Deep Convolutional and Generative Networks for Ocean Synoptic Feature Extraction and Super Resolution from Remotely Sensed Images

Abstract

Accurate extraction of Synoptic Ocean Features and Downscaling of Ocean Features is crucial for climate studies and the operational forecasting of ocean systems. With the advancement of space and sensor technologies, the amount of remote-sensing ocean data is rising sharply. There is a need for precise and reliable algorithms to extract information from such remotely sensed datasets. Deep learning algorithms have shown significant superiority over traditional physical or statistical methods for several remote-sensing applications. Two important applications are ocean synoptic feature extraction (needed to extract useful information submerged in data) and downscaling of satellite images (needed due to insufficient resolution of current imaging sensors). This thesis introduces two novel deep learning algorithms: W-Net (for Ocean Feature Extraction) and PF-GAN-SR (for Downscaling of Sea Surface Temperature Satellite Images). Ocean Synoptic Feature Extraction: For operational regional models of the North Atlantic, skilled human operators visualize and extract the Gulf Stream and Rings (Warm and Cold Eddies) through a time-consuming manual process. There is a need for an automated dynamics-inspired system to extract Gulf Stream and Rings. We have developed a deep learning system (W-Net) that extracts the Gulf Stream and Rings from concurrent satellite images of sea surface temperature (SST) and sea surface height (SSH). Our approach's novelty is that the above extraction task is posed as a multi-label semantic image segmentation problem solved by developing and applying a deep convolutional neural network with two parallel Encoder-Decoder networks (one branch for SST and the other for SSH), implemented as a WNet. W-Net is the first neural architecture and deep learning system developed for automated synoptic ocean feature segmentation. For the Gulf Stream, we obtain 82.7% raw test accuracy and a low error of 4.39% in the detected path length. For the Rings, we obtain more than 71% raw eddy detection accuracy. Downscaling of Sea Surface Temperature Satellite Images: The unavailability of high-resolution remotely sensed images affects the quality of ocean forecasting and ocean feature extraction. Typical downscaling for geophysical applications is achieved using bi-linear/ bicubic interpolation, which is not good for large downscaling ratios. To improve the current state of the art, we developed a Bayesian algorithm for Super-Resolution (Downscaling) of lower resolution geophysical fields observed by satellites. The key novelty is the development and use of Generative Adversarial Networks (GAN) to learn the prior probability distribution of the high-resolution geophysical fields from historical data and/or model forecasts. The trained GAN is used to sample from the high-resolution prior and a Particle Filter along with the low resolution data (observation) is used to obtain the posterior high-resolution geophysical field. The resultant algorithm has been named the Particle Filter Generative Adversarial Network super-resolution (PF-GAN-SR) algorithm. PF-GAN-SR is applied to downscale sea surface temperature fields in the northwest Atlantic Ocean. Results show consistent performance across different downscaling ratios. Notably, the high-resolution fields obtained from PF-GAN-SR have a better similarity score with the true high-resolution field as compared to existing Super- Resolution methods.