Target detection in hyperspectral thermal images

This thesis explores target detection methods using hyperspectral thermal images. Hyperspectral thermal images contain data collected from the long wavelength infrared region of the electromagnetic spectrum. These images provide detailed spectral information about the material properties of the targets, which is useful in target detection. However, changes in the weather conditions of the environment where the targets are located make it challenging to detect the target in long-wave infrared (LWIR) hyperspectral images (HSI). Based on this observation, a scene transfer method is proposed in this thesis which transforms the radiance data in the images where the target is difficult to detect, into the images with higher discrimination and hence, enables the detection of targets in more difficult conditions. In addition, target detection studies are carried out with deep learning-based models on radiance data and it is observed that the long-short-term memory (LSTM) based model gives the best results. Also, as a result of target detection studies performed with signature-based matching methods on the radiance data and on the emissivity data obtained by applying temperature-emissivity transform (TES) to the radiance data, it is observed that the emissivity data perform slightly better. The heating and cooling profiles of the objects during the day are different from each other due to the characteristic radiation properties arising from the thermal inertia and heat capacity of the objects. In order to benefit from the temporal radiation characteristics of the materials, target detection is performed on the temperature profiles of the objects in the proposed method, and on the temporal LWIR HSI's created by combining the radiance data in another proposed method. Finally, dimension reduction techniques are applied on LWIR HSI's, and the wavelengths that are important in detecting targets are determined. In the error analysis based on the difference between the radiation emitted from the objects and the black body radiation at the same temperature as the objects, it is observed that the wavelengths for distinguishing the target objects are similar to the wavelengths found by band selection techniques.