THE USE OF MACHINE LEARNING FOR THE SYNTHESIS OF STREAM DISCHARGE – GAGE HEIGHT RATING CURVES

The conventional method of measuring stream discharge is to install a gaging station that automatically measures gage height, which is the elevation of the water surface above a fixed datum. The gage height is then converted into stream discharge by means of a rating curve, which is developed by carrying out multiple simultaneous field measurements of stream discharge and gage height over a wide range of discharge values. The major cost of a gaging station is the manpower required for the development of the rating curve. As a consequence, gaging stations can be prohibitively expensive in remote and impoverished areas of the world. The problem with the conventional method is that, for each new gaging station, a rating curve is developed from field measurements as if it has never been done before. By contrast, the U.S. Geological Survey has developed rating curves for 61,240 gaging stations, which has involved 3.68 million simultaneous measurements of gage height and discharge. The objective of this research is to use machine learning for the synthesis of rating curves from easily measurable hydrogeologic parameters. The first step has been to create a conceptual framework that identifies the relevant hydrogeologic parameters. Frequent reverse flow or flood waves preclude the existence of a rating curve (unique relationship between gage height and discharge). If a rating curve exists, then a stable channel has a power-law rating curve. Deviations from the power-law curve result from deposition (power-starvation) or scouring (sediment-starvation), which could occur at the high or low range of discharge or both. The eight types of deviation (including no deviation) from the power-law curve can be regarded as eight functional forms of rating curves, which can be represented as lines, parabolas or cubic polynomials on plots of the Z-scores of the logarithms of gage height and discharge. Rating curves can be classified into the eight types based on the hydrogeologic criteria of (1) stream slope (2) relative erodibility of the stream banks (3) distance to the nearest upstream and downstream confluences with relatively significant discharge. The caret package in the R programming language is being used to predict rating curves from the above hydrogeologic parameters in the state of Utah. Further results will be reported at the meeting.