Professional Documents
Culture Documents
Worden Et Al. - A Revised Ground-Motion and Intensity Interpolation Scheme For Shakemap
Worden Et Al. - A Revised Ground-Motion and Intensity Interpolation Scheme For Shakemap
Worden Et Al. - A Revised Ground-Motion and Intensity Interpolation Scheme For Shakemap
1785/0120100101
Introduction
ShakeMap is a system for rapidly characterizing the these reports are received and converted to seismic intensity
extent and distribution of strong ground shaking following by the USGS. With minor filtering and with sufficient num-
significant earthquakes worldwide (e.g., Wald et al., 2008). bers, the intensity data reported through DYFI have been
Current ShakeMap systems are deployed and operating at found to be a remarkably consistent and reliable measure
several regional networks within the United States (U.S.), of earthquake effects (e.g., Atkinson and Wald, 2007). In
in various networks worldwide, and at the U.S. Geological fact, DYFI data have been shown to be consistent with USGS
Survey (USGS) National Earthquake Information Center MMI assignments over the entire range of intensities (Dewey
(NEIC) in Golden, Colorado, for the production of Shake- et al., 2002), with only minor differences at the lowest inten-
Maps for important global earthquakes. sities in which DYFI data appear to be more consistent and
The increased availability of worldwide seismic inten- complete.
sity data immediately following significant earthquakes In addition, for reconstructing shaking distributions for
offers the opportunity to supplement (oftentimes scarce) historical events, macroseismic observations can provide
instrumental data for the rapid characterization of ground valuable constraints; often they are abundant, whereas
shaking. Internet users worldwide can also now report the strong-motion recordings are sparse for such events. An
effects of earthquakes on themselves and their surroundings example of the value of adding historic macroseismic data
through the “Did You Feel It?” (DYFI) Web site (see Data and as constraints for ShakeMaps in the process of developing
Resources section; see Wald, Quitoriano, Dengler, et al. earthquake loss-models is documented in the ShakeMap
[1999] for a discussion of the DYFI system and Dewey et al. Atlas (Allen et al., 2008) and subsequent exposure and loss
[2000] for the relation of these internet felt reports to tradi- catalog (Allen, Marano, et al., 2009; Allen, Wald, et al.,
tional Modified Mercalli Intensity [MMI] reports). Within 2009). While we expect that the interpolation methodology
minutes of major earthquakes, hundreds to thousands of described herein will work with traditional MMI as well as
3083
3084 C. B. Worden, D. J. Wald, T. I. Allen, K. Lin, D. Garcia, and G. Cua
intensity from DYFI, our testing (see Discussion and Testing) to ensure that better constrained observations or predictions
uses DYFI data exclusively. are weighted and treated appropriately.
However, direct incorporation of macroseismic data into With the introduction of intensity observations and their
ground shaking maps is not possible because of the differ- utilization for constraining PGM maps, a more rigorous
ences in the units of measurement. Thus, another important consideration of the uncertainties involved is required. Thus,
advancement in the combined use of ground-motion and our goal was twofold: (1) combine traditional instrumental
macroseismic data is the ability to convert intensity to peak observations and predicted ground motions with converted
ground motion (PGM) and vice versa (e.g., Wald, Quitoriano, observations of another ground-motion parameter and (2) pro-
Heaton, et al., 1999; Atkinson and Kaka, 2007; Faenza and cess the intensity map directly, using intensity data as the
Michelini, 2010). These ground-motion/intensity conversion native observation and converted peak ground acceleration
equations (GMICE) allow us to supplement instrumental (PGA) and PGV as supplemental data.
ground-motion data (e.g., peak ground velocity, PGV) with Therefore, we desired a systematic means of combining
another independent data set (e.g., MMI converted to various data types in a given map. Ebel and Wald (2003)
PGV), although doing so introduces the uncertainty of the proposed a Bayesian approach to weighting the various
conversion. contributions, using probability density functions derived
Since its inception, ShakeMap has used well- from the uncertainties associated with those contributions.
