Apply MistNet segmentation to a polar volume

Source: R/apply_mistnet.R

Applies the MistNet segmentation model to a polar volume file on disk and loads the resultant segmentation as a polar volume (pvol) object.

apply_mistnet(
  file,
  pvolfile_out,
  verbose = FALSE,
  mount = dirname(file),
  load = TRUE,
  mistnet_elevations = c(0.5, 1.5, 2.5, 3.5, 4.5),
  local_install,
  local_mistnet
)

Arguments

file

Character. Path to a polar volume (pvol) file.

pvolfile_out

Character. (optional) File name. When provided, writes a polar volume (pvol) file to disk that includes the Mistnet segmentation results.

verbose

Logical. When TRUE, vol2bird stdout is piped to the R console.

mount

Character. Directory path of the mount point for the Docker container (deprecated).

load

Logical. When TRUE, returns a pvol object.

mistnet_elevations

Numeric vector of length 5. Elevation angles to feed to the MistNet segmentation model, which expects exactly 5 elevation scans at 0.5, 1.5, 2.5, 3.5 and 4.5 degrees. Specifying different elevation angles may compromise segmentation results.

local_install

Character. Path to local vol2bird installation (e.g. your/vol2bird_install_directory/vol2bird/bin/vol2bird) to use instead of the Docker container.

local_mistnet

Character. Path to local MistNet segmentation model in PyTorch format (e.g. /your/path/mistnet_nexrad.pt) to use instead of the Docker container.

Value

When load is TRUE, a polar volume (pvol) object with the Mistnet segmentation results. When load is FALSE, TRUE on success.

Details

MistNet (Lin et al. 2019) is a deep convolutional neural network that has been trained using labels derived from S-band dual-polarization data across the US NEXRAD network. Its purpose is to screen out areas of precipitation in weather radar data, primarily legacy data for which dual-polarization data are not available. Because the network has been trained on S-band data, it may not perform as well on C-band.

MistNet requires three single-polarization parameters as input: reflectivity (DBZH), radial velocity (VRADH), and spectrum width (WRADH), at 5 specific elevation angles (0.5, 1.5, 2.5, 3.5 and 4.5 degrees). Based on these data it can estimate a segmentation mask that identifies pixels with weather that should be removed when interested in biological data only.

MistNet will calculate three class probabilities (from 0 to 1, with 1 corresponding to a 100% probability) as additional scan parameters to the polar volume:

  • BACKGROUND: Class probability that no signal was detected above the noise level of the radar.

  • WEATHER: Class probability that weather was detected.

  • BIOLOGY: Class probability that biological scatterers were detected.

MistNet will calculate three class probabilities (from 0 to 1, with 1 corresponding to a 100% probability) as additional scan parameters to the polar volume:

  • BACKGROUND: Class probability that no signal was detected above the noise level of the radar

  • WEATHER: Class probability that weather was detected

  • BIOLOGY: Class probability that biological scatterers were detected

These class probabilities are only available for the 5 input elevations used as input for the MistNet model. Based on all the class probabilities a final weather segmentation map is calculated, stored as scan parameter CELL, which is available for all elevation scans.

  • CELL: Final weather segmentation, with values > 1 indicating pixels classified as weather and values equal to 1 indicating pixels that are located within 5 km distance of a weather pixels.

A pixel is classified as weather if the class probability WEATHER > 0.45 or when the average class probability for rain across all five MistNet elevation scans at that spatial location > 0.45.

MistNet may run more slowly on Windows than on Linux or Mac OS X.

References

Please cite this publication when using MistNet:

  • Lin T-Y, Winner K, Bernstein G, Mittal A, Dokter AM, Horton KG, Nilsson C, Van Doren BM, Farnsworth A, La Sorte FA, Maji S, Sheldon D (2019) MistNet: Measuring historical bird migration in the US using archived weather radar data and convolutional neural networks. Methods in Ecology and Evolution 10 (11), pp. 1908-22. doi:10.1111/2041-210X.13280

Examples

# \donttest{
# make sure you have installed the MistNet libraries and model, using:
if (requireNamespace("vol2birdR", quietly = TRUE)){
if(!vol2birdR::mistnet_exists()){
   vol2birdR::install_mistnet()
   vol2birdR::install_mistnet_model()
}
# start a temporary file to store polar volume
tempfile=tempfile("KBGM_example")
# Download a NEXRAD file and save as KBGM_example
download.file(
  "https://noaa-nexrad-level2.s3.amazonaws.com/2019/10/01/KBGM/KBGM20191001_000542_V06",
  method="libcurl", mode="wb", tempfile
)

# Calculate MistNet segmentation
mistnet_pvol <- apply_mistnet(tempfile)

# Print summary info for the segmented elevation scan at the 0.5 degree,
# verify new parameters BIOLOGY, WEATHER, BACKGROUND and CELL have been added
scan <- get_scan(mistnet_pvol, 0.5)
scan

# Project the scan as a ppi
ppi <- project_as_ppi(scan, range_max = 100000)

# Plot the reflectivity parameter
plot(ppi, param = "DBZH")

# Plot the MistNet class probability [0-1] for weather
plot(ppi, param = "WEATHER")

# Plot the MistNet class probability [0-1] for biology
plot(ppi, param = "BIOLOGY")

# Plot the final segmentation result, with values >1 indicating
# areas classified as weather, and value 1 pixels that fall within an
# additional 5 km fringe around weather areas
plot(ppi, param = "CELL")

# Remove file
file.remove(tempfile)
}
#> Filename = /tmp/RtmpYC4uIi/KBGM_example1a4511426481, callid = KBGM
#> Reading RSL polar volume with nominal time 20191001-000542, source: RAD:KBGM,PLC:BINGHAMTON,state:NY,radar_name:KBGM
#> Running vol2birdSetUp
#> Warning: radial velocities will be dealiased...
#> Warning: disabling single-polarization precipitation filter for S-band data, continuing in DUAL polarization mode
#> Warning: using MistNet, disabling other segmentation methods
#> Running segmentScansUsingMistnet.
#> Warning: Requested elevation scan at 1.500000 degrees but selected scan at 1.318360 degrees
#> Warning: Requested elevation scan at 3.500000 degrees but selected scan at 3.120115 degrees
#> Warning: Requested elevation scan at 4.500000 degrees but selected scan at 3.999025 degrees
#> Warning: Ignoring scan(s) not used as MistNet input: 2 4 8 9 ...
#> Running MistNet...done
#> [1] TRUE
# }