Abstarct: Abstract Hyperspectral imaging, a technique that records reflected light from objects in narrow frequency bands, finds applications in diverse fields such as medical imaging, mining, and agriculture. Typically captured within the 400 to 2500 nanometer range, these images enable the identification of unique spectral signatures for different objects. A key application lies in remote sensing, where it facilitates the identification and classification of various land covers from remotely captured imagery. Variations in imaging conditions and differences in hyperspectral sensors can cause a model trained with one hyperspectral image to perform poorly when identifying the same classes in a new image. This phenomenon, termed domain shift, presents a significant challenge for image classification, as labeling new hyperspectral images is both difficult and time-consuming. Furthermore, the presence of unknown samples in real-world scenarios presents another significant challenge. This research will demonstrate that these two issues—domain shift and the presence of unknown samples—are interconnected, and failing to address one can exacerbate the other. This dissertation proposes a deep learning-baxsed method to simultaneously address domain shift and identify unknown samples. The method unfolds in three progressive approaches: first, simultaneous unknown detection with domain adaptation in homogeneous hyperspectral images; second, generalizing the domain adaptation method for heterogeneous images; and third, concurrently addressing domain adaptation and unknown sample identification in heterogeneous images. The proposed method was tested on homogeneous datasets, including Salinas, Botswana, and RPaviaU-RPaviaC, as well as heterogeneous datasets, including RPaviaU-DPaviaC and Ehangzhou-RPaviaHR. For the first approach, target image accuracies reached 91.5% for the Botswana dataset, 62.5% for RPaviaU-RpaviaC, and 81.4% for Salinas. The second approach yielded accuracies of 90.16% for the RPaviaU-DPaviaC dataset and 99.12% for the Ehangzhou-RPaviaHR dataset. In the third approach, using the RPaviaU-DPaviaC dataset with unknown samples present and the network lacking the ability to identify them, accuracy was 70.58%; this figure rose to 73.17% upon equipping the network with unknown identification capabilities. Similarly, for the Ehangzhou-RPaviaHR dataset under identical conditions where unknown samples were present and the network initially could not identify them, accuracy improved from 80.97% to 91.95% after integrating the unknown identification capability into the network.