Abstract: Typically, large-scale training data is needed for Deep Neural Network (DNN) training. This is also the case for Natural Disaster Management (NDM) applications, as e.g., for flood and forest fire detection and segmentation. Unfortunately, annotated image datasets are scarce, since their annotation is very time-consuming, particularly for fire (flame/smoke), burnt areas, and flood images. To this end simulations of natural disasters can be of great importance, providing accurate, photorealistic, and pre-annotated images. Using tools such as Unreal Engine 5, the AirSim plugin for Unmanned Aerial Vehicle (UAV) emulation, and Procedural Content Generation (PCG) tools, creating environments with high variety can be effortless. Such Big Visual Data can be vital for augmenting NDM datasets that may have limited images in terms of count and the variety of scenery. This lecture focuses on utilizing these tools to create a simple pipeline capable of generating such data enabling DNNs to achieve even greater performance in NDM applications.
Figure 1: Flood simulation.
Figure 2: Forest fire Simulation.
Flood-and-Wildfire-Simulations-v2.0