Plastic-film mulching is widely applied in arid and semi-arid regions, yet its feasibility for wheat production and carbon footprint mitigation in dry semi-humid areas remains unclear under global aridification. In this study, the Denitrification Decomposition (DNDC) model was used to simulate wheat yield and greenhouse gas (GHG) emissions over 30 years under plastic-mulched ridge-furrow rainwater harvesting (RF) and conventional flat planting (FP) across varied irrigation scenarios. Further, Random Forest (RandForest) coupled with SHAP (SHapley Additive exPlanations) quantified nonlinear effects of water-related factors (including planting practice, irrigation, and pre-sowing and growing-season precipitation) on yield and GHG emissions. Results showed DNDC accurately simulated yield and GHG dynamics, while RandForest-SHAP revealed significant nonlinear water-related influences. Overall, FP yield exceeded RF by 25.8% as mulching reduced sowing area, while yields increased with irrigation. Although RF reduced CO2 emissions by 35.4%, it decreased CH4 uptake by 41.2% and increased N2O emissions by 46.1%. Specifically, the plastic film buffered the effects of precipitation on CO2 and CH4 fluxes but intensified its effect on N2O emissions. While RF reduced indirect emissions by 6.3% through lower agricultural inputs (e.g., irrigation, owing to planting furrow-only application), direct emissions, predominantly from N2O, increased by 62.8%, leading to an 11.6% net increase in the farm carbon footprint. Furthermore, RF only outperformed FP in product carbon footprint and net profits when growing-season precipitation was below 150 mm. Given the region's average precipitation (385.1 mm pre-sowing; 196.6 mm growing-season), the comprehensive sustainability of RF for wheat production in dry semi-humid areas is limited.