Lightweight Remote Sensing Image Denoising via Knowledge Distillation

Since multispectral remote sensing images (RSIs) contain the abundant information of the surface environment, they have been widely applied in diverse research areas, such as earth observation, agricultural monitoring, and geological exploration. However, RSIs generally suffer from the interference of random noise during the recording and transmission, which significantly impacts the accuracy and reliability of subsequent tasks. Existing neural network-based denoising methods often rely on the heavy parameters to model the generation process of the clean image from its noisy counterpart. However, it is time-consuming deploying such model in real application scenarios. To address the above issue, we present a lightweight denoising model based on the proposed Simplified Residual Spatial-Spectral Module (SRSSM) to effectively extract the spatial and spectral features of RSIs, while maintaining a low computation complexity. A training strategy based on the knowledge distillation is designed to constrain the feature distribution by referring to a larger teacher model. Extensive experiments demonstrate the effectiveness and efficiency of the proposed method compared against existing models. Moreover, the downstream tasks including object detection and clustering are performed on the denoised images to further validate the necessity of denoising process.