SUPER-RESOLUTION FOR ENHANCED AREA ESTIMATION AND TARGET DETECTION

Abstract

Low-resolution of satellite images restrict the performance and applications of satellite image processing. Due to inability of the methods to convert low-resolution satellite imagery to high-resolution, there might be a large amount of temporal data loss. Satellite imagery deals with the problem of mixed pixels and aims to minimise it. This study focuses on the review of the methods to classify subpixels inside the mixed pixels, along with the background involving endmember extraction and abundance mapping. Further, the study utilises two subpixel mapping algorithms, i.e. Pixel Swap (PS) Method and Inverse Euclidean Distance (IED) Method for the subpixel distribution of binary classes. The model is trained on resampled satellite image to compute the best possible method among these to perfom binary target detection. Considering various accuracy metrics, IED performs better than PS. This IED algorithm is further used on Landsat-8 data, with a scale factor 3, to obtain the high-resolution image. This high-resolution output image is used in the application of area estimation. In area estimation, the area corresponding to pure target pixels only is calculated for the reference image, the output image and the low-resolution image. Accuracy of IED for testing was not as high as training. But, considering the increase in area estimation of the target from original low-resolution imagery to output high resolution imagery, the method is reliable enough for subpixel mapping.