documented empirical ground-motion prediction equations Cua and Wald (2008) suggested that a simpler, weighted-
(GMPEs) (e.g., that of Boore et al., 1997) to supplement average approach provides a more direct, analytic solution
sparse observational data. Prior ShakeMap implementations to the problem while accomplishing the same ends as the full
combined recorded and predicted ground motions, giving probability based approach of Ebel and Wald. Here, we fol-
priority to recorded values over predicted values at and low Cua and Wald (2008) and take advantage of the recent
nearby the recording station locations. Intensity values development of intensity prediction equations (IPEs) to
(the intensity grid), in contrast, were derived by simply com- extend Cua and Wald’s approach to incorporate intensity as
bining and converting the peak acceleration and velocity map a native ground-motion parameter. Thus, our treatment of
observed intensity is fully analogous to our treatment of
grid values via the relations of Wald, Quitoriano, Heaton,
PGM and pseudospectral acceleration (PSA).
et al. (1999). To incorporate macroseismic observations into
The estimation of uncertainties depends upon the map
the PGM maps, the observations were converted to peak
layer in question. For native and converted observations we
acceleration and velocity values (using the inverted equations
incorporate an uncertainty-with-distance-from-observation
of Wald, Quitoriano, Heaton, et al. [1999]) and were treated
formula presented by Boore et al. (2003, appendix), though
as if they were recorded ground motions. This simple ap-
our framework allows for other interstation amplitude corre-
proach was useful to provide constraints for the intensity
lation functions. As mentioned previously, additional
and ground-motion maps in areas of sparse seismic station
variance is included for converted observations. For GMPE-
coverage, but it suffered from several approximations and based predictions (or IPE-based predictions for the intensity
limitations. These are addressed in the improved ShakeMap layer), we use the published variance of the GMPE (IPE).
approach that we describe here. Naturally, the variance caused by unknown fault geometry
First, the appropriate use of observations in each grid greatly increases this uncertainty (Electric Power Research
layer should consider whether or not the parameter is native Institute, 2003; Wald et al., 2008). Hence, we continue to
or converted. For example, ground-motion observations are modify the variance when the source dimensions are not
well constrained in the ground-motion layer, and intensity known, as described in Wald et al. (2008) but now extend that
observations converted to ground-motion have greater uncer- treatment to weight the contribution of the GMPE to the aver-
tainty because of added uncertainty associated with the aged ground motions.
conversion process. In contrast, for the intensity grid, macro- In the sections that follow, we first describe the metho-
seismic values are native, and recorded ground motions con- dology of our new approach to computing ShakeMaps.
verted to intensity values carry additional uncertainties. We then discuss some of the advantages of this new approach
Second, the conversion of intensity to ground motion is con- and its implementation over our previous versions of
trary to the intended use of many GMICE. For example, ShakeMap. Finally, we discuss a series of validation exercises
Wald, Quitoriano, Heaton, et al. (1999) developed a relation that demonstrate the viability and benefits of our new
for deriving intensity from PGM but using it in the reverse implementation.
direction is not necessarily as well constrained (and violates
its intended purpose). Finally, the uncertainty of predicted
Methodology
values depends on whether or not the source dimensions
are known. With an unknown source dimension, predicted Under our new approach, the input to ShakeMap for any
ground-motion values may be less certain than converted in- particular ground-motion parameter (PGA, PGV, PSA, inten-
tensity values (because the distance to the source is poorly sity) consists of three distinct types of data: (1) measured
constrained). These uncertainties must be treated rigorously instrumental ground-motion or macroseismic observation
A Revised Ground-Motion and Intensity Interpolation Scheme for ShakeMap 3085
data at specific sites; we call these data “native observa- etc.) Thus, the predicted amplitude, Y obs;xy , at a grid point
tions;” (2) converted data, based on an observation of one x; y from a nearby observation, Y obs , is
type that is converted (by use of an empirical or theoret-
ical relationship) to another type [e.g., MMI fPGV or
Y GMPE;xy
PGV gMMI]; we call these data “converted observa- Y obs;xy Y obs × × Csite ; (1)
Y GMPE;obs
tions;” and (3) ground-motion estimates, based on a GMPE
(or IPE) or from numerical waveform modeling. Our over-
arching goal is to combine the three types of data in a manner where Csite is the term to correct the relative site amplifica-
that (a) is consistent across the different map types (PGM, tion (or site amplification at grid point x; y=site
PSA, and intensity); (b) weights the various data appropri- amplification at location of obs), Y GMPE;obs and Y GMPE;xy
ately given how well they are constrained; and (c) produces are the ground-motion predictions at the observation point
a computed variance for each grid point. and the point x; y, respectively, and their ratio corrects for
(In the discussion that follows, we will frequently refer the relative distance-to-source (i.e., rxy versus robs in Fig. 1a)
to the GMPE or GMPE-based estimates. It should be noted and other source and path effects when the predictions come
that the discussion applies generally to both ground motions from more sophisticated methods.
