Automatic detection of Mediterranean water eddies from satellite imagery of the Atlantic Ocean
A new machine learning approach is presented for automatic detection of Mediterranean water eddies from sea surface temperature maps of the Atlantic Ocean. A pre-processing step calculates the direction of the temperature gradient at each point of the map. Given a map point, information on the surrounding thermal gradient field is extracted through a binary mask and organised as a numerical vector of gradient angles. This vector is forwarded to a multi-layer perceptron classifier that is trained to recognise patterns of gradient angles generated by positive and negative instances of meddy structures. The proposed system achieved high recognition accuracy with fast and robust learning results over a range of different binary masks. Results were also characterised by a very low detection rate of false positives. The simple and modular structure of the image processing routines reduce computational costs and make the system easily modifiable.
thanks for you question. i see a few problems. first, training data. i am not aware whether there are enough training data for the phenomenon. second, the system detects temperature gradients, i don't know whether tsunamis are characterised by a temperature gradient. altimetric maps of the ocean could be a better choice for tsunami detection. finally, time. the system uses SST maps that in my case i got from the oceanographic institute. i don't know how long it takes to produce one of such maps, then the system may take some 10/20 minutes to scan the SST map. this is ok for eddies, since we need to analyse one image every 6 hours, but i think in the case of tsunamis we would need something faster.
It is clear that the given technique can predict wether a specified point in the satellite image is a meddy centre or not. Hence in order to find the meddies, it is necessary to identify possible meddy centers within the map. Is there any procedures availbe to identify potential meddy centers, or it is necessary to traverse the whole image?
dear charles,
thank you for your question. in section 6 i mentioned 2 possible ways of using the algorithm. in the first case, the whole map is scanned and a map of meddy detection results is produced. since the algorithm can distinguish meddy points from open sea points, we should get positive meddy detections scattered in an area at the centre of meddy regions. in the second case, we can segment the map into "tiles" of the size of the feature extraction mask and get a map of meddy detection regions. in a couple of months we should be able to start the implementation of this last part. however, the algorithm is quite robust. i personally selected the meddy "centres" and i can tell you they are very rough estimates. so i would expect the area of positive identification results to be quite large around the actual centre. there are no procedures to identify beforehand meddy centres. if we had any, we wouldn't need the proposed meddy detection algorithm.
Hi Castellani,
Thanks for the reply. What i ment is, is there any way to identify points in the map which have a high probability of being a meddy centre. You have mentioned that you personally choosed rough estimates. What i am asking is, is that possible to automate this selection process? Not by checking each and every point through the neural network. I fell like some image processing technique could be helpfull for this.
dear charles,
as far as i know, there are no ways to identify beforehand points that are meddy centres. the main problem is the high variability of meddy structures. meddies vary in temperature (i.e., they can be warmer or colder than the neighbouring regions), their size can vary fom 20kms diameter to almost 200kms diameter, they can happen almost everywhere in the area under consideration.
from my experience and from what i read in the literature, the only way to identify a meddy centre is first to locate the meddy and then estimate its centre. most commonly, the centre is estimated from analytical models of circular water motion structures or using circle detectors. but first the meddy has to be identified.
anyway. in our case, we are only interested in detecting the presence of meddies, possibly as a map of meddy locations, and present the idenfication results to the experts of the oceanographic institute for further analysis.
which image processing techniques do you think may help? i have unsuccessfully tried several combinations of edge detectors, histogram analysis techniques, thresholding, etc. the result was always unsatisfactory, typically i wasn't able to isolate ditinctive meddy features. the only solution i found so far is the texture analysis approach i have proposed in the paper.
I wonder if the system can be used to detect or predict Tsunami problem. Thanks.