Comparative Study Between Real-Time and Non-Real-Time Segmentation Models on Flooding Events

Scene understanding of aerial imagery is essential for proper emergency response during catastrophic events such as hurricanes, earthquakes, and floods. Unmanned Aerial Vehicles (UAVs) capture aerial images and analyze the context by passing images into a semantic segmentation model for monitoring damaged areas. However, the state-of-the-art semantic segmentation models are mainly trained and evaluated on ground-based datasets such as Cityscapes, MS-COCO, and CamVid, unsuitable for aerial image segmentations. For example, extracted features from objects in aerial perspective are distinct from objects on the ground view. Hence, neural networks cannot properly segment an aerial scene, especially on deformed or damaged objects during disasters. This research analyzes current semantic segmentation models to explore the feasibility of applying these models for emergency response during catastrophic events. We compare the performance of real-time semantic segmentation models with non-real-time counterparts constrained by aerial images under adversarial settings. Furthermore, we train several models on the FloodNet dataset, containing UAV images captured after Hurricane Harvey, and benchmark their execution on special classes such as flooded-buildings vs. non-flooded buildings or flooded-roads vs. non-flooded roads. In this research, real-time UNet-MobileNetV3 yields 59.3% test mIoU while non-real-time PSPNet [1] attains 79.7% test mIoU on the FloodNet, demonstrating the trade-off between accuracy and efficiency in the segmentation models.