and intensities, but to avoid excessive repetition the reader Equation 1 holds whether the observation is native or
should assume that GMPE means GMPE or IPE.) converted. In the case of a converted observation, Y obs in
equation 1 is replaced by Y conv fY obs , where f repre-
sents the application of a GMICE.
How Observations Influence Nearby
Ground-Motion Estimates
Uncertainty due to the Influence of Observations
The approach here follows Cua and Wald (2008) in
on Nearby Ground-Motion Estimates
weighting the relative contribution of a datum by a function
of its uncertainty (i.e., its variance, σ2 ). The essential geome- Observations are assumed to have a radius of influence
try of this approach is illustrated in Figure 1a. The observa- within which variance decreases as a function of distance,
tion, Y obs , is scaled to remove the observation point’s site reaching zero at the observation site itself (using the function
amplification effects relative to the point x; y, where a pre- in the Boore et al., 2003, appendix). Similarly, converted
diction is desired. The amplitude is also scaled by a factor to observations have a variance associated with the conversion
account for the difference in distance to the source from the itself, which is added to the distance-dependent variance.
observation and the point of interest. (This distance correc- Predictions from GMPEs will have a variance, typically
tion generalizes well to a geometry correction in situations (although not universally) constant for a given ground-
where the predictions include effects like radiation pattern, motion parameter, that is determined by the study used to
directivity, source-time functions, wave propagation effects, develop the GMPE.
r∆
r∆
rmax
rxy rROI
Source
Figure 1. Illustrations of the relative geometry of observations and grid points: (a) a grid point, x; y, relative to an observation (Obs),
defining the various distance measures rΔ , rxy , and rObs ; (b) the three possible configurations of grid points relative to an observation point:
rΔ < rROI , rROI < rΔ < rmax , and rΔ > rmax .
3086 C. B. Worden, D. J. Wald, T. I. Allen, K. Lin, D. Garcia, and G. Cua
10
The value of the standard deviation associated with the
GMPE, σGMPE , as provided by the relevant study’s authors, is 9
often broken down into interevent and intraevent terms that 8
are combined to produce a total uncertainty for the GMPE.
7
If enough observations are available, ShakeMap attempts
to compute an interevent bias for each event (i.e., a magni-
Weight Ratio
6
tude adjustment that minimizes the residual between the
5
observations and predictions; see the Discussion and Testing
section for more on computing the bias). Applying this event 4
bias should effectively compensate for the interevent uncer- 3
tainty of the GMPE. Thus, if an event bias can be obtained
from the observations, and if the GMPE in question provides 2
which reduces to
Pn Y obs;xy;i
P
Y conv;xy;j
1
σ2GMPE
× σ2 1
i1 Y GMPE;obs;i m
j1 Y GMPE;conv;j × σ2
1
Y
vations, equation (5) (with k 3, ranging over GMPE, The site corrections and the dividend Y GMPE;obs;i
obs;xy;i
(and the
native, and converted observations) can be written as similar dividend for the converted values) are computed once
for each observation in simple On operations. The two
Y GMPE;xy P Y P Y conv;xy;j summations in the numerator are computed independently
σ2GMPE
ni1 σ2obs;xy;i mj1 σ2conv;xy;j
Y xy Pn obs;xy;i
Pm ; (7) of one another (to simplify the computation of σT ) while
1
σ2GMPE
i1 σ2 1
j1 σ2 1 keeping track of the total weights needed for the summations
obs;xy;i conv;xy;j
in the denominator. The entire quotient in equation (10) is
computed for every point in the output grid, and then that
(after Cua and Wald, 2008, equation 6), where Y GMPE;xy and
grid is multiplied by a grid of GMPE predictions (represented
σ2GMPE are the amplitude and variance, respectively, at the
by Y GMPE;xy ). As a final step, a site amplification factor is
point x; y as given by the GMPE, Y conv;xy;j are the converted
applied to each output point to scale the amplitude from
amplitudes scaled to the point x; y, and σ2conv;xy;j is the var-
the GMPE-native site condition to that of the site itself.
