Neural network labeling of the Gulf Stream
Abstract
An expert analyst can study an infrared (IR) image of sea surface temperature (SST) in the western North Atlantic offshore of Cape Hatteras and define a Gulf Stream north wall where the surface is not cloud covered, but human analyses are time consuming, tedious, qualitative, and subjective. Numerical depictions range from a latitude-longitude grid (40,000 points) to a complex vector form (132 latitude-longitude pairs along-stream) to a complex empirical orthogonal function (CEOF) or normal mode sum, in which 10 modes (2650 values) `explain'' most of the variability. A new representation is needed every 2 to 3 days; a new CEOF series only requires new coefficients (20 values). The CEOFs can be used to interpolate to give reasonable Gulf Stream position analyses from spotty observations provided by an analyst. The next step was to use neural network technology to get CEOF coefficients directly from SST imagery. Neural networks have high computation rates, are typically more fault tolerant, and make weaker assumptions about the underlying distributions than traditional statistical classifiers. A two-layer perceptron network trained by the back-propagation algorithm was used for this study. The authors'' attempts proved the concept''s feasibility; a neural network was trained to output the first 3 coefficients using the Gulf Stream vectors as inputs. Image- type input was then simulated by mapping the vectors onto a 40 km grid; the CEOF coefficients were highly correlated to the correct ones for the training set, but less so for the test set. The next stage of work aimed at improving performance. The hardware and software was upgraded to reduce the training time by a factor of 6, and the input grid resolution and the number of output modes were increased. Doubling the resolution while confining input to the Gulf Stream''s historical envelope gave a 100 resolution increase with a 64 increase in input nodes. The paper gives details of the results, conclusions, and recommendations. This work shows that neural networks can yield coefficients which can produce realistic Gulf Streams, accurate enough to use with edge labeling and optical interpolation schemes.
- Publication:
-
Applications of Artificial Neural Networks II
- Pub Date:
- August 1991
- DOI:
- 10.1117/12.44996
- Bibcode:
- 1991SPIE.1469..637L