effectively measure suspended sediment
loads in storm events, particularly in small basins. Continuous turbidity measurements can be
used, along with discharge, in an automated system that makes real-time sampling decisions to
facilitate sediment load estimation. The Turbidity Threshold Sampling method distributes
sample collection over the range of rising and falling turbidity values and attempts to sample all
significant turbidity episodes. A data logger activates a pumping sampler when specified
turbidity conditions are met. The resulting set of samples can be used to accurately determine
suspended sediment loads by establishing a relation between sediment concentration and
turbidity for any sampled period with significant sediment transport. Relations between
turbidity, concentration, and particle size are examined at five sites in northern California, USA.
Despite the influence of particle size, turbidity is in all cases superior to flow as a surrogate for
sediment concentration.

Key words particle size; regression; sampling; suspended sediment; turbidity


INTRODUCTION

Accurate measurement and estimation of suspended sediment transport is dependent on the
timing and frequency of data collection. It is common in streams and rivers for most of the
annual suspended sediment to be transported during a few, large runoff events. Automated data
collection is essential to effectively capture such events. Although it is possible to rely solely on
manual measurements, important flows are infrequent, unpredictable, and when they do occur,
trained personnel may not be available to collect the required information.
There is currently no practical method to directly measure the full range (submicron to
2 mm) of suspended sediment concentration (SSC) in the field. Pumped or manual samples must
be transported to a laboratory for analysis. However, a number of companies offer turbidity
sensors that can be deployed on a continuous basis in streams. While turbidity cannot replace
SSC, it can be of great benefit as an auxiliary measurement. The continuous turbidity record can
reveal sediment pulses unrelated to flow, providing information about the timing and magnitude
of sediment inputs. And turbidity can be used in an automated system that makes real-time
sampling decisions to facilitate sediment load estimation. Such a system, called Turbidity
Threshold Sampling (TTS), has been used at a growing number of gaging stations in northern
California, USA, since 1996 (Lewis & Eads, 2001). Its design objectives are to:
(a) Facilitate accurate estimation of suspended sediment loads at a reasonable cost.
(b) Provide an adequate number and distribution of physical samples to validate every
significant rise in turbidity and calibrate turbidity against SSC for each period being
estimated.
(c) Provide a continuous estimate of sediment concentration and flux based upon turbidity. 2
The sampling algorithm ensures that a wide range of SSC is sampled in each transport event, so
that reliable turbidity-SSC relations can be developed. These are then applied to the near-
continuous turbidity data to produce a corresponding time series of estimated SSC.
While turbidity is virtually always a better SSC surrogate than flow, its quality is less
consistent due to fouling from biological organisms, detritus, and waterborne debris. Mechanical
wipers can prevent fouling from small contaminants such as fine organics, sediment, algae, and
macroinvertebrates, but larger debris must be manually removed. Fouling from larger debris is
best controlled by overhead suspension of the sensor on an articulating boom (Eads & Lewis,
2001, 2002). Data affected by fouling are in many cases difficult to distinguish from episodes of
sediment transport. Some types of fouling can be readily identified on data plots with
experience. However, fouling that occurs during storm events can often be identified only by
plotting the turbidity against SSC from corresponding physical samples, or by comparing the
turbidity with independent readings from a second sensor. By activating a pumping sampler
during each significant change in turbidity, TTS provides physical samples that can be used to
validate the turbidity.


SAMPLING PROTOCOL

The TTS algorithm attempts to collect physical samples at specific turbidity thresholds.
Thresholds are chosen so that the square roots are evenly spaced to adequately define loads for
small storms without oversampling large storms. A programmable data logger, recording at 10
or 15-min intervals, instructs an automatic pumping sampler to collect a sample whenever a
threshold is crossed. To avoid sampling ephemeral turbidity spikes caused by passing debris, a
threshold must be met for two intervals before signaling the sampler. Because most sediment is
discharged while turbidity is in recession, more thresholds are utilized while turbidity is falling
than when it is rising. Reversals are detected when the turbidity drops 10% below the preceding
peak, or rises 20% above the preceding trough. In addition, the change must be at least 5 NTU,
and the new course must continue for at least two intervals before declaring a reversal. At the
time a reversal is detected, a sample is collected if a threshold has been crossed since the
preceding peak or trough, unless that threshold has already been utilized in the past five
intervals. The user can modify the specific thresholds and numerical parameters mentioned here.
The above rules provide reasonable assurance of avoiding extraneous sampling in the
presence of normal turbidity fluctuations. However, it will not prevent oversampling when
debris snags on the sensor or its mounting apparatus, causing extended fluctuations.
Oversampling due to fouling can cause a pumping sampler to quickly reach its bottle capacity
and fail to sample the next important event. Sites experiencing fouling require more frequent
field visits. Remote telemetry provides the most effective means of detection.


SIMULATIONS

Lewis (1996) simulated the above sampling protocol with varying threshold scales, fitting
procedures, and sample sizes to evaluate the effectiveness of TTS and associated regression
models for load estimation. The sampled populations consisted of five Caspar Creek (north
coastal California) storm events for which both turbidity and SSC were measured at 10-min 3
intervals. Sampling based on the square-root threshold scale generally produced more accurate
results than cube-root or logarithmic scales; and regression variable transformations (square root,
cube root, and logarithm) tended to increase estimation errors. But there was not a great deal of
sensitivity to either the threshold scale type or the choice of regression model. The most
important result was that RMS errors were small, less than 10% in nearly every combination
simulated, with mean sample sizes of 4-13 per storm. For samples sizes of at least five, RMS
was generally no more than 5% of the load. Estimates based on log-log discharge-SSC rating
curves had RMS 1.9-7.5 times larger than those based on linear turbidity-SSC regressions, fitting
a single curve to each storm.
In one of the storm events, applying separate turbidity-SSC regressions to periods of rising
and falling turbidity significantly improved the estimation. In another storm, quadratic
regression had a slight 1-2% edge over linear regression. But, in most cases, a single linear
regression performed nearly as well or better than other methods, and caution should be
exercised in applying nonlinear fits or multiple fits, particularly in the presence of outliers.
Extrapolating nonlinear curves can lead to large errors and dividing the data is inefficient unless
there clearly are multiple relations.
Power functions representing the turbidity-SSC relation based on log-log fits were less prone
to extrapolation error than polynomials. Log-log models performed nearly as well as linear
models in estimating sediment load; and they have two advantages over linear models: (1)
predictions are always positive, and (2) residual variance is often less dependent on turbidity.
This last feature can improve estimation of the variance of the sediment load.
The variance of the load estimate can be estimated without bias if the regression model
assumptions are satisfied. Formulas are given for the linear regression model by Lewis (1996)
and for log-log regressions by Gilroy et al. (1990). Lewis investigated the errors associated with
applying linear regressions to realistic SSC data generated with log-normal error models and
concluded that, for typical TTS sample sizes of 4-11 samples per storm, variance estimation was
unreliable regardless of the model applied. The variance estimator associated with log-log
models had little bias, but its RMS error ranged from 52 to 110%. The variance estimator
associated with linear models performed even worse, with RMS from 73 to 244%. In contrast,
both models produced very good load estimates, with RMS from 5.2 to 7.9% for log-log models
and 5.6 to 8.3% for linear models fit to the log-normal data. For larger sample sizes, variance
estimation would improve, and the log-log model should produce reliable estimates if the
residuals can be normalized by log-transformation.


EXAMPLE

The TTS method is illustrated by a February 2000 storm event at Caspar Creek (Fig. 1). The
range of turbidity measured during the 5.5-day storm