| dc.description.abstract | Remote sensing is the process of obtaining informative data about an object from afar. While this applies to many different methods of data collection, as well as different domains to collect data from, there is a universal constant that the technology and recording instruments utilized towards this purpose are being improved day by day. However, for every piece of higher quality data collected from these new instruments, there still exists much more data from now lower quality measurement instruments that can not provide the historic significance nor insights that will begin to be made with increases in accuracy.
Machine learning has shown the capability to recognize very subtle patterns between different types of data. In recent years, one such method known as Generative Adversarial Networks (GANs) has displayed much success in artificially creating new data based on given input by learning from corresponding example output. Through this research, we show the potential for using the complex generative abilities of GANs to improve the accuracy and quality of remote sensing data taken from older instruments by using more precise data from newer technology as examples to learn from.
We take data obtained from two atmospheric satellites utilizing two Ozone measurement instruments TROPOMI and its predecessor OMI that collect Nitrogen Dioxide (NO\textsubscript{2}) readings in the troposphere, early indicators of pollution, and use it to create paired datasets based on location and the time a location was crossed over by each satellite. Using this training set, we train an Enhanced Super Resolution GAN (ESRGAN) to improve both the resolution and measured values of OMI data used as input, inspired by TROPOMI training examples. | |
