[A-CG41] 衛星による地球環境観測

[ACG41-04] Enhancing Hyperspectral Imaging Experiments in CubeSats: The ARC-NET Model for High-Resolution Image Fusion

*Wen-Qian Chang^1,2、Pei-Yuan Li¹、Bing-Chen Lai^2,1、Cheng-Ling Kuo^1,2、Tang-Huang Lin^1,3、Loren C. Chang^1,2、Chi-Kuang Chao^1,2、Jann-Yenq Liu^1,2 (1.Center for Astronautical Physics and Engineering, National Central University, Taoyuan City 320317, Taiwan、2.Department of Space Science and Engineering, National Central University, Taoyuan City 320317, Taiwan 、3.Center for Space and Remote Sensing Research, National Central University, Taoyuan City 320317, Taiwan)

キーワード：hyperspectral images, multispectral images, data fusion, Convolutional Neural Network

HyperSCAN (Hyper Spectral Camera ANalyzer) is a hyperspectral imager designed to monitor the Earth's environment and is mounted on the 12U CubeSat SCION-X which is scheduled to be launched in 2026. As a low-cost cubesat payload, hyperspectral imager often faces the problem of insufficient spatial resolution due to the size constraints of CubeSats, which limit the focal length of the camera (longer focal lengths generally provide higher spatial resolution). As a result, fusing hyperspectral images with high spatial resolution images becomes a very important issue. Obtaining hyperspectral images with higher spatial resolution provides more accurate observation data and enables a wider range of data applications. With the aid of the deep learning model, we reconstructed the high spatial resolution hyperspectral images (HR-HSI) with high spatial resolution multispectral images (HR-MSI) and low spatial resolution hyperspectral images (LR-HSI). We propose an Attention-Based Residual Convolutional Neural Network (ARC-NET) for hyperspectral and multispectral image fusion. Characterized by its three different attention mechanisms for hyperspectral images (Channel attention), multispectral images (Self-attention), and fusion of hyperspectral and multispectral features (Fusion attention), ARC-NET can effectively capture pixel-to-pixel and band-to-band correlations, amplifying important information and reducing unimportant information. Furthermore, ARC-NET is designed with a residual connection to retain the original data in case the information is lost during the training process. To verify the performance of ARC-NET, we compared the performance of ARCNET with other models including SSRNET, TFNET, MSDCNN, SSFCNN, ResTFNet and SpatCNN on four datasets including Urban, Indian pine, Pavia Center and Pavia University. The results show that ARCNET performs better on all four parameters of Peak signal-to-noise ratio (PSNR), RMSE (Root Mean Squared Error), ERGAS (Erreur Relative Globale Adimensionnelle de Synthèse) and SAM (Spectral Angle Mapper) compared to other models, which demonstrates the competitiveness of ARC-NET in image fusion. ARCNET has introduced an advanced image fusion algorithm that incorporates a fusion attention mechanism that provides a promising solution for integrating hyperspectral data from HyperSCAN with multispectral images. This approach will be used to combine data from newly developed hyperspectral imagers and RGB cameras in the planned 6U CubeSat mission. The algorithm is intended to significantly increase the usefulness and adaptability of the acquired data sets to support a wider range of applications, including accurate environmental monitoring, disaster assessment and effective resource management.