THe specific pattern recognition problem treated in this half-posthumous paper ( Fu died early this year ) is that of the detection of bright spots in one-dimensional seismic traces and in two-dimensional seismograms. Any analyst not specialized in seismology will move very cautiously through this strange and earthquaking land, where not only the terminology is very peculiar, but a lot of things are supposed to be understood, without explanation. What is, to begin with, a "bright spot"? From the conclusions of the paper, one can see that in seismograms it is possible to recognize flat spot patterns, pinchout patterns, bright spots, and so on. From the Introduction, a definition of the bright spot can be extracted : candidate bright spot = high amplitude signals + low frequency and / or negative polarity ; bright spot = candidate bright spot + continuous reflection layer. However, the exact meaning of each characteristic is rather vague, and the included seismograms do not contribute tobrighten the spot.
A first block diagram of a 1 D synactioc pattern recognition system ( why syntactic? ) is presented, in order to detect candidate bright spots in testing traces selected from the input seismograms, using a tree classification technique. A likelihood radio test is used, unfortunately invaded by a lot of synactic, notational, and mathematical errors ( e.g. , P ( r ) instead of p n ( r ) and p 1 ( r ); &sqrt;&sgr; instead of &sgr;; ∧ ( r ) = p ( r &slash; H 1 ) &slash; p ( r &slash; H 0 ) instead of p n &slash; p 1; threshold &bgr; and quotient &eegr; without definition, etc.). Also, a “Levenshtein distance” [1] is calculated, using an 8-level optimal quantization and a very special encoding system. Both suffer from a presentation which is far from satisfactory.
Then a second block diagram for recognition of bright spots in a 2D seismogram of candidate bright spots is presented. It is based on three kinds of string distance computation (shifted string matching method, modified distance, and substring matching computation). The recognition of bright spot strings are based on a distance less than a threshold t ⋍ M / 3.
Two experiments on real seismograms (at Mississippi Canyon and at High Island), each containing 64 traces of 2 s of duration, seem to support the proposed method. Even so, the block diagrams are very incomplete; the functions indicated in each step are not explained, so the reconstitution of the program and the verification of the claimed results are impossible.
It is lamentable that such an important and delicate signal processing/pattern recognition problem, comparable to that of searching for a needle in a haystack, could be so deficiently presented, notwithstanding the scientific credits of the authors.
