Intelligent Enhancement for Low-Light Earth Remote Sensing Based on Physical Process Modeling

Earth remote sensing is advancing towards achieving all-day availability and enhanced intelligence. However, the presence of low light conditions during nighttime poses a significant challenge as visible remote sensing imaging information becomes heavily intertwined with complex noise components, often leading to its submergence in noise. This issue severely hampers the effective utilization of data. Traditional denoising methods typically rely on assumptions of a universal noise model, which fails to address the intricate noise problem unique to remote sensing applications. Deep neural networks (DNNs) based on learning offer promising solutions; however, obtaining real training sample pairs for space-based low-illumination remote sensing images remains challenging. To overcome these obstacles, this paper proposes a method for enhancing low-light remote sensing based on high-level modeling of physical processes. By meticulously modeling, measuring and calibrating imaging physical processes and detector noise characteristics, we establish an extensive collection of high-fidelity training sample pairs consisting of both noisy and noise-free inputs. These samples are then utilized to effectively train the network model framework, enabling stable and high-performance capabilities specifically tailored for low-light detection using specific hardware configurations.