Target Detection with LWIR Hyperspectral Scene Transfer Based on Deep Learning

Objects exhibit changing spectral characteristics based on weather conditions, and the dynamic spectral behavior of materials under varying weather conditions poses challenges in detecting them in hyperspectral images (HSIs). In this paper, we propose a generative adversarial network (GAN) based scene transfer (ST) method to reduce the negative effects of changing weather conditions on the detection performance of long wave infrared (LWIR) HSIs. The proposed method initially converts the test scene at an arbitrary air temperature to the reference scene at a reference air temperature with the GAN-based ST method. Then, it performs the detection on the transferred scene. When the proposed ST-based method is used, detection results improve even for the HSIs captured at night, and the overall area under the curve (AUC) scores increase by about 8%. Hence, our proposed ST method emerges as a promising solution to enhance the reliability of thermal LWIR HSI analysis, particularly in the presence of different weather conditions between the capturing time of the test and reference HSIs.