SoilSpecData

Load soil spectral data, such as the Open Soil Spectral Library (OSSL) and Thermo OMNIC .spa files, as spectra and soil properties ready for machine learning.

OSSL is an open database of soil spectra, with the lab-measured soil properties of the same samples. Version 1.2 combines 10 spectral libraries, from the US KSSL library (86,603 samples) and the European LUCAS survey (40,175) to smaller regional ones. It holds 135,651 samples, with mid-infrared (MIR) spectra for 85,684 of them and visible and near-infrared (VISNIR) spectra for 64,644. OSSL publishes the whole database as one CSV file of about 1 GB.

SoilSpecData downloads and caches that file, and gives the spectra in wavenumbers. It aligns spectra with soil properties by sample. One line then gives you X and y for a model. The OSSL documentation describes the database and how OSSL builds it.

The SoilSpecData documentation covers the full API.

Installation

pip install -U soilspecdata

To install the development version, with the development tools nbdev and twine, run this in the project root:

pip install -e .[dev]

Features

• Download and cache the OSSL database, at data level L0 or L1
• VISNIR (visible and near-infrared) and MIR (mid-infrared) spectra, in wavenumbers, for any wavenumber range
• Soil properties and metadata as pandas DataFrames, indexed by sample ID
• Spectra and target variables aligned by sample, ready for machine learning
• Spectra from Thermo OMNIC .spa instrument files, interpolated onto a shared grid with replicates averaged, as a Spectra
• Further datasets to come …

Quick start

from soilspecdata.all import * imports load_ossl, OSSL, Spectra, get_cache_path, read_spa and load_spa. Each also has its own module, such as soilspecdata.datasets.ossl, for explicit imports. Load OSSL, then get the MIR spectra and CEC values of every sample that has both. The first call downloads about 1 GB. Later calls read the cached file instead. Loading took 24 seconds on an Apple silicon Mac. Memory use peaks at about 12 GB while pandas reads the file. The loaded data then takes 3.5 GB:

from soilspecdata.all import *

ossl = load_ossl()

X, y, ids = ossl.mir().xy('cec_usda.a723_cmolc.kg')
X.shape, y.shape

((57062, 1701), (57062, 1))

Loading the data

load_ossl() loads the L0 data. L0 keeps the lab method of each source dataset. load_ossl(level='L1') loads the L1 data. L1 converts properties measured with different lab methods to one target method.

load_ossl caches each downloaded file in ~/.soilspecdata/, under the file’s own name, such as ossl_all_L0_v1.2.csv.gz. To download the file again:

ossl = load_ossl(force_download=True)

Spectra

OSSL MIR spectra cover 600 to 4000 cm⁻¹. VISNIR spectra cover 350 to 2500 nm, or 4000 to 28571 cm⁻¹. mir and visnir return both in wavenumbers, in increasing order. They return only the samples with a value at every wavenumber in the range.

Their wmin and wmax parameters select a wavenumber range. By default, mir returns 600 to 4000 cm⁻¹, which covers all 85,684 MIR spectra. By default, visnir returns 4000 to 25000 cm⁻¹ (400 to 2500 nm). That range covers all 64,644 VISNIR spectra. Only 23,880 of them start at 350 nm.

The MIR spectra over the default range:

mir_data = ossl.mir()

The VISNIR spectra between 8000 and 25000 cm⁻¹:

visnir_data = ossl.visnir(wmin=8000, wmax=25000)

Both methods return a Spectra, with these fields:

• wavenumbers: wavenumbers in cm⁻¹, in increasing order
• spectra: one row per sample, one column per wavenumber
• measurement_type: ref (reflectance) or abs (absorbance)
• sample_ids: the sample ID of each row

Printing a Spectra shows the start of each field:

print(visnir_data)

Spectra attributes:
------------------
Available attributes: wavenumbers, spectra, measurement_type, sample_ids

Wavenumbers:
-----------
[8000, 8013, 8026, 8039, 8052, 8065, 8078, 8091, 8104, 8117]
Shape: (426,)

Spectra:
-------
[[0.4707, 0.4702, 0.4697, 0.4692, 0.4687],
 [0.6087, 0.6081, 0.6074, 0.6068, 0.6062],
 [0.6257, 0.625 , 0.6243, 0.6237, 0.6231],
 [0.6598, 0.6589, 0.6581, 0.6574, 0.6567],
 [0.6655, 0.6644, 0.6634, 0.6624, 0.6615]]
Shape: (64644, 426)

Measurement type (Reflectance or Absorbance):
--------------------------------------------
ref