iance associated with converted observation j. For points
The implementation described here requires a consider-
zero or near-zero-distance from an observation, σobs;xy is
ably greater computational effort than previous versions of
fixed at a small, nonzero value that ensures that the observa-
ShakeMap. The implementation also lends itself very well
tion will dominate the total ground motion, while avoiding
to array operations. Because we work extensively with geo-
the singularity in the equation. Similarly,
graphic grids, we have for some time found the Generic
Mapping Tools (GMT; Wessel and Smith, 1995) to be con-
1 venient for most of the mapping and visualization, but with
σ 2Y Pn Pm ; (8)
xy 1
σ2GMPE
i1 σ2
1
j1 σ2
1 this new implementation we have delegated much of the
obs;xy;i conv;xy;j
computational work to GMT’s grdmath program. This ap-
proach provides us with very efficient computation, while
is the total variance at a given point. keeping most of our development in the convenient Perl
Because GMPEs for various tectonic regimes are rela- programming language.
tively more common than IPEs, we allow for virtual-IPEs that
are the combination of a GMPE and a GMICE. The variance Improvements over Previous Versions of ShakeMap. Read-
for the IPE in this case is determined by combining the var- ers familiar with previous versions of ShakeMap will recog-
iance of the GMPE with the variance of the GMICE. nize that this new implementation is a substantial departure
3088 C. B. Worden, D. J. Wald, T. I. Allen, K. Lin, D. Garcia, and G. Cua
Count
200 200
Count
Count
200
50 50 50
0 0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
Error (ln(Obs) − ln(Pred); percent−g) Error (ln(Obs) − ln(Pred); percent−g) Error (ln(Obs) − ln(Pred); percent−g)
Figure 3. Histograms of log PGA raw error [lnobservation lnprediction] for three cases: (a) unbiased GMPE, (b) biased GMPE,
and (c) ShakeMap predictions. std = standard deviation.
from the old. The output products and many of the support- function within surface. The new approach gives us com-
ing programs are essentially the same (with a few important plete control of the method.
differences), but the computational element is essentially all- • The previous use of a coarse grid of estimates could miss
new and based on the methodology described herein. Some near-fault details because of the low resolution of the es-
of the advantages of this new approach are listed as follows: timate grid. While this limitation was mitigated somewhat
• Our previous implementation of ShakeMap used a surface by the ability to use a finer grid of predictions near the
of continuous curvature, fit to the data and estimates, to source, our new approach utilizes estimates at every point
produce an interpolated grid. However, it is debatable as in the output grid, guaranteeing that source features will
to whether such a surface correctly represents the distribu- not be missed or aliased beyond the limitations imposed
tion of ground motion. Additionally, because the new by the output grid resolution.
approach interpolates each point independently of its • Similarly, the coarse grid of estimates could miss GMPE
neighbors, it does not impose curvature or artificial con- corner distances or other real features produced by the
straints on point-to-point continuity, as a surface-fitting in- model, which would be smoothed over by the fitting (even
terpolation does. simple interpolation would break down here). Again, by
• Similarly, with a surface fit, there is a tradeoff between computing estimates at every grid point, the new approach
fitting the observations precisely and generating high- will reflect any predicted features within the resolution of
frequency artifacts. This tradeoff is represented by the ten- the output grid.
