Landscape fragmentation drives rice heavy-metal health risk and reveals actionable thresholds.

The contamination of staple crops with heavy metals is predominantly assessed through a geochemical lens, treating farmland as a static container while overlooking the organizing influence of landscape structure. Here we advance a "landscape risk filter" perspective and a threshold-identification workflow that links landscape configuration to spatially explicit health risk. Using Jiangxi Province (China) as a case study, we integrated 30-m land-use metrics with paired soil-rice measurements from 377 sites and a standard non-carcinogenic risk assessment, then applied interpretable machine learning (Random Forest-SHAP) together with segmented regression and Bai-Perron structural break tests. The mean Hazard Index (HI) was 1.022, and 35.3% of sites exceeded the safety threshold (HI > 1). Risk was dominated by Cadmium (mean THQ_Cd = 0.812), whereas Arsenic alone remained below the threshold but contributed to cumulative risk at a subset of sites. RF-SHAP identified cropland fragmentation, quantified as land-use/cover patch density (LUCC_PD), as the strongest predictor of HI, exceeding the contributions of soil pH, soil organic matter, and soil metal concentrations. The SHAP response to LUCC_PD was distinctly non-linear, and two robust breakpoints at approximately 2.17 × 106 and 4.38 × 106 LUCC_PD units delineated buffering, escalation, and saturation regimes, statistically confirmed by Bai-Perron tests. Geographically weighted regression further revealed strong spatial non-stationarity, with the fragmentation effect peaking in the west-central hotspot corridor. Together, these results provide quantitative, landscape-based targets for risk zoning and land consolidation policies to prevent the formation of high-risk agricultural mosaics and improve rice food safety.