THE DEEP LEARNING APPROACH FOR PAN SHARPENING ASTER SWIR DATA

Abstract

Summary Multispectral satellite images are effective for solving a wide range of problems in the various application areas. In particular, ASTER spectroradiometer images recorded in the Short Wave Infra-Red (SWIR) and Thermal Infrared Ranges (TIR) ranges can provide the lithological mapping and mineralogical composition of rocks. The scale of the resulting maps is based on the spatial resolution of the recorded digital images corresponding to the sensor. The ASTER spectroradiometer corresponds to the thematic maps with 1: 100000 - 1: 500000 scales. The goal is to increase the spatial resolution (upsampling) ASTER images in the SWIR and TIR bands with saving their spectral characteristics. The problem is the inability to use most known Pan-Sharpening methods because of ASTER images specification. The stack of imagery does not have the appropriate high spatial resolution panchromatic reference image Using any image (from visible or near-infrared (VNIR) spectral ranges) as a reference is not a good idea because they contain the data captured in the narrow spectral range (only in green, or red, or infrared). Therefore, the applying of standard approaches causing significant distortion of the spectral characteristics of SWIR images after the Pan-Sharpening procedure. In this paper, the main idea is using Convolutional Neural Networks (CNN) and Deep Learning techniques for superresolution resampling Aster multispectral imagery. Accuracy assessment of proposed model shows better RMSE metric values (up to 50%) in the spectral characteristics of transformed images, as compared with standard pan-sharpening technics.