Ocean fronts are sharp boundaries between different water masses and different types of vertical structure, that are usually accompanied by enhanced horizontal gradients of temperature, salinity, density, nutrients and other properties. Ocean front recognition is important, as it can provide essential information for understanding the properties and dynamics of the oceans and atmosphere. In order to understand the oceanographic processes, ocean fronts have been a subject of study for many years, a variety of methods and algorithms have been proposed to address the problem of ocean front. However, most of the existing ocean front recognition methods are built upon human expertise in defining the front based on subjective thresholds of relevant physical variables. In recent years, with the inexpensive availability of a large and rapidly expanding data set of remotely sensed sea surface temperature (SST), arises a growing interest and demand for automatic techniques to recognize and detect fronts. The main objective of this thesis is to introduce a new approach for ocean front recognition that is based on Deep Learning methods, in particular, is the Convolutional Neural Networks (CNN) ，which is able to automatically recognize and classify the front into strong and weak ones by using remote sensing data. Deep Learning (DL), is a subfield of machine learning that uses multiple layers of connections to reveal the underlying representations of data. Deep learning has been used extensively in a wide range of fields, and have shown dramatic improvement over traditional methods not only in classic problems, such as speech recognition, natural language processing, object recognition and detection, but also in many other practical applications, including remote sensing. Convolutional Neural Networks (CNN) is one of these DL methods that are receiving increased attention in recent years, and has been intensively used in several distinct tasks in different domains, but are not yet widely explored for the task of ocean front recognition. Due to the requirement of a large number of annotated samples which prohibits its wide use in the ocean front recognition task, collecting and labeling a large amount of data is difficult due to the challenges in obtaining the data from the ocean front domain. Exploring deep CNN''s methods in ocean front recognition is a challenging task because the training data are very scarce. We overcome this challenge by using three different strategies to explore the deep learning methods. Firstly, we propose a deep learning method for an ocean front recognition task, using a pre-trained CNN''s model and a sequence of transfer learning steps via fine-tuning. The core idea is to extract deep knowledge of the CNN''s model from a large dataset, and then transfer the knowledge to our ocean front recognition task on limited remote sensing images. Secondly, we propose a deep network with fewer layers compared to the existing deep architectures for the front recognition task. The proposed network requires less data, less computer power, and less time to be trained and has a total of five learnable layers. Furthermore, a small network has the advantage of avoiding over-fitting, especially when the amount of training data is small. These methods proposed throughout the thesis have been evaluated in several experiments using various image datasets. The experimental results demonstrate the efficiency and effectiveness of the proposed methods for an ocean front recognition. The proposed methods are able to efficiently learn representations of the front from labeled data and automatically recognize and classify the front as weak and strong based on features learned. In contrast, the existing ocean front method requires human expert carefully examining the front to define its region.
Estanislau Baptista Lima
展开 >
Deep Learning Convolutional Neural Network Ocean front recognition Remote Sensing
NETL
