This study presents an advanced pothole detection system utilizing ensemble learning (YOLOv9 instance segmentation and Mask R-CNN) and a Multi-Criteria Decision Making (MCDM) framework to improve detection reliability. The system combines YOLOv9 for rapid instance segmentation and Mask R-CNN for precise segmentation, experimenting with adjusted confidence thresholds to enhance detection rates in challenging scenarios. For Yolov9 instance segmentation model achieved a mean Average Precision (mAP) of 0.908 at 0.5 IoU and an F1-score of 0.58 at a confidence threshold of 0.282. The F1-confidence curve highlights a strong balance between precision and recall, but further work is needed to ensure generalization. Dynamic weights are used to merge outputs, leveraging the strengths of both models. The MCDM framework refines detections by evaluating pothole features such as size, position, and shape. While the system demonstrates high detection accuracy of 20%, narrowly and over-specific defined MCDM criteria may lead to overfitting, limiting adaptability to diverse conditions. The study underscores the importance of balancing accuracy and adaptability for reliable performance in varied environments.