This is the first of two papers that describe a procedure for pattern recognition of seismic data in order to locate deposits of hydrocarbons. [For a review of the second paper, see <CR> this issue, Rev. 8609-0852.] This paper is concerned with the generation of training pattern and with feature selection. To obtain a sufficient variety of training patterns, seismic traces are synthesized from density and velocity data of actual well logs which provide the basis for an a priori probability distribution. Out of 29 commo- nly used features that are extracted from the traces, sets of three to four are selected by a “bottom-up” procedure. In the given example, an 85 percent discrimination accuracy is obtained for the given example.