sion parameter in GMT’s program surface. With the new • Uncertainty was computed using the distance from the
approach, a station’s data dominate the result near the sta- nearest station only (possibly modified by the unknown-
tion, but the approach does not employ a mathematical source aleatory), but the current approach reflects the
spline that could generate artifacts elsewhere. combined uncertainty from proximity to multiple nearby
• Also because of fitting with surface, the results at each observations, further constrained by the GMPE.
output point were dependent on the dimensions of the in-
terpolated grid, the data within the mapped area, and any 1.8
padding added to the gridded area. The new approach gen-
Mean Absolute Residual |ln(Obs) − ln(Pred)|
ShakeMap Prediction
1.6 Biased GMPE
erates a result at each grid point that is independent of grid Unbiased GMPE
size, mapped area, or any data beyond rmax .
1.4
• The prior surface fitting was something of a black box,
dependent on the specific implementation of the fitting 1.2
Table 1 1
Residual {abslnObserved lnPredicted} of
ShakeMap Predictions vs. the Biased and Unbiased 0.8
GMPE, Using PGM and Converted MMI Observations
0.6
Biased Unbiased
Parameter n ShakeMap* GMPE* GMPE*
0.4
PGA 3245 Mean 0.4571 0.5631 0.9347
Std† 0.3827 0.4583 0.6441 0.2
PGV 3105 Mean 0.4466 0.5391 0.7691 0 10 20 30 40 50 60
Std† 0.3679 0.4313 0.5656 Event
*Units for PGA are lnpercent-g, and for PGV are lncm=s. Figure 4. Mean PGA error for events with at least ten predic-
†
Std = standard deviation. tions, sorted by increasing mean error of ShakeMap predictions.
A Revised Ground-Motion and Intensity Interpolation Scheme for ShakeMap 3089
1.6
Table 3
Mean Absolute Residual |ln(Obs) − ln(Pred)|
ShakeMap Prediction
Biased GMPE
Unbiased GMPE
Summary Statistics of PGA Residual as a Function of the
1.4
Number of Nearby Stations Contributing to a Prediction
• Our previous version would cull predictions and converted n Samples is the number of predictions contributing to the mean in
observations near seismic stations in order to ensure that each bin.
the best-quality data would dominate the solution near
the observation. Now these data are combined with the As mentioned earlier, ShakeMap attempts to compute a
observations, via weighting by quality, utilizing all avail- GMPE bias factor for each event. If enough stations (config-
able information and reducing overall uncertainty. urable, but usually at least six) are near the source (i.e., with-
in the valid range of the GMPE), the event magnitude (as
provided to the GMPE) is adjusted to minimize the misfit
Discussion and Testing between the observed and predicted values. The definition
of “misfit” is also configurable, but here we have used the
We can test our approach by computing the ShakeMap absolute deviation (i.e., the L1 norm) of the difference in
predicted amplitudes at the location of ground-motion obser- log amplitudes. It should be noted that in southern California,
vations (rather than on a uniform grid), excluding the obser- from where our test data originate, the interevent bias is
vation itself, and then comparing our prediction to the usually well constrained by dense station coverage; it may
observation. This is a test of our ShakeMap methodology not perform as well in areas with sparse observations. Also
specifically and not of the accuracy of the maps themselves; as mentioned earlier, when a GMPE bias is available for an
in operation, ShakeMap produces estimates on a rectangular event, the uncertainty of the GMPE is reduced to the intrae-
grid, which would not, in general, be expected to be identical vent term (if so supplied), rather than the combined intrae-
vent and interevent terms used otherwise. This reduced
with the ground motion at a nearby station or stations.
uncertainty will have a very minor effect on the relative con-
However, the results discussed here give us confidence that
tribution of the predictions of the GMPE in the averaging pro-
our predictive grids will be better, in general, than GMPE or
cess (because the contributions of the observations are
IPE predictions, especially in the proximity of one or more
relative to the uncertainty of the GMPE; see equations 2a
seismic stations or intensity observations. and 5) but will reduce the computed uncertainty for every
output point in the map. In cases where a bias is computed,
Table 2 but separate intraevent and interevent GMPE uncertainty
Residual (abslnObserved lnPredicted) of ShakeMap terms are not available, the computed ShakeMap uncertainty
Predictions and the Biased and Unbiased GMPE, will be somewhat greater than the actual uncertainty as, to
Using PGM Observations Only some extent, the interevent uncertainty contribution has been
Biased Unbiased removed by the bias correction (although it cannot be in-
Parameter n ShakeMap* GMPE* GMPE* cluded mathematically).
