CSE Team Received Best Presenter Award at SLAAI-ICAI 2026

[20th November 2026, at University of Sri Jayewardenepura]
The Department of Computer Science and Engineering, University of Moratuwa, is proud to announce that Team HydraPath was awarded the Best Presenter Award at the 10th SLAAI International Conference on Artificial Intelligence, in “Core AI and Machine Learning” track, for their publication titled “A Road State Error Taxonomy for Flood Road Segmentation in UAV Imagery”.
The team consists of Mr. Ashidu Dissanayake, Mr. Dasun Illangasinghe, Mr. Sehan Dissanayake, and Mr. Shanil Diwakar from Batch 22, supervised by Prof. Dulani Meedeniya. The Department congratulates them on achieving this recognition and commends their continued dedication to advancing AI for societal benefit!

Abstract: Flood road accessibility mapping from unmanned aerial vehicle (UAV) imagery requires distinguishing errors with unequal consequences. A flooded road reported open can expose a vehicle to standing water, while an open road reported flooded may cause a detour. Mean intersection over union (mIoU) does not retain this direction of confusion. This study collapses ten class segmentation outputs into three road states and partition disagreements within a road relevant region into mutually exclusive False-Open, False-Block and Hallucination families, separating missed and fabricated Hallucination. Pixel and component rates rescore ordinary predictions directly. They have evaluated DeepLabv3+, PSPNet and Mask2Former on FloodNet under two flooded road emphasis schemes and three seeds. Under unweighted training, only two of six taxonomy rates reproduce the mIoU architecture ordering and flooded road predicted open reverses it (rank correlation −1.0). The highest mIoU configuration has the largest rate of this confusion. Boosting the rare, flooded road class reduces erased flooded road pixels by 18.5–41.7% but increases fabricated flooded road budget share by up to 54%, while mIoU changes by about 0.6 points or less. These diagnostics expose error tradeoffs that aggregate accuracy conceals.
