High-resolution downscaling of multi-band satellite images without co-located high-resolution data: a new approach based on training images
Abstract
Last-generation satellite imagery offers extremely sharp representation of the Earth surface up to submetric resolution. These products are useful to monitor earth-surface natural processes and human activity, but they are still very expensive and limited in coverage, since they require the use of specific space-borne devices. Conversely, lower-resolution images are available with larger spatial and temporal coverage. In this paper, we propose a technique to downscale coarse-resolution multi-spectral satellite images using the information contained on a limited training set of coarse-to-fine resolution images (TI). The technique, based on the Direct Sampling algorithm [1], aims at simulating the fine-scale image by sampling data from the TI where a similar coarse-scale data pattern is found. The peculiarity of the method is that it avoids the use of co-located fine-resolution data, as usually required by data fusion techniques. The method is tested on the downscaling of a series of 4-band (R,G,B,NIR) satellite images from the Planet product (3m resolution). Co-registered WorldView-3 and -4 images (4 bands, 1.2m resolution) are used as hi-res equivalents. A continuous portion in these images is omitted and regenerated with the technique. The region considered includes both natural and urban landscapes, depicting relatively stationary heterogeneity. The results show that the technique can generate realistic images with respect to the hi-res reference, preserving fairly well statistical properties such as the intensity histogram, the number of objects and their shape. Moreover, being a stochastic method, it allows uncertainty analysis by generating multiple equiprobable images. The method is only at its early stage of development, but it holds great potential to extend the coverage of hi-res images limited in space or time. [1] G.Mariethoz et al. (2010), Water Resour. Res., 10.1029/2008WR007621.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.H34B..04O
- Keywords:
-
- 0434 Data sets;
- BIOGEOSCIENCESDE: 1855 Remote sensing;
- HYDROLOGYDE: 1926 Geospatial;
- INFORMATICSDE: 1942 Machine learning;
- INFORMATICS