1. Exploratory analysis


1.1 Data Inputs

NOAA SEFSC visual data goes back to 1992, but as shown in the figure below, many predictor variables are only available starting in 2003, therefore earlier visual data is currently excluded from further analyses.

Note: Future work could use monthly climatologies (averages) so that older sightings data could be used. Some dynamic drivers like eddy and front locations would not be able to be considered using that approach.

Visual data predictor variable availability:


1.1.1 Splitting into testing and training sets

The data are split into training and testing sets. In this case, visual data from 2009 and acoustic data from 2013 were used only for testing. Only observations from 2003 or later were used for modeling due to covariate limitations.


1.1.2 Map of visual sightings data

The visual data selected for modeling are displayed on the map below. Data from 2009 were held back for testing. Blue markers indicate HARP locations.


1.1.3 Time series of acoustic data

The time series below show timeseries of estimated densities from passive acoustic data used for modeling (Densities were calculated following methods detailed in [1]). Data from 2011 and 2012 were used for training, and 2013 data was held back for testing.


Acoustic Timeseries:


1.2 Examination of covariates


1.2.1 Covariate distribution check


Distributions of covariates from acoustic observations (training data only):


Distributions of covariates from the visual observations (training data only):


Some of these covariates are more or less interrelated. Correlations are examined in the figure below. Numbers closer to 1 above the diagonal in the figure below represent correlation coefficients. If a pair of covariates is highly-correlated only one should typically be used in the model.


Covariate Correlations:


1.2.2 Transformation of predictor variables

Some variables, including chlorophyll, mixed layer depth and distance to fronts are highly skewed and were log-transformed for input to GAMs.

Below, the two sets of covariates have been combined and transformed:


1.2.3 Preliminary check of predictive power

To get an idea of the basic predictive power of these covariates, we can look at presence/absence relative to each variable. This also provides an opportunity to look at the range of values observed for each covariate in the visual and acoustic datasets. In the plots below dotted lines indicate the distribution of each covariate when Stenella spp. were present, and solid lines indicate the distribution when Stenella spp. were absent. Note that these plots do not account for effort.


Acoustic kernel densities:



Visual kernel densities:


1.2.4 Estimation of relative weights

To train the model, we need to know how much power the various data points have relative to one another. This is important because the duration, spatial coverage, and detection probabilities are quite different between the visual and acoustic data sets. If an animal is seen or heard, we know for certain that the species was present. However, if it was not heard, then it either wasn’t present, or it was present but missed.

“Zero inflation” or an excess of false zeros more common in the visual survey data because each data point represents a 10km transect section, traversed at survey speed (>10 knots) or approximately 30 minutes of observation effort. In contrast, the acoustic data are binned by day with a stationary instrument, therefore the probability of missing a group over the course of a day is lower.

For each data type, we estimated the probability of a missed detection to account for differences in zero-inflation, downweighting zeros according to the probability of a recording false negative.

The visual data represent wether or not Stenella spp. were seen during each transect segment. The probability of missing a sighting of Stenella spp. was estimated as

\[P_{V}(detect|present) = \mu_{det} * g0\ = 0.28 * 0.94 = 0.26\]

where _{det} is the mean detection probability as estimated by a model fit using the mrds package, and g0 is the probability of observing an animal on the transect line [2]. We assume that reported absences are likely to be true absences 26 % of the time, therefore zeros are given a weight of 0.26 on a scale of [0,1].

The acoustic data represent presence or absence of Stenella spp. in one day bins. Given that a group of animals is present near the sensor, the probability of detecting them in a 5 minute period within a 5 km range is estimated at 0.15, therefore the probability of missing an encounter is 1 - 0.15 = 0.85 [1,3]. Given that animals were present, the probability of missing a group for a full day (288 5-minute periods) is estimated as \[P_{A}(detect|present) = (1-0.15)^{288} \approx 1\]

Therefore we assume that there are no false negative days in the passive acoustic timeseries, and all acoustic observations are given weight = 1.


Best visual detection probability model for Stenella spp.:


2. Model Fitting

Models were fit using avnnet from the caret package in R.


Case weights:

## png 
##   2


2.1 Run Models

Run NNs Acoustic only, Visual only, and joint Acoustic/Visual datasets.

