The use of support vectors from support vector machines for hydrometeorologic monitoring network analyses

Abstract

Hydrometeorologic monitoring networks are ubiquitous in contemporary earth-system science. Network stakeholders often inquire about the importance of sites and their locations when discussing funding and monitoring design. Support vector machines (SVMs) can be useful by their assigning each monitoring site as either a support or nonsupport vector. A potentiometric surface was created from synthetic data and 800 random observation locations (sites) as an analog to a groundwater-level network. Using generalized additive models for potentiometric surface prediction, simulations show that a subsample of support vectors from the 800 sites will out perform random samples of sample size equaling the support vector count. Support vector percentages from simulation quantify the recurrence that SVMs assign each site as a support vector, and these percentages in turn measure site importance. An example application of support vector percentages identifies important monitoring sites needed to regionalize the 0.1 annual exceedance probability peak streamflow. The results indicate that 152 of 283 streamgages with support vector percentages equalling 100 percent have not operated since about 2000 and generally have much smaller drainage areas than the greater streamgage network in Texas. The drainage area disparity is an indication of historical imbalance in peak streamflow data acquisition from various stream sizes in Texas.