dasio

Lightweight, standalone IO + basic processing for DAS (Distributed Acoustic Sensing) data: vendor HDF5 readers (ASN/OptoDAS, OptaSense/QuantX, AP Sensing), a Proc concatenate/downsample format, a Basic format for converted analysis products, an Event format, a numpy DASdata container, a file catalog (DASdb), and signal processing including a C++/OpenMP bandpass.

Install (development)

This package contains a compiled C++ extension. Always install editable so the extension builds in place:

pip install -e .

Build requirements: a C++14 compiler, CMake ≥3.15, and OpenMP (e.g. apt-get install build-essential libomp-dev). pybind11 and scikit-build-core are pulled automatically by the build.

Optional extras

Ambient-noise cross-correlation and phase picking have heavier, optional dependencies, exposed as install extras (still editable):

pip install -e '.[viewer]'      # interactive viewer (dasio.viewer.view) -> fastplotlib, ipywidgets
#                                 run it from a Jupyter kernel you already have
pip install -e '.[noise]'       # ambient-noise cross-correlation (dasio.noise) -> PyTorch
pip install -e '.[pick]'        # PhaseNet-DAS P/S picking (dasio.pick_phases) -> phasenet (+ PyTorch)
pip install -e '.[noise,pick]'  # both

Both are imported lazily, so the rest of dasio works without them installed.

Quickstart

There is a Jupyter notebook to read, trouble shot and processing data from Silixa iDAS.

One file:

from dasio import DASFile

d = DASFile('file.h5').read().to_physical()          # format auto-detected -> microstrain (or /s)
d.bandpass(1.0, 10.0).subtract_common_mode().plot()  # filter -> denoise -> waterfall

A directory, queried by time — DASdb catalogues the files once, then read stitches any window across file boundaries:

from datetime import datetime, timedelta, timezone
from dasio import DASdb

db = DASdb.from_dir('/data/das/100Hz')       # scan once; format auto-detected
db.to_file('dasdb.csv')                      # DASdb.from_file() next time

t0 = datetime(2025, 1, 1, 0, 0, tzinfo=timezone.utc)
d = db.read(t0, t0 + timedelta(seconds=120)) # concatenated, gaps zero-filled
d = d.to_physical()                          # vendor units -> microstrain (or /s)

Readers return the instrument's own units (OptaSense counts, AP Sensing radian/s, ASN strain/s); to_physical() is the one conversion step, and a no-op if already converted. Raw OptaSense can roll over at 2**32 — call d.unwrap() first, on the concatenated window rather than per file.

Command line

Three tools, all run with python -m, all with --help.

Catalog a directory of DAS files. Run the same command again later and it adds only what is new:

python -m dasio.dasdb --from /data/das/100Hz --dasdb dasdb.parquet

Downsample to 25 Hz. --fmax is the low-pass corner in Hz and sets the output rate with it, at about 2.5x fmax:

python -m dasio.desample --from /data/das/raw --to /data/das/25Hz --fmax 10 \
    --dasdb raw.parquet --all --nworkers 4 --nthreads 6

--all processes every window missing from --to. A window counts as done when its output file exists, so a stopped job restarts by re-running the same command. After an unclean kill, delete the partial file first: it sits at its final path and counts as done. To pick one window instead, give --since and --until together, with an explicit UTC offset.

Cut one file per event from an event catalog CSV:

python -m dasio.extract_events --catalog events.csv --dasdb dasdb.parquet \
    --to /data/das/events