Sample IDs:
----------
['594ef6c549138844cf79a0fece6de7d2', '7417caf4328373c2f6b2977817f508b6',
 'e01efc61d0222fc5ca44656483dfa665', '5c1dcddd9cfe540857cc2bfea3d9a44b',
 '393ab838badfeb6e12520f09c83e479b', 'c8f162486487b8525403e55c00d60cfb',
 'e3b9c5d0e649e49726e1fed397f18417', 'c48a2a1f0cc04b26537e0c653530cbf2',
 'bf2ea7ea2964e7d0bd42443e095d1b56', '2507c34c14815cb6ec303e5bf9156072']
Total samples: 64644

The first 20 MIR spectra, plotted with matplotlib:

from matplotlib import pyplot as plt

plt.figure(figsize=(12, 3))
plt.plot(mir_data.wavenumbers, mir_data.spectra[:20,:].T, alpha=0.3, color='steelblue', lw=1)
plt.gca().invert_xaxis()
plt.grid(True, linestyle='--', alpha=0.7)

plt.xlabel('Wavenumber (cm⁻¹)')
plt.ylabel('Absorbance');

Soil properties and metadata

Selecting columns of ossl returns a pd.DataFrame, indexed by sample ID (id). Sample IDs are OSSL’s id.layer_uuid_txt values. They are unique across all source datasets. dropna keeps the samples that have a value. The CEC (cation exchange capacity) of every sample that has one, in cmolc/kg:

properties = ossl[['cec_usda.a723_cmolc.kg']].dropna()

properties.head()

Any column in ossl.meta_cols can serve as a target or as metadata. Soil-lab column names follow the pattern <property>_<method>_<unit>. For example, cec_usda.a723_cmolc.kg is the CEC measured with USDA method A723, in cmolc/kg. The OSSL database description describes all the variables. To find the columns for one property, filter on its name:

[c for c in ossl.meta_cols if c.startswith('cec')]

['cec_iso.11260_cmolc.kg', 'cec_usda.a723_cmolc.kg']

Metadata such as coordinates can have missing values. Without dropna, the selection keeps every sample:

metadata = ossl[['longitude.point_wgs84_dd', 'latitude.point_wgs84_dd']]

Selecting samples

ossl behaves like a pandas DataFrame indexed by sample ID. Filter samples with ordinary pandas conditions on its columns. The result supports every method above. The KSSL samples only:

kssl = ossl[ossl['dataset.code_ascii_txt'] == 'KSSL.SSL']
len(kssl), len(ossl)

(86603, 135651)

Training data

xy returns the spectra X, the targets y and the sample IDs ids of the samples that have both, in the same row order. It accepts several targets, and works on any selection of samples. The KSSL samples with both CEC and organic carbon values:

X, y, ids = kssl.mir().xy(['cec_usda.a723_cmolc.kg', 'oc_usda.c729_w.pct'])

X.shape, y.shape, ids.shape

((50825, 1701), (50825, 2), (50825,))

Upgrading from 0.0.x

Version 0.1.0 changed the API. To update code written for version 0.0.x:

• get_ossl(url, force_download, level) is now load_ossl(level, url, force_download).
• get_mir and get_visnir are now mir and visnir.
• ossl.get_aligned_data(spectra, targets) is now spectra.xy(targets).
• ossl.get_properties(cols) is now ossl[cols]. Add .dropna() for require_complete=True.
• OSSLData, SpectraData and properties_cols are now OSSL, Spectra and meta_cols.
• ossl.df is indexed by sample ID (id.layer_uuid_txt) instead of by row number.

Version 0.1.0 still accepted the old names. Version 0.2.0 removes them. If your code runs on 0.1.0 without a DeprecationWarning, and doesn’t use OSSLData, SpectraData, properties_cols or get_properties, it runs on 0.2.0 unchanged.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

SoilSpecData is under the Apache 2.0 licence.

OSSL publishes its data under CC BY 4.0, unless a dataset states otherwise. The dataset.license.title_ascii_txt column gives the licence of each source dataset. In v1.2, three datasets have their own licence:

• LUCAS.SSL and LUCAS.WOODWELL.SSL: JRC License Agreement
• AFSIS1.SSL: ODC Open Database License
• AFSIS2.SSL: CC0

Check these terms before you publish or share results that use those datasets.

Citation(s)

• OSSL Library: Safanelli, J.L., Hengl, T., Parente, L.L., Minarik, R., Bloom, D.E., Todd-Brown, K., Gholizadeh, A., Mendes, W. de S., Sanderman, J., 2025. Open Soil Spectral Library (OSSL): Building reproducible soil calibration models through open development and community engagement. PLOS ONE 20, e0296545. https://doi.org/10.1371/journal.pone.0296545
• SoilSpecData: Albinet, F., 2026. SoilSpecData, version 0.2.0. https://github.com/franckalbinet/soilspecdata