fit_ssm: Correlated Random Walk Filter

In ianjonsen/ssmTMB: A fast state-space model for filtering Argos satellite tracking data using TMB

Description

Correlated Random Walk Filter

Usage

fit_ssm(d, subset = rep(TRUE, nrow(d)), tstep = 2.4/24, nu = 5,
  gamma = 0.5, soth = TRUE, common.tau = FALSE, parameters = NULL,
  optim = c("nlminb", "optim"), verbose = FALSE, amf = amfCRAWL(),
  span = 0.3)

Arguments

d	a data frame of observations
subset	a logical vector indicating the subset of data records to be filtered
tstep	the time step to predict to (in days)
nu	degrees of freedom parameter
gamma	the autocorrelation parameter used to estimate initial parameters
soth	if TRUE states occur on the hour, starting with the hour immediately preceeding the time of the first observation; if FALSE states occur relative to the time of the first observation
common.tau	estimate a single error sd for both longitude and latitude
parameters	the TMB parameter list
optim	numerical optimizer
verbose	report progress during minimization
amf	ARGOS error multiplication factors
span	the span parameter for the loess fits used to estimate initial locations

Details

Fit a correlated random walk to filter a track and predict locations on a regular time step.

The input track is given as a dataframe where each row is an observed location, with columns

'date'

observation time (as GMT POSIXct),

'lon'

observed longitude,

'lat'

observed latitude,

'lc'

location class.

The TMB parameter list can be specified directly with the parameters argument. Otherwise suitable parameter values are estimated from predicted locations estimated by fitting loess smooths to the raw observations.

The filtering model assumes the errors in longitude and latitude are proportional to scale factors determined by the location class. The scale factors are specified through the amf argument. By default the function uses the same scaling factors for location accuracy as used in the R package crawl for ARGOS data.

Value

a list with components

predicted	a data.frame of predicted location states
fitted	a data.frame of fitted locations
par	model parameter summmary
data	the input data.frame
subset	the input subset vector
tstep	the prediction time step
common.tau	has a common tau been estimated
opt	the object returned by the optimizer
tmb	the TMB object
aic	the calculated Akaike Information Criterion

Examples

# Fit DCRW model for state filtering and regularization
data(ellie)
fit <- fit_ssm(ellie, tstep = 6 / 24)

## plot predicted track over observations
plot(lat ~ lon, fit$data, pch = 3, col = "firebrick")
points(lat ~ lon, fit$predicted, pch = 19, cex = 0.5, col = "dodgerblue")

## plot residuals for longitude and latitude
layout(matrix(1:2, 2, 1))
lon.res <- fit$data$lon - fit$fitted$lon
lat.res <- fit$data$lat - fit$fitted$lat
plot(fit$data$date, lon.res, xlab = "")
abline(h = 0, lty = 2, col = "red")
plot(fit$data$date, lat.res, xlab = "date")
abline(h = 0, lty = 2, col = "red")

ianjonsen/ssmTMB documentation built on July 5, 2019, 3:03 p.m.