Machine Learning and Explainable AI for Transferable Estimation of Average Daily Traffic (ADT) on Local Roads in Thailand

Road network designing, planning, and management require accurate and reliable traffic information, particularly average daily traffic (ADT). However, direct ADT surveys are often costly and fail to comprehensively cover local roads. This research aims to develop a transferable ADT estimation model for local roads in Thailand using practical variables i.e. easily accessible data and low operational costs. The study covers eight representative areas, with a total of 963 traffic count locations, across several cities in Thailand that vary in economic and geographical characteristics. The methodology compared traditional linear regression with several advanced Machine Learning (ML) techniques, including Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). To address model transferability, models were constructed under two frameworks: (a) the city-specific model (CSM) which is calibrated using data uniquely identified for each individual city, and (b) the city-type model (CTM) which employs stratification based on a distinction between a provincial center city and a secondary city. SHapley Additive exPlanations (SHAP) was applied for feature interpretability. Key findings indicate that the XGBoost algorithm within the CSM framework achieved the highest accuracy (R² ≈ 0.65). However, the CTM, notably with the RF model (R² ≈ 0.62), demonstrated superior transferability, affirming its potential for reliable application in other cities of similar type without the need for comprehensive local data recalibration. SHAP analysis further confirmed that road physical characteristics and accessibility are more critical to estimation accuracy than population-related variables.