An elastic contour matching model for tropical cyclone pattern recognition.

In this paper, an elastic graph dynamic link model (EGDLM) based on elastic contour matching is proposed to automate the Dvorak technique for tropical cyclone (TC) pattern interpretation from satellite images. This method integrates traditional dynamic link architecture (DLA) for neural dynamics and the active contour model (ACM) for contour extraction of TC patterns. Using satellite pictures provided by National Oceanic and Atmospheric Administration (NOAA), 120 tropical cyclone cases that appeared in the period from 1990 to 1998 were extracted for the study. An overall correct rate for TC classification was found to be above 95%. For hurricanes with distinct "eye" formation, the model reported a deviation within 3 km from the "actual eye" location, which was obtained from the aircraft measurement of minimum surface pressure by reconnaissance. Compared with the classical DLA model, the proposed model has simplified the feature representation, the network initialization, and the training process. This leads to a tremendous improvement of recognition performance by more than 1000 times.