Models have the following characteristics:

  • R trainRepeats averaged repeats with random node initalization

  • Include 9 covariates

  • One hidden layer

  • Weighted training data

  • Hidden node layer sizes from 4 to 14 were tested in 2 node increments to search for optimal network size.




2.2 Model Comparisons

Models were compared using mean absolute error (MAE) to compare predicted and observed density in the test data.

## [1] "mean absolute error scores (lower is better)"
##                                  4       6       8     10      12      14
## Acoustic - Train             0.056   0.053   0.052   0.05   0.049   0.048
## Acoustic - Test            254.073 265.437 264.803 260.77 262.911 255.632
## Visual - Train              79.855  76.134  72.379  69.90  67.237  68.301
## Visual - Test              121.369 114.219 113.643 107.56 110.774 116.179
## Joint - Train              137.384 132.202 127.227 123.99 122.903 120.547
## Joint - Test               253.221 251.841 252.712 250.62 254.130 260.170
## Acoustic - Joint test data 296.520 297.297 304.122 305.49 309.558 309.689
## Visual - Joint test data   250.808 250.274 242.167 244.50 244.522 246.760
## [1] "Best acoustic model has 4 nodes."
## [1] "Best visual model has 8 nodes."
## [1] "Best joint model has 10 nodes."


2.3 Variable Importance

For the best model in each category, the importance of each input variable was calculated across the 50 model iterations.

AcOnly VisOnly Joint
SST 18.6 9.9 17.3
SSH 20.1 9.8 13.6
log10_CHL 22.6 12.2 17.0
log10_HYCOM_MLD 3.4 13.2 5.0
HYCOM_SALIN_0 8.6 13.7 12.9
log10_HYCOM_MAG_0 8.3 9.1 10.2
HYCOM_UPVEL_50 4.4 10.2 6.3
Neg_EddyDist 5.4 12.2 6.7
Pos_EddyDist 8.6 9.7 11.0


Example network:


3. Model Predictions

3.1 Temporal predictions

Predictions were made on the acoustic test dataset, and compared with actual observations for 2013. The predicted density of animals at each site was compared with the estimated daily density from the pasive acoustic record.


3.1.1 Acoustic-only prediction

Predicted and observed densities at passive acoustic monitoring sites using the acoustic-only model:


3.1.2 Visual-only prediction

Predicted and observed densities at passive acoustic monitoring sites using the visual-only model:


3.1.3 Joint prediction

Predicted and observed encounter probabilities at passive acoustic monitoring sites using the joint model:


3.2 Spatial predictions

Models were evaluated for summer (July 2009) and winter(January 2009) across the entire Gulf of Mexico (US EEZ beyond the 200m contour).

Note: As an alternative to training joint models, an average surface was computed across the acoustic and visual-only predictions.

Each grid cell in the following maps represents a 10x10 km square. Densities are therefore shown as estimated number of animals per 100 km2.


3.2.1 Summer 2009 predictions

Summer 2009 predicted distribution and test sightings:

Summer 2009 prediction uncertainty:


3.2.2 Winter 2009 predictions

Winter 2009 predicted distribution:

Winter 2009 prediction uncertainty:


3.3.3 How does hidden layer size change spatial predictions?

4. Monthly model predictions

Spatial model predictions were generated using climatological means of oceanographic variables, averaged by month between 2003 and 2015.

1. Frasier KE, Wiggins SM, Harris D, Marques TA, Thomas L, Hildebrand JA. Delphinid echolocation click detection probability on near-seafloor sensors. The Journal of the Acoustical Society of America. 2016;140: 1918–1930. doi:http://dx.doi.org/10.1121/1.4962279

2. Palka DL. Summer abundance estimates of cetaceans in us north atlantic navy operating areas. Northeast Fish Sci Cent Ref Doc. 2006; 06–03.

3. Hildebrand J, Baumann-Pickering S, Frasier K, Tricky J, Merkens K, Wiggins S, et al. Passive acoustic monitoring of beaked whale densities in the gulf of mexico during and after the deepwater horizon oil spill. Nature Scientific Reports. 2015;5: 16343.