Super-Resolution of Advanced Himawari Imager Data Using SRCNN
Yohei Kato(a),Masayuki Matsuoka(a*)

a) Department of Information Engineering, Mie University, 1577 Kurima-machiya,Tsu, 514-8507
*matsuoka[at]info.mie-u.ac.jp

Abstract

Recent advancements in deep learning have enabled notable improvements in satellite imagery resolution through super-resolution techniques. This study focuses on enhancing the spatial resolution of Himawari Advanced Imager (AHI) data using a convolutional neural network, specifically the Super-Resolution Convolutional Neural Network (SRCNN). The objective is to generate high-resolution images from their low-resolution ones while preserving key structural and radiometric information. The dataset consists solely of Himawari imagery, covering six spectral bands: B01, B02, B04, B05, B06, and B15. High-resolution (HR) and low-resolution (LR) image pairs were created and divided into smaller tiles based on original image sizes for model training and evaluation. The SRCNN model was trained to learn the mapping from LR to HR features. To assess super-resolution performance, several image quality metrics were employed. Global metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) were used to evaluate fidelity and structural consistency. Brightness and contrast were also analyzed for radiometric evaluation. Furthermore, Local Contrast was introduced to assess the preservation of local details and textures. The results showed that the SRCNN-based method improved both visual and quantitative image quality across all bands. Comparisons for each band indicated that enhancement effectiveness varied depending on spectral characteristics. These findings suggest future improvements could be achieved through network optimization and band-aware training strategies. This research demonstrates the feasibility of applying deep learning-based super-resolution to Himawari satellite images. It holds potential for enhancing applications such as weather monitoring, disaster assessment, and environmental observation. In particular, generating high-resolution data from existing images can support faster, more accurate analysis in time-critical scenarios like natural disasters.

Keywords: CNN-based super resolution, image quality evaluation, spectral band

Topic: Topic A: General Remote Sensing