Performance Comparison of LRT and YOLO for Floodwater Segmentation from CCTV Images: A Case Study in Nan Province, Thailand

Abstract

This research article presents an approach for detecting water-covered areas in closed-circuit television (CCTV) images to support flood surveillance in Nan Province, Thailand. The proposed method employed a Likelihood Ratio Test (LRT) model under the assumption of equal-variance Gaussian distributions between classes. Pixel intensities from grayscale images within a predefined Region of Interest (ROI) were used as input features to classify each pixel as either “water” or “non-water.” The LRT model was compared with two widely used deep learning-based segmentation models, YOLOv11n-seg and YOLOv11x-seg. Experimental results show that the LRT model achieved an F1 Score of 0.84. That was comparable to or better than the two YOLO variants trained using domain adaptation. Notably, LRT demonstrated a significant advantage in processing speed, reaching up to 679.54 frames per second (fps) on a CPU, making it highly suitable for real-time applications on resource-constrained systems. While the LRT model yielded strong performance within the specific context of the study area, the YOLO models exhibited better generalization capabilities under domain shift conditions. Therefore, YOLO models may be more appropriate for scenarios involving diverse or uncertain data characteristics. This research highlights a comparative perspective between simple-structured models and deep learning models, providing practical insights for the development of effective and context-appropriate flood surveillance systems.