Synergizing deep learning and representation learning for hyperspectral image analysis

Abstract

Hyperspectral imaging has emerged as an important remote sensing technology due to its ability to capture detailed spectral information across hundreds of contiguous wavelength bands. This rich spatial-spectral information has enabled a wide range of real-world applications, including environmental monitoring, precision agriculture, mineral exploration, land-cover mapping, and the assessment of greenhouse gas emissions. However, in practice, hyperspectral data acquisition is often affected by sensor limitations, atmospheric disturbances, illumination changes, and transmission errors. These factors introduce different types of noise that degrade image quality and make subsequent analysis more challenging. In addition, methane (CH4) plume detection and segmentation is difficult due to weak spectral signatures, strong inter-band redundancy, complex backgrounds, and limited annotated data. These issues motivate the need for learning-based approaches that can improve hyperspectral image denoising (HSID) and CH4 monitoring performance. This work focuses on developing deep learning (DL) frameworks for HSID, CH4 plume segmentation, and spatial-spectral representation (SSR) learning. The main objective is to improve hyperspectral image (HSI) analysis by making better use of spatial, spectral, contextual, and frequency-domain information. The contributions are organized into three main directions.