IRIS: DB_HCLASS returned as packed int16 words instead of per-bin uint8 classes

[aladinor]

Summary

For IRIS/Sigmet RAW files, DB_HCLASS (type 55, 8-bit hydrometeor class) is returned as unsplit int16 words — each element packs TWO 8-bit class values — instead of one class per range bin like the other 8-bit types.

MRE (public IDEAM file)

import fsspec, numpy as np
from xradar.io.backends.iris import open_iris_datatree

url = "s3://s3-radaresideam/l2_data/2024/06/01/Guaviare/GUA240601000308.RAWFXK0"
with fsspec.open(f"simplecache::{url}", s3={"anon": True}) as f:
    local = f.name

ds = open_iris_datatree(local)["sweep_0"].ds
print(ds["DB_HCLASS"].dtype, np.unique(ds["DB_HCLASS"].values)[:8])
# int16 [   0    9   17  105  106  108 2304 2313]

4369 == 0x1111 and 27756 == 0x6C6C are two packed uint8 classes. A byte-level decode of the same rays gives per-bin classes {0, 9, 17, 105–108}, and word == byte0 + 256*byte1 holds element-for-element over the first nbins // 2 columns — the rest of the range dim is not meaningful per-bin data.

Expected

Per-bin uint8 classes (the same 2-bins-per-word unpacking the scaled 8-bit types get). Likely cause: SIGMET_DATA_TYPES[55] has func=None, and the func-less path skips the pair unpacking.

xradar 0.12.1.dev9+gd723857f2 (main), Python 3.12.

[kmuehlbauer]

Thanks @aladinor, IIRC, I've left everything at the defaults, returning the full 2 bytes.

Reserved:

• 0 (0x00) for No data
• 255 (0x255) Area not scanned

1-byte of binary DB_HCLASS data is split into 3 bit segments as follows:

Each segment can be associated with an identification method (echo classifier). For a list of
the available classification methods listed as enumerated logical names, see the public
IRIS/RDA header sig_data_types.h.
The bytes Enums of the task_dsp_info structure (see task_dsp_info Structure
(page 56)) contain the list of classification methods allocated to each bit segment of the
HClass data. The information is associated to each IRIS RAW data file header, and can be
inspected with the productx utility.
The bits in each segment are interpreted as an unsigned integer value belonging to an
enumerated class type (the class base), specific to each classifier. The enumerated class
types are defined in sig_data_types.h with unambiguous logical names that are
meteorologically explicit.

So decoding isn't trivial, maybe the reason why I skipped it back then.