65 Habitat Occupancy Models
Description: Habitat Occupancy
Found in: State of the Ecosystem - Gulf of Maine & Georges Bank (2018), State of the Ecosystem - Mid-Atlantic (2018)
Indicator category: Database pull with analysis; Extensive analysis; not yet published; Published methods
Contributor(s): Kevin Friedland
Data steward: Kevin Friedland, kevin.friedland@noaa.gov
Point of contact: Kevin Friedland, kevin.friedland@noaa.gov
Public availability statement: Source data are available upon request (see Survdat, CHL/PP, and Data Sources below for more information). Model-derived time series are available here.
65.1 Methods
Habitat area with a probability of occupancy greater than 0.5 was modeled for many species throughout the Northeast Large Marine Ecosystem (NE-LME) using random forest decision tree models.
65.1.1 Data sources
Models were parameterized using a suite of static and dynamic predictor variables, with occurrence and catch per unit effort (CPUE) of species from spring and fall Northeast Fisheries Science Center (NEFSC) bottom trawl surveys (BTS) serving as response variables. Sources of variables used in the analyses are described below.
65.1.1.1 Station depth
The NEFSC BTS data included depth observations made concurrently with trawls at each station. Station depth was a static variable for these analyses.
65.1.1.2 Ocean temperature and salinity
Sea surface and bottom water temperature and salinity measurements were included as dynamic predictor variables in the model, and were collected using Conductivity/Temperature/Depth (CTD) instruments. Ocean temperature and salinity measurements had the highest temporal coverage during the spring (February-April) and fall (September-November) months. Station salinity data were available between 1992-2016.
65.1.1.3 Habitat descriptors
A variety of benthic habitat descriptors were incorporated as predictor variables in occupancy models (Table 65.1). The majority of these parameters are based on depth (e.g. BPI, VRM, Prcury, rugosity, seabedforms, slp, and slpslp). The vorticity variable is based on current estimates, and the variable soft_sed based on sediment grain size.
Variables | Notes | References |
---|---|---|
Namera_vrm | Vector Ruggedness Measure (VRM) measures terrain ruggedness as the variation in three-dimensional orientation of grid cells within a neighborhood based on The Nature Conservancy Northwest Atlantic Marine Ecoregional Assessment (“NAMERA”) data. | Hobson (1972); Sappington, Longshore, and Thompson (2007) |
Prcurv (2 km, 10 km, and 20 km) | Benthic profile curvature at 2km, 10km and 20 km spatial scales was derived from depth data. | Winship et al. (2018) |
Rugosity | A measure of small-scale variations of amplitude in the height of a surface, the ratio of the real to the geometric surface area. | Friedman et al. (2012) |
seabedforms | Seabed topography as measured by a combination of seabed position and slope. | [http://www.northeastoceandata.org/ |
Slp (2 km, 10 km, and 20 km) | Benthic slope at 2km, 10km and 20km spatial scales. | Winship et al. (2018) |
Slpslp (2 km, 10 km, and 20 km) | Benthic slope of slope at 2km, 10km and 20km spatial scales | Winship et al. (2018) |
soft_sed | Soft-sediments is based on grain size distribution from the USGS usSeabed: Atlantic coast offshore surficial sediment data. | [http://www.northeastoceandata.org/ |
Vort (fall - fa; spring - sp; summer - su; winter - wi) | Benthic current vorticity at a 1/6 degree (approx. 19 km) spatial scale. | Kinlan et al. (2016) |
65.1.1.4 Zooplankton
Zooplankton data are acquired through the NEFSC Ecosystem Monitoring Program (“EcoMon”). For more information regarding the collection process for these data, see Kane (2007), Kane (2011), and Morse et al. (2017). The bio-volume of the 18 most abundant zooplankton taxa were considered as potential predictor variables.
65.1.1.5 Remote sensing data
Both chlorophyll concentration and sea surface temperature (SST) from remote sensing sources were incorporated as static predictor variables in the model. During the period of 1997-2016, chlorophyll concentrations were derived from observations made by the Sea-viewing Wide Field of View Sensor (SeaWIFS), Moderate Resolution Imaging Spectroradiometer (MODIS-Aqua), Medium Resolution Imaging Spectrometer (MERIS), and Visible and Infrared Imaging/Radiometer Suite (VIIRS).
65.1.2 Data processing
65.1.2.1 Zooplankton
Missing values in the EcoMon time series were addressed by summing data over five-year time steps for each seasonal time frame and interpolating a complete field using ordinary kriging. Missing values necessitated interpolation for spring data in 1989, 1990, 1991, and 1994. The same was true of the fall data for 1989, 1990, and 1992.
65.1.2.2 Remote sensing data
An overlapping time series of observations from the four sensors listed above was created using a bio-optical model inversion algorithm (Maritorena et al. 2010). Monthly SST data were derived from MODIS-Terra sensor data (available here).
65.1.2.3 Ocean temperature and salinity
Date of collection corrections for ocean temperature data were developed using linear regressions for the spring and fall time frames; standardizing to collection dates of April 3 and October 11 for spring and fall. No correction was performed for salinity data. Annual data for ocean temperature and salinity were combined with climatology by season through an optimal interpolation approach. Specifically, mean bottom temperature or salinity was calculated by year and season on a 0.5° grid across the ecosystem, and data from grid cells with >80% temporal coverage were used to calculate a final seasonal mean. Annual seasonal means were then used to calculate combined anomalies for seasonal surface and bottom climatologies.
An annual field was then estimated using raw data observations for a year, season, and depth using universal kriging (Hiemstra et al. 2008), with depth included as a covariate (on a standard 0.1° grid). This field was then combined with the climatology anomaly field and adjusted by the annual mean using the variance field from kriging as the basis for a weighted mean between the two. The variance field was divided into quartiles with the lowest quartile assigned a weighting of 4:1 between the annual and climatology values. The optimally interpolated field at these locations was therefore skewed towards the annual data, reflecting their proximity to actual data locations and associated low kriging variance. The highest kriging variance quartile (1:1) reflected less information from the annual field and more from the climatology.
65.1.3 Data analysis
65.1.3.1 Occupancy models
Prior to fitting the occupancy models, predictor variables were tested for multi-collinearity and removed if found to be correlated. The final model variables were then chosen utilizing a model selection process as shown by Murphy, Evans, and Storfer (2010) and implemented with the R package rfUtilities
(Evans and Murphy 2018). Occupancy models were then fit as two-factor classification models (absence as 0 and presence as 1) using the randomForest
R package (Liaw and Wiener 2002).
65.1.3.2 Selection criteria and variable importance
The irr
R package (Gamer, Lemon, and Singh 2012) was used to calculate Area Under the ROC Curve (AUC) and Cohen’s Kappa for assessing accuracy of occupancy habitat models. Variable importance was assessed by plotting the occurrence of a variable as a root variable versus the mean minimum node depth for the variable (Paluszynska and Biecek 2017), as well as by plotting the Gini index decrease versus accuracy decrease.
catalog link No associated catalog page