SPRINT: Spectra Preserving Radiance Image Fusion Technique using holistic deep edge spatial attention and Minnaert guided Bayesian probabilistic model

Abstract

Remote Sensing Image Fusion produces a high resolution multispectral image by merging panchromatic image with the corresponding low resolution multispectral counterpart. The limitation of the existing image fusion techniques is that it lacks in maintaining the spectral characteristics of the multispectral image in the fused output. The motivation of our proposed work is to develop balanced and robust image fusion method, named Spectra Preserving Radiance Image Fusion Technique (SPRINT). The technique developed extracts deep edge map from medoid intensity matched image using Holistic Nested Edge Detection (HNED). Minnaert function is applied on denoised panchromatic radiance image along with Digital Elevation Model (DEM) and solar angles’ computation to determine the surface topography. SPRINT’s core design emphasizes on holistic deep edges for spatial attention and terrain guidance using minnaert parameter at each image pixel in a Spectra Preserving Bayesian Probabilistic Model. The unique data pre-processing engine generates fusion ready representative datasets to trigger SPRINT processing workflow. The Indian Cartosat-1 panchromatic and Resourcesat-2/2A multispectral sensors’ datasets along with IKONOS and USGS-NASA’s Landsat-8 OLI images covering diverse landscapes are used for image fusion evaluation and assessment. It has been found that SPRINT’s fusion performance is superior to the state-of-the-art (SOTA) image fusion methods in terms of both visual effects and quantitative metrics. The Normalized Difference Vegetation Index (NDVI) using SPRINT’s fused radiance image tends to have negligible deviation at various classes with respect to reference NDVI. It has been observed that SPRINT derived surface reflectance values have close agreement with original reflectance measurements.