Evaluation of CNN-Based Regression Models for Automated SNR Estimation of High-Resolution Satellite Imagery

초록

Accurate estimation of the Signal-to-Noise Ratio (SNR) in high-resolution satellite imagery is particularly challenging in homogeneous regions where edge-based method is inapplicable. This study proposes a convolutional neural network (CNN)-based regression framework designed to automatically estimate SNR in such regions. A synthetic dataset was constructed by deriving reference SNR from natural edges and generating degraded images with Gaussian noise. Four CNN architectures-VGG-16, ResNet-18, DenseNet-121, and EfficientNet-B0-were evaluated with transfer learning from ImageNet. Experiments with KOMPSAT-3A imagery demonstrated that DenseNet-121 achieved the best overall performance, confirming the effectiveness of deep learning for automated SNR estimation in homogeneous areas of satellite imagery. © ACRS 2025.All rights reserved.

키워드