Deep interpolation based hyperspectral-multispectral image fusion via anisotropic dependent principal component analysis.

In remote sensing, the information present in hyperspectral images (HSI) and multispectral images (MSI) often contrasts with each other. HSI has a higher spectral resolution than spatial resolution, while MSI is rich in spatial details. Image fusion aims to integrate crucial information from these source images into a single fused image that enhances both spatial and spectral features. In this work, we introduce a deep interpolating convolutional neural network (DICNN) model that utilizes anisotropic-dependent principal component analysis (PCA) for HSI-MSI fusion. Initially, our goal is to enhance the HSI resolution by training the DICNN on numerous examples. This training aligns the spatial resolutions between HSI and MSI data. Subsequently, we replace the highest variance HSI principal components with their corresponding MSI counterparts obtained after band selection. Additionally, we preserve significant edge information through the incorporation of anisotropic filtering. The proposed methodology demonstrates superior results compared to other existing state-of-the-art methods, both qualitatively and quantitatively. [ABSTRACT FROM AUTHOR]