M4SFWD: A Multi-Faceted synthetic dataset for remote sensing forest wildfires detection

Forest wildfires are one of the most catastrophic natural disasters, which poses a severe threat to both the ecosystem and human life. Therefore, it is imperative to implement technology to prevent and control forest wildfires. The combination of unmanned aerial vehicles (UAVs) and object detection algorithms provides a quick and accurate method to monitor large-scale forest areas. Nevertheless, most available datasets on forest wildfires comprise single -mode ground-fixed-angle pictures that inadequately represent the intricate terrain, high humidity, low visibility meteorological conditions, and multiscale light flux densities of forest wildfires. To address these limitations, we developed the Multiple scenarios, Multiple weather conditions, Multiple lighting levels and Multiple wildfire objects Synthetic Forest Wildfire Dataset (M4SFWD), which provides remote sensing data on forest fires across diverse terrain types, weather conditions, light flux densities as well as different numbers of wildfire objects. Researchers can employ this dataset to improve the efficacy of fire and smoke detection algorithms, promoting continuous forest monitoring. This paper presents a Multi-Faceted Synthetic Forest Wildfire Dataset based on Unreal Engine 5. We first constructed eight forest scenes with different terrains, weather conditions, and texture effects. We also simulated the light flux density at different times of the day by utilizing real -time ray tracing technology, which created realistic lighting and shadows. Secondly, we introduced a range of wildfire targets with varying scales and numbers into each scenario to enable multiple -angle shooting simulations from a UAV's viewpoint. During evening hours and in foggy conditions, many objects resemble wildfires. To enhance the dataset's precision and reliability for fire and smoke detection, 3,974 images were undergone pixel-level manual annotation using tools like labelImg. This annotation yielded 17,763 bounding boxes, which were subsequently statistically analyzed to ascertain their positions and proportions. Finally, we assessed the applicability of M4SFWD in singlestage, two-stage, and lightweight object detection algorithms by inputting the dataset into various algorithms with different parameter sizes. Based on the experimental results' visualization, M4SFWD exhibited superior performance in scenarios with standard light flux density and large-scale wildfire objects. However, due to its complex contextual information and multiscale object features, false detections and missed detections occurred in other complex multi-faceted scenarios. Thus, optimizing the existing object detection algorithms will be necessary for future research. The dataset is available at: https://github.com/Philharmy-Wang/M4SFWD.