzymatic deficits (BG 90.0, NAG 40.7, XYL 15.0, beta-D-Cellobiohydrolase 20.0, acid phosphatase 439 nmol/g/h). Spatial analysis revealed that 39.2 % (127/324) of sites exhibited multiple constraints, predominantly in coastal areas where salinity indirectly reduced SOC through enzymatic and microbial suppression (beta = -0.150, p < 0.05). This study establishes that high microbial enzymatic activity, driven by Y-strategist microbial communities characterized by rapid growth and efficient resource processing, constitutes the core mechanism of soil health. These insights provide a robust, ML-driven framework that identifies critical constraint thresholds to guide precision management and prioritize microbial-based restoration in heterogeneous agricultural landscapes.