Session
Weekend Session 1: Advanced Technologies - Research & Academia I
Location
Utah State University, Logan, UT
Abstract
Early detection of wildfires is crucial for preventing the spread of fires and protecting lives and properties. In recent years, satellite-based wildfire detection is becoming popular because of the coverage area and cost limitations of traditional detection methods such as ground-based observation and aerial surveillance. In particular, using CubeSat has the advantage of real-time monitoring and early detection of wildfires in a large area at a low cost. However, the CubeSats have limited image quality due to physical limitations such as size, weight, and power, which reduce detection performance. Therefore, this paper proposes a novel approach for early wildfire detection with CubeSat images using deep learning and super-resolution techniques. Considering the limitations of CubeSat, a dataset of three-channel RGB images was used for binary classification. Landsat-8 images of ten bands were preprocessed into RGB images and enhanced by 4x using Real-ESRGAN. The study utilized transfer learning for wildfire detection using two pre-trained deep learning models, MobileNetV2 and ResNet152V2. The results proved that the super-resolution of the satellite images improved the wildfire detection precision, recall, and f1-score by about 3~5%, depending on the models.
Early Wildfire Detection With CubeSat Images Using Single Image Super-Resolution
Utah State University, Logan, UT
Early detection of wildfires is crucial for preventing the spread of fires and protecting lives and properties. In recent years, satellite-based wildfire detection is becoming popular because of the coverage area and cost limitations of traditional detection methods such as ground-based observation and aerial surveillance. In particular, using CubeSat has the advantage of real-time monitoring and early detection of wildfires in a large area at a low cost. However, the CubeSats have limited image quality due to physical limitations such as size, weight, and power, which reduce detection performance. Therefore, this paper proposes a novel approach for early wildfire detection with CubeSat images using deep learning and super-resolution techniques. Considering the limitations of CubeSat, a dataset of three-channel RGB images was used for binary classification. Landsat-8 images of ten bands were preprocessed into RGB images and enhanced by 4x using Real-ESRGAN. The study utilized transfer learning for wildfire detection using two pre-trained deep learning models, MobileNetV2 and ResNet152V2. The results proved that the super-resolution of the satellite images improved the wildfire detection precision, recall, and f1-score by about 3~5%, depending on the models.