PGA 3066 Mean 0.4576 0.5648 0.9503 In the process of computing the bias, ShakeMap finds
Std† 0.3847 0.4668 0.6501 and removes outliers (another configuration choice; here,
PGV 3069 Mean 0.4431 0.5389 0.7714 we consider an amplitude an outlier if it lies in excess of
Std† 0.3657 0.4322 0.5658
four standard deviations from the biased GMPE estimates).
*Units for PGA are lnpercent-g, and for PGV are lncm=s. Outliers are removed from the misfit calculation, and the bias
†
Std = standard deviation. is recomputed in an iterative process until no additional
3090 C. B. Worden, D. J. Wald, T. I. Allen, K. Lin, D. Garcia, and G. Cua
Count
Count
0 0 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
Error (Obs − Pred) Error (Obs − Pred) Error (Obs − Pred)
Figure 6. Histograms of intensity raw error (observation prediction) for three cases: (a) unbiased IPE, (b) biased IPE, (c) ShakeMap
predictions. std = standard deviation.
A Revised Ground-Motion and Intensity Interpolation Scheme for ShakeMap 3091
0.7
is consistent over other datasets, we may find it productive to ShakeMap Prediction
Biased IPE
devise a means of compensating for this bias.
if we substitute the GMPE of Boore and Atkinson (2008) affect the overall accuracy of the maps, and it therefore be-
(BA08) for CY08. These results are directly comparable hooves the ShakeMap operator to make a careful selection
to Table 2. While BA08 does somewhat less well than of GMPE. Those seeking guidance in this matter and in the
CY08 (except for the unbiased PGA residual), the difference matter of GMICE selection are referred to Cua et al., 2010,
is not dramatic, and we note that BA08 was developed for which reviews the options for a wide range of geographic
earthquakes of magnitude 5.0 and greater (unlike our CY08 areas and tectonic settings.
module, which has adjustments for small-magnitude to mod- In all cases previously mentioned, the GMICE was used
erate-magnitude earthquakes per Chiou et al., 2010), and our in the reverse manner for which it was designed, that is, we
dataset contains many earthquakes of M 4.0–M 5.0. As one used it to generate ground motions from known intensities,
might expect of a GMPE derived for larger events, the un- rather than intensities from known ground motions. Neither
biased BA08 tends to overpredict for our dataset. The choice of the GMICEs we use here (Wald, Quitoriano, Heaton, et al.,
of an inappropriate GMPE, however, can have greater conse- 1999; Atkinson and Kaka, 2007) was regressed in this direc-
quences. In areas of sparse data, ShakeMap predictions may tion, and our usage must be considered decidedly off label.
be largely based on the GMPE, and in some cases it may be We await with anticipation the development of GMICEs that
the unbiased GMPE. Even in cases where data-availability is are two-way (i.e., they intend both PGM-to-intensity and
sufficient to compute a bias, the bias procedure itself may not intensity-to-PGM conversions), however we find the current
respond well to an inappropriate GMPE and could produce equations to behave admirably under these less-than-ideal
spurious results. We reiterate that the choice of GMPE will circumstances. (We note that Faenza and Michelini [2010]
Figure 8. ShakeMap intensity map of the M 5.4 Diamond Bar earthquake. This map was made using the old ShakeMap methodology of
PGM converted to intensity. The small triangles on the map indicate the location of seismic stations reporting PGM.
