Converting image between different satellite sensors: a statistical approach
Abstract
Currently, more satellites are observing the Earth than ever before. The past 5 years have seen substantial changes, with the launch of Sentinel 2 that provides 10 m multi-spectral imagery and the Planet Labs constellation that provides daily worldwide imagery at 3 m, among others. At the same time, there is still an important reliance on the legacy of older missions, such as the Landsat program which is crucial for long-term change assessment. This large variety of satellites includes a host of different sensors, resulting in a huge variety of spectral measurements. This variability can rapidly make cross-sensor studies cumbersome, especially if some other variations are involved, such as the revisit frequency or spatial resolution of the data. Here we propose a solution based on multiple-point statistics to convert an image to its equivalent for another sensor. The method uses training images of the target sensor to produce a simulated image based on a pattern matching approach. This method shows good results and can also be used to extend some information to missing bands of the original sensor, for example adding a NIR band to a RGB image. In addition, since it is based on geostatistics the method provides an uncertainty quantification, which can be propagated further depending on the application. More generally, this framework can provide a significant gain in automatizing short- and long-term change detection studies where an important number of sensors is used.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.H31H1980G
- Keywords:
-
- 0434 Data sets;
- BIOGEOSCIENCESDE: 1855 Remote sensing;
- HYDROLOGYDE: 1926 Geospatial;
- INFORMATICSDE: 1942 Machine learning;
- INFORMATICS