Evaluation of CNN-Based Regression Models for Automated SNR Estimation of High-Resolution Satellite Imagery
Hongjun Youn(a), Jihyeon Lim(b), and Taejung Kim(b*)

a) Program in Smart City Engineering, Inha University
100 Inha-ro, Incheon 22212, Republic of Korea
b) Dept. of Geoinformatic Engineering, Inha University
100 Inha-ro, Incheon 22212, Republic of Korea
*tezid[at]inha.ac.kr

Abstract

With the growing use of satellite imagery, the need for quantitative image quality assessment has become more pronounced. Signal-to-Noise Ratio (SNR) quantifies the ratio of useful signal power to noise power in an image and serves as a key metric for assessing image quality. In particular, high-resolution satellite imagery can exhibit SNR variation due to factors such as atmospheric condition and sensor degradation. There is a strong need for automated evaluation of SNR per image. Traditional SNR calculation methods rely on statistical analysis or high-pass filters within regions with uniform Digital Number (DN) values, often combined with manual operations. These approaches are limited by the difficulty of applying them to images with complex textures or boundaries. Furthermore, manual evaluation lacks consistency and scalability. This study aims to analyze the applicability of a Convolutional Neural Network (CNN) based regression approach for automatic quantitative estimation of SNR in satellite imagery. To this end, homogeneous regions were selected from high-resolution images and augmented with varying levels of artificial Gaussian noise to construct a training dataset. Regression models utilizing existing CNN architectures were then trained and evaluated. All CNN models were pretrained on ImageNet before fine-tuning. The performance of CNN regression models was compared across major architectures, including DenseNet-121, ResNet-50, and EfficientNet-B0. The experiment results showed that DenseNet-121 achieved high predictive accuracy with an RMSE of 7.87 and an R^2 of 0.82. In several homogeneous regions such as ocean and bare land, the model produced stable estimations, indicating that it captured underlying SNR characteristics beyond local intensity or contrast-based cues. Compared to existing statistics-based SNR estimation methods, the proposed regression model maintained precision for different noise levels and demonstrated its applicability in complex sce

Keywords: SNR- Convolution Neural Network- Natural Target-based assessment- Satellite Image Quality- Quality assessment

Topic: Topic C: Emerging Technologies in Remote Sensing