A Revised Ground-Motion and Intensity Interpolation Scheme for ShakeMap 3093
do have a two-way relation for Italian intensity data, derived Wald, Quitoriano, Heaton, et al. (1999). (Note that here we
by orthogonal distance regression, but they employ Mercali- are using the GMICE as it was intended, although we use the
Cancani-Seiberg intensities and for a limited, lower range of IPE for magnitudes below the M 4.9–M 7.9 range for which
intensity values.) We do, however, restrict the use of PGM it was developed.) Figure 6 shows the residual histograms
derived from small intensities because our work here had in- of the unbiased (Fig. 6a) and biased IPE (Fig. 6b), and
dicated that these derived amplitudes are a significant source ShakeMap predictions (Fig. 6c). As with PGM, we see the
of error. This higher error is likely the result of the lower bias of the IPE, and the ShakeMap methodology leading to
intensities deriving from a wider range of PGMs than that tighter, more centered distributions. The unbiased IPE tends
of the higher intensities. This wider PGM range is apparent, to overpredict with the center of the distribution at a mean
for example, in figure 4 of Atkinson and Kaka (2007). position of (0:2235), while the biased IPE and ShakeMap
ShakeMap, therefore, only uses PGVs derived from intensi- predictions are near-zero at (0:0030) and (0:0023), re-
ties greater than or equal to 4, and PGAs from intensities 3 spectively. Table 5 presents the summary statistics for the
and larger (these thresholds were determined from our test- absolute error. Here again, we see a dramatic reduction in
ing). In addition, many GMICEs do not provide conversions the residual mean and variance with the implementation
for PSA, so ShakeMap PSA predictions are not always sup- of a per-event IPE bias and additional improvement through
plemented by nearby intensity observations. the use of the data-informed ShakeMap methodology. It is
Our results for macroseismic intensity (here we use notable, however, that even the unbiased IPE is able to predict
DYFI MMI-equivalent specifically) are similar to those for mean intensities within one-quarter intensity unit of the data,
PGM. In the case of intensity, however, we use linear rather with a moderate variance, suggesting that intensity is consid-
than logarithmic differences for our residual computations erably more spatially stable than PGM. This is perhaps not
because the MMI scale is constrained to a limited range of unexpected in that MMI is by definition an average of obser-
values from 1 to 10. Here, we chose the IPE developed by vations over some (variable) spatial extent. Finally, Figure 7
Allen et al. (unpublished manuscript), and the GMICE of presents the per-event mean prediction error for those events
with at least 10 data-informed amplitudes. As with PGM, the An interesting example of the value of native intensity
ShakeMap predictions are superior to the IPE for most of the approach over the converted PGM approach is found in the
events in our dataset. M 5.4 earthquake of 29 July 29 2008 near the city of
In our earlier discussion of the development of the new Diamond Bar, California (event ID 14383980). Figure 8
ShakeMap methodology, we asserted the importance of treat- shows the intensity map derived by the previous ShakeMap
ing intensity as a native data source similar to PGM, rather system, showing an area of intensity VI southwest of the
than as a derived parameter as had been our previous ap- epicenter running to the coast. This feature was much more
proach. We can test the merits of this decision by approximat- subtle and of lower intensity on the DYFI maps for the same
ing the approach of earlier versions of ShakeMap: we produce event (Fig. 9), and upon examination, we found that the high
PGM predictions at the locations of the intensity observations ShakeMap intensities were derived from PGVs that were
and then convert those ground motions into intensities. Table 6 likely the result of energy amplified within the basin but
presents the summary statistics from this approach, which of a low enough frequency that it did not generate commen-
may be compared directly with the statistics in Table 5. The surate increased macroseismic effects. Figure 10 shows the
converted PGM approach shows a marked inferiority to the map that the new system would create using both PGM and
native intensity approach, producing greater absolute error DYFI data. The revised map is much more consistent with the
and variance under all data-availability scenarios. observed intensities (Fig. 9).
Figure 10. ShakeMap intensity map of the M 5.4 Diamond Bar earthquake. This map was made using the new ShakeMap methodology
described in this paper. The small triangles on the map indicate the location of seismic stations reporting PGM, and the small circles show the
locations of intensity data from DYFI.
A Revised Ground-Motion and Intensity Interpolation Scheme for ShakeMap 3095