Automatic detection of Mediterranean water eddies from satellite imagery of the Atlantic Ocean

M. Castellani

CENTRIA - Departamento Informática, Faculdade Ciências e Tecnologia, Universidade Nova Lisboa, Portugal.

RENA project

• Remote Detection of Mediterranean Water Eddies in the Northeast Atlantic (RENA)
• Institutions
• Instituto de Oceanografia, Faculdade de Ciencias da Universidade de Lisboa (IO) - P
• Centro de Inteligencia Artificial (CENTRIA), Faculdade de Ciencias e Tecnologia, Universidade Nova de Lisboa
• www.io.fc.ul.pt/fisica/research.htm

RENA project objectives

• Census of meddies in the Northeast Atlantic
• Estimation of their generation frequency
• Determining physical mechanisms controlling interaction between Mediterranean undercurrent and coastal upwelling current system off Iberian Peninsula.

• Problem domain
• Meddies
• Sea Surface Temperature Maps
• Main approaches
• Proposed algorithm
• Experimental results
• Discussion
• Conclusions and further work

Mediterranean water diffusion in the Atlantic Ocean

• Warm and salty tongue of Mediterranean water in the Atlantic
• Mediterranean water
• travels thousands of kms with little mixing
• strongly influences hydrology and dynamics of Atlantic Ocean
• carries particles, suspended material, live organisms
• Meddies play a crucial role in the process

Main features of meddies

• Mesoscale (50-200 km) lens-like oceanic structures
• Solid body cyclonic/anticyclonic rotation
• Typical diameter 50kms (variable)
• Rotation period 6-8 days
• Translation speed 30cm/s

How to detect meddies

• Sea surface temperature (SST) maps
• temperature signature
• Altimetric maps
• anomaly of sea surface level
• Drifters
• trajectory of buoys

Meddy identification approaches

• 2-steps process
• Image pre-processing
• Identification
• Pre-processing step
• extract structure/flow information
• Identification step
• Analytical models
• fit mathematical models
• Machine learning
• artificial neural network classifiers

Image pre-processing

• Isothermal lines
• geometry and distribution
• Edge detectors
• borders of oceanic structures
• Dominant orientation maps
• water temperature gradient flow
• Fourier power spectrum
• edge orientation

Analytical approach

Analytical approach - example

 

 

Analytical approach - advantages

• Description of structure
• Expert knowledge
• heuristics
• Parameters/equations available
• Modelling
• Interpretation

 Analytical approach - drawbacks

• High structural variability of meddies
• Sensitive to noise
• Enhance gradient = more noise
• Complex models
• Not robust

Machine learning approach

Artificial Neural Networks (ANNs)

• Extract numerical features
• ANN identification

Machine leaning approach - example

 

 

Mach. learning approach - advantages

• ANN parallel processing
• Fast
• Can handle large amounts sensory information
• Approximate matching / Generalisation
• Robust to noise
• Robust to structural variations
• Pre-processing less critical
• ANN Learning
• No need for knowledge design
• Easier to re-configure

Mach. learning approach - drawbacks

• Learned knowledge not transparent
• Can not be interpreted
• Can not be used for modelling
• Expertise can not be used for initialisation
• Training may be difficult
• Structure design not straightforward
• Trial and error

Proposed algorithm

• Eddies as textural patterns in image
• Pre-processing
• Temperature gradient direction
• Feature extraction
• Numerical features characterising meddies
• ANN identification
• Multi-Layer Perceptron (MLP) classifier

Textural patterns - visual example

 

 Gradient-based identification

• Water flow direction in meddies rotatory
• Temperature gradient vectors perpendicular to isotherm lines
• Gradient vectors inward or outward concentric patterns 
• Other oceanic phenomena have other patterns
• Temperature gradient field characterises meddies
  

Gradient based identification: pre-processing algorithm

• Analysis of oriented texture
• Image is smoothed (low pass filter)
• Horizontal and vertical gradient of temperature are calculated
• Horizontal and vertical gradient measures are locally averaged
• Local direction of thermal gradient vector is calculated
• Gradient direction map is created

Step 1 – convolution kernels

 

 Step 2 – Gradient direction

• Parallelogram law
• Temperature gradient angle with respect to horizontal polar axis

Gradient based approach feature extraction

• Calculate gradient angle for each pixel p(i,j)
• A feature vector can be defined by picking out values at fixed points
• Binary extraction mask is used

Feature extraction visual example

 

Feature extraction masks

 

The multi-layer perceptron classifier

• Input layer
• Size of feature vector
• Hidden layer(s)
• One or two hidden layers
• Hypertangent transformation function
• Output layer
• Two neurons (meddy/”non-meddy”)
• BP training algorithm
• Learning rate 0.01
• Momentum term 0.1

Training and validation sets

• 105 examples of meddies
• 445 examples of non-meddies (e.g., thermal fronts, open ocean, etc.)
• 550 total data points
• 80% training
• 20% validation
• Each data point centre of oceanic structure
• Training set is balanced by duplicating examples of meddies

Learning results 

Visual comparison

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

summary of results

• high identification accuracy
• robust learning performances
• low rate of false positives
• important since meddies occupy small portion of SST map area
• larger feature extraction masks may allow better accuracy 
• larger feature extraction masks allow smaller MLP architectures
• larger feature vectors may convey more information
• no noticeable difference between one-layer and two-layers MLP architectures

Summary of presentation

• Problem of meddy detection
• Approaches for automatic detection
• Proposed approach
• Image texture analysis
• MLP identification
• Results
• Good accuracy
• Robustness

Further work - feature generation

• Image pre-processing
• Different convolution masks
• Different texture analysis methods (Laws, Fourier power spectrum, Eulerian vector fields)
• Optimise code for speed
• Sampling
• Feature extraction and selection
• Try other binary extraction masks
• Feature selection (statistical analysis, evolutionary algorithms)

Further work - identification algorithm

• Improve ANN classifier
• automatic structure generation
• evolutionary algorithms
• constructive/destructive approaches
• Better training algorithms
• conjugate gradients
• evolutionary algorithms
• Different classification algorithm
• RBF
• KNN