279 
AUTOMATIC EDDY EXTRACTION FROM SST IMAGERY 
USING ARTIFICIAL NEURAL NETWORK 
Jin Hai a,b *, Yang Xiaomei b , Gong Jianming a,b , Gao Zhenyu a,b 
a School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, Hubei, China 
institute of Geographical Science and Natural Resources, Chinese Academy of Sciences, Beijing, China - 
(jinh, yangxm, gongjm, gaozy)@lreis.ac.cn 
KEY WORDS: Image Interpretation, 2-D Feature Extraction, Image Understanding, Feature Detection, Oceanography 
ABSTRACT: 
Mesoscale eddies have a significant impact on the exchange of material and energy in the ocean, and thus their knowledge are of 
great importance for the study of oceanic circulation. Images of sea surface temperature (SST) created from satellite infrared sensors 
are used to detect mesoscale eddies that have a surface signature in temperature. Various techniques, including texture analysis, 
wavelet transform, mathematical morphology, etc., have been used to identify mesoscale eddies from SST images. However, mainly 
due to the strong morphological variation of eddies which causes the absence of a valid analytical model, these approaches either 
have many limitations or are rather complex. The paper proposes a new methodology for automatic detection of mesoscale eddies 
from SST images using artificial neural network (ANN) and edge detection, and it can be summarized in the following steps: 1) pre 
processing to reduce noise and to obtain maps of temperature gradient, its direction and magnitude; 2) using artificial neural network 
to detect the possible eddy centres; 3) removing the false eddy centres; 4) detecting the edge points of the eddies and fitting them 
into ellipses. This approach has been applied to the detection and extraction of mesoscale eddies in the Gulf Stream area using 
NOAA GOES 10 & 12 SST images, and the experiment has proved that this method has the following advantages: 1) It’s effective 
and robust with high detection accuracy (over 90%), especially for the cold-core eddy (over 95%) since the training set used for the 
neural network is mainly composed of cold-core eddies. If more samples of eddies and non-eddies are used to train the neural 
network, the detection accuracy can be further improved. 2) Not only are eddies detected by the approach, but also the parameters of 
eddies such as centre location, size and direction are also calculated at the same time, which can be rather useful for detecting the 
change of eddies in sequential SST images. 3) The procedure is rather simple, efficient and easily reconfigurable, without the need 
of a valid analytical model. It can be adapted to different conditions such as different sizes of eddies, different cores (cold or warm), 
and different resolution of SST images. Therefore, the proposed approach is rather suitable for automatically detecting and 
extracting eddies from satellite SST images. 
1. INTRODUCTION 
Mesoscale oceanographic phenomena such as mesoscale eddies, 
fronts and upwellings have a significant impact on the exchange 
of material and energy in the ocean, and thus their knowledge 
are of great importance for the study of oceanic circulation. 
Compared to ordinary methods of in-situ observation, the 
technology of marine remote sensing has advantages in 
synoptic coverage of large areas of instant oceanic information, 
collection of long sequence data of global oceans, and 
measurement of marine features. Thus, it has become an 
important method of marine environment monitoring. The 
remote sensing imagery, through data processing and inversion, 
can provide lots of marine feature information related to 
mesoscale oceanographic phenomena, including sea surface 
temperature (SST), chlorophyll concentration, sea surface wind 
speed field, etc. Among them, SST, the first marine 
environment parameter obtained by marine remote sensing, is 
widely used in the research on mesoscale oceanographic 
phenomena, ocean-atmosphere heat exchange, global climate 
change, fishery resources and pollution monitoring (Feng, 
1999). Since almost all the marine dynamic processes are 
related to SST directly or indirectly, mesoscale oceanographic 
phenomena such as mesoscale eddies, fronts, and upwellings 
can be effectively detected by using SST. With the massive 
increase of SST data, it is necessary and urgent to detect the 
marine phenomena in SST images automatically. 
Various techniques, including texture analysis, wavelet 
transform, mathematical morphology, etc., have been used to 
identify mesoscale eddies from SST images. For instance, 
Alexanin and Alexanina (2000) described the SST images as a 
set of oriented textures in the temperature field, calculated the 
dominant orientation of the radiation contrasts and fitted it into 
the elliptical model of eddies. However, mainly due to the 
strong morphological variation of eddies which causes the 
absence of a valid analytical model, these approaches either 
have many limitations or are rather complex (Marcello, J. et al., 
2004). 
This paper proposes a new methodology for automatic detection 
of mesoscale eddies from SST images using artificial neural 
network (ANN) and edge detection. 
2. APPROACH 
The approach can be summarized in the following steps: 1) pre 
processing to reduce noise and to obtain maps of temperature 
gradient, its direction and magnitude; 2) using artificial neural 
network to detect the possible eddy centres; 3) removing the 
false eddy centres; 4) detecting the edge points of the eddies 
and fitting them into ellipses.