Comparative multi-algorithm AI framework for real-time carbon emission optimization in a medium-scale irrigation project in Thailand

This study develops and validates a real-time carbon monitoring framework that integrates automated meteorological data with operational construction logs to enhance environmental impact assessment (EIA) methodology for tropical infrastructure projects. Leveraging 2332 h of synchronized climate–operational data from a reinforced-concrete sluice gate project in Thailand, the framework addresses critical gaps in carbon governance for medium-scale irrigation construction, a sector underrepresented in current EIA literature yet vital to climate adaptation in emerging economies. Three machine learning models (artificial neural networks, random forest, and extreme gradient boosting) were used to capture complex emission patterns. Under the tested monsoon scenarios, all models achieved high predictive fidelity (R 2 = 0.997–0.998) with low normalized errors (mean absolute scaled error < 0.05). The framework demonstrated 5 %–30 % potential emission reduction across operational–climatic scenarios, confirming its robustness for tropical infrastructure management and showing robust cross-algorithm consensus. Feature importance analysis highlighted embodied carbon in concrete (59.7 %) and steel reinforcement (24.8 %) as the primary emission sources, while environmental conditions emerged as influential factors. Furthermore, lower embodied‑carbon levels were observed during operational periods under ambient temperatures below 28 °C. The framework facilitates climate-responsive construction planning via real-time emission forecasting, scenario testing, and adaptive resource optimization. A three-tier implementation strategy encompassing climate-adaptive material scheduling, climate-responsive planning, and scenario-based emission testing demonstrates 5 %–30 % reduction potential while accommodating the technological constraints of tropical developing economies. This approach advances EIA methodology from retrospective assessment toward predictive, climate-responsive decision support, offering a scalable framework for integrating real-time carbon management into digital EIA platforms and aiding national net-zero infrastructure goals. • Comparative AI (XGBoost/ANN/RF) for site-level CO₂e in tropical irrigation. • 2332 hourly records; R 2 ≈ 0.99 for real-time CO₂e prediction from field logs. • Concrete 59.7 % + steel 24.8 % dominate; prioritize material actions. • Periods <28 °C associated with lower concrete/steel-related emissions. • Scenario tests indicate ∼14 % CO₂e reduction via optimized practices.