Detection of Ocean Eddies using Deep Convolutional Neural Networks

Abstract

The ocean eddies are circular water currents with a lifetime of 10 to 100 days and radial scales of 25 to 250 kilometers. Eddies are important for mixing and moving heat, salt, and biogeochemical tracers across the world’s seas. Additionally, eddies have been found to affect marine ecosystems as well as nearby near-surface winds, clouds, rainfall, and moderate the effects of climate change. The cyclonic and anti-cyclonic eddies correspond to negative and positive sea surface height anomalies. As a result, satellite-measured sea surface height images provide information on the eddy character-istics. Data from altimeters that measure sea surface height are utilized in mesoscale eddy detection. These statistics accurately portray the level of sea surface height. Un-derstanding ocean eddies is crucial for research on marine biological ecosystems and climate change. However, manual monitoring of eddies is not feasible. Therefore, the development of automatic techniques is essential, for the detection of ocean eddies. The current automatic eddy detection approaches suffer from unreliable predictions and high computational complexity. Hence, in this dissertation, the problem of automatic detec-tion of eddies from sea surface height images using deep convolutional neural networks is addressed. Here, different supervised learning methods are proposed for the detection of ocean eddies using altimeter data. The developed architectures outperform the ex-isting techniques. At first, a deep convolutional neural network approach based on an attention mechanism is proposed to detect and classify ocean eddies and enhance model performance. The feature representation is further improved by appending the input data and the data produced from the attention method. The different sizes and shapes of eddies make automatic eddy segmentation challenging. U-Net makes a dense prediction to solve this problem. However, the network architecture is very intricate. In this thesis, a dilated convolution based U-Net is developed for the detection of ocean eddies using sea surface height data. This technique decreases architectural complexity without sac-rificing performance. Further, a new residual path is also proposed to cascade encoder outputs with the decoder. The residual path replaces the conventional skip connection between the encoder and decoder modules. Second, due to cascaded convolutions and nonlinearities, spatial details usually get lost in high-level feature maps. This makes reducing false detections for small objects with a lot of shape diversity difficult. This problem is addressed by using the attention mechanism to choose pertinent spatial data from low-level feature maps and passed to the decoder. A VGG16-based U-Net with an attention mechanism which is referred to as modified U-Net or MU-Net is proposed to address this issue. The proposed attention modules are merged with the base network VGG16 to enhance feature representations of ocean eddies. Third, the existing auto-matic eddy detection approaches suffer from high model and computational complexity and have poor multi-scale context fusion. A novel architecture is proposed to tackle the challenging task of extracting objects of varied sizes from remote-sensing images. An attention mechanism is designed to obtain the eddies class context information. A series pyramid pooling module is proposed to aggregate the global context data. To improve the effectiveness of both class context and global context of feature maps and to boost model performance, a feature enhancement approach is proposed. Finally, a dual en-coder and decoder-based architecture is proposed for eddies detection. Since encoded information lacks semantic information, they hinder the segmentation performance. To address this issue, a novel attention module that accumulates semantic data while sup-pressing irrelevant data is proposed. Further, a novel tracking algorithm is proposed to track anti-cyclone and cyclone eddies. The key characters like eddy coordinates and their contours which provide oceanographic representations of an eddy movement are effectively captured. In an experimental analysis of two datasets containing 4383 and 5480 SSH images, the findings of the experiments indicate that the proposed techniques consistently outperformed the current eddy detection approaches.