Machine Learning Approaches to Improve Resolution of Sentinel-1 Data.
Suggested elective courses: radar remote sensing/advance image analysis
No fieldwork is included/required
Being freely available, weather- and sunlight-independent, Copernicus Sentinel-1 data are increasingly being used for a variety of remote sensing applications. While Sentinel-1 imagery at medium resolution provides unprecedented visual impressions of the Earth's surface, curiosity is high for more detailed ground monitoring. This detailed inspection can be accomplished with very high-resolution data. These data are typically provided by commercial satellites, which can limit remote sensing scientists' access to these rich sources of information (due to acquisition costs). Under these circumstances, improving the spatial resolution of Sentinel-1 is a compromise solution to create clear view of SAR image with a better pixel size. Currently, some super-resolution techniques have already been introduced to sharpen the optical image and photos of normal cameras. In this proposal, for the first time, a machine learning method specifically tailored to the properties of the SAR image will be developed to improve the resolution of the Sentinel-1 images.
The main goal of this proposal is to use a convolutional neural network or machine learning methods to improve the resolution of the Sentinel-1 SAR image and produce high-resolution data. A set of Sentinel-1 data and high-resolution SAR images (e.g., RADARSAT-2) are available to train the model. Thus, the trained model can later be used to create a high-resolution image of any area on Earth based on the current Sentinel-1 medium-resolution data. Currently, machine learning and deep learning solutions have been proposed in the literature for remote sensing image processing. The idea is to take a state of the art machine learning method and possibly make it specific to the characteristics of the data to create a high-resolution mapping model. Once the model is built, the results can be evaluated in one or more specific topics (e.g., heat map of individual buildings, activity trend of aircraft at airports) that require high-resolution data