Predicting Roadblock Occurrences Using Machine Learning with AHP for Feature Prioritization and Confusion Matrix Evaluation

Abstract

Roadblocks in Bhutan are common and significant challenges that impact transportation, public safety, and the economy. Predicting these roadblocks is difficult because of the complex interplay between geological, climatic, and topographical factors. This research proposes to develop a predictive model using Adaptive Neuro-Fuzzy Inference System (ANFIS) to create a roadblock susceptibility map of Bhutan. The study incorporates fourteen influencing factors such as rainfall, soil type, elevation, slope, aspect, settlement area, profile curvature, plane curvature, distance to rivers, distance to fault, topographic position index (TPI), Topographic Wetness Index (TWI), Terrain Ruggedness Index (TRI) and Normalized Difference Vegetation Index (NDVI), with roadblock inventory data. These factors were processed using Quantum Geographic Information System (QGIS) to build a geodatabase. The data are split into 70% for training the ANFIS model and 30% for validating the results. The ANFIS model incorporates the neural networks and fuzzy logic principles and it gives better predictions. The performance accuracy, evaluated using the confusion matrix, was 0.8408, indicating good predictive ability. The Analytical Hierarchy Process (AHP) was used to determine the relative weight of each factor, which was then applied in the Weighted Over Method (WOM) to create a susceptibility map. This map, generated for both the entire country and individual districts, can aid in mitigation efforts and serve as a preliminary tool for future infrastructure planning.