mahynski/crassostrea-virginica-chem

Summary

This is a subset of results from NOAA's MusselWatch Program. Only Crassostrea virginica (American Oysters) have been selected from 1986-2021. The figure below summarizes the number of measurements made in each year.

A "measurement" is an entry (row) in raw.csv and corresponds to a single analyte being tested in a single sample. Each sample has a unique Sample_ID and it is possible for a given sample to have the same analyte tested for multiple times. Thus, the data in raw.csv has been condensed by Sample_ID to form more user-friendly datasets.

This dataset contains 184,740 measurements spanning 1,879 unique Sample_IDs and 442 unique analytes.

Metadata

The file raw.csv contains the following metadata on each measurement.

Name	datatype	Description
Study	string	Name of study (all "Mussel Watch")
NST_Site	string	Short name for sampling site
General_Location	string	General location of sampling site
Specific_Location	string	More detailed location of sampling site
Sample_ID	string	Unique identifier for a sample
Latitude	float	Latitiude of sampling site
Longitude	float	Longitude of sampling site
Fiscal_Year	int	Year sample was collected
Matrix	string	Description of what was sampled
Scientific_Name	string	Binomial nomenclature designation (all "C. virginica")
Collection_Date	date	YYYY-MM-DD of sampling
Method	string	Analytical chemistry method used
Laboratory	string	Name of laboratory that performed the analysis

Each measurement is also accompanied by a description with the following fields:

Name	datatype	Description
Parameter	string	Name of the analyte being measured
Value	float	Level or amount of analyte found
Unit	string	Units value is reported in
Qualifier	string	Comments about measurement (e.g., if below the limit of detection)

Units

Refer to to units.json for a list of units for each analyte. An analyte is reported in the same units across the entire dataset.

Notes

When a Qualifier is NaN this means there were no issues with the measurement; anything potentially problematic with the measurement is reported here, if present.

Duplicates

The following Sample_IDs have duplicate measurements reported in raw.csv. The processed datasets averaged the duplicate measurements.

1. 'MW2015CHBLCV'
2. 'MW2017AESPCV'
3. 'MW2017CBJBCV'
4. 'MW2017CKBPCV'
5. 'MW2017GBHRCV'
6. 'MW2017VBSPCV'

The file condensed.csv is formed by merging raw.csv by Sample_ID and averaging these duplicates. This has N = 1,879 unique samples and P = 442 unique analytes.

Sparsity

The majority of the analytes are not measured for all samples. The figure below is a missingno snapshot of the dataset.

Observe that only a small number of analytes are missing less than 10% of the time.

The user may create more advanced interpolation or data recovery / filling schemes starting with condensed.csv, but we avoided this for the training datasets provided here. In these cases, we simply dropped all rows with any NaN (missing) values. As a result, including many analytes as features (columns) increases the likelihood a row will be dropped. Increasing the threshold, t, in the above figure ("Fraction of Samples where Analyte is Missing") tends to decrease the number of rows, N, and increase the number of columns, P. The figures below show this tradeoff.

We selected 2 thresholds (A) t = 0.14 and (B) t = 0.26 (shown above) as a fair balance, which resulted in the following analytes:

(A): t = 0.14 has 26 analytes: 2,4'-DDD, 2,4'-DDE, 2,4'-DDT, 4,4'-DDD, 4,4'-DDE, 4,4'-DDT, Aldrin, Alpha-Chlordane, Arsenic, Cadmium, Chromium, Copper, Dieldrin, Gamma-Hexachlorocyclohexane, Heptachlor, Heptachlor-Epoxide, Hexachlorobenzene, Iron, Lead, Mercury, Mirex, Nickel, Selenium, Silver, Tin, Trans-Nonachlor

(B): t = 0.26 has 76 analytes: 1,6,7-Trimethylnaphthalene, 1-Methylnaphthalene, 1-Methylphenanthrene, 2,4'-DDD, 2,4'-DDE, 2,4'-DDT, 2,6-Dimethylnaphthalene, 2-Methylnaphthalene, 4,4'-DDD, 4,4'-DDE, 4,4'-DDT, Acenaphthene, Acenaphthylene, Aldrin, Alpha-Chlordane, Anthracene, Arsenic, Benz[a]anthracene, Benzo[a]pyrene, Benzo[b]fluoranthene, Benzo[e]pyrene, Benzo[g,h,i]perylene, Benzo[k]fluoranthene, Biphenyl, C1-Chrysenes, C1-Dibenzothiophenes, C1-Fluoranthenes_Pyrenes, C1-Fluorenes, C1-Naphthalenes, C1-Phenanthrenes_Anthracenes, C2-Chrysenes, C2-Dibenzothiophenes, C2-Fluorenes, C2-Naphthalenes, C2-Phenanthrenes_Anthracenes, C3-Chrysenes, C3-Dibenzothiophenes, C3-Fluorenes, C3-Naphthalenes, C3-Phenanthrenes_Anthracenes, C4-Chrysenes, C4-Naphthalenes, C4-Phenanthrenes_Anthracenes, Cadmium, Chromium, Chrysene, Copper, Dibenzo[a,h]anthracene, Dibenzothiophene, Dibutyltin, Dieldrin, Endrin, Fluoranthene, Fluorene, Gamma-Hexachlorocyclohexane, Heptachlor, Heptachlor-Epoxide, Hexachlorobenzene, Indeno[1,2,3-c,d]pyrene, Iron, Lead, Manganese, Mercury, Mirex, Monobutyltin, Naphthalene, Nickel, Perylene, Phenanthrene, Pyrene, Selenium, Silver, Tin, Trans-Nonachlor, Tributyltin, Zinc

Once the rows with missing values (NaN) were dropped the remaining datasets had the following dimensions.

(A): N = 1481 rows and P = 26 analytes.
(B): N = 965 rows and P = 76 analytes.

Training Datasets

The training sets reported here are reduced versions of (A) and (B) based on the minimum number of times a General Location must be observed to be included. The figure below shows the number of General Locations left in the dataset when a location is required to have a certain minimum number of observations, c. As the minimum number increases, the number of General Locations decreases monotonically and tends to show a sharp decline around some critical minimum number.

We provide 4 different datasets for version (A) and version (B) based on different values of c. If you wish to perform your own filtering, simply use the "v0" of each. Note, however, for any simple (stratified) test/train splitting of this dataset a minimum number of 2 observations would be required.

v	A	B
v0	c = 0, (1481, 26), 62 Locs.	c = 0, (965, 76), 61 Locs.
v1	c = 10, (1466, 26), 57 Locs.	c = 6, (958, 76), 58 Locs.
v2	c = 15, (1217, 26), 38 Locs.	c = 9 (823, 76), 40 Locs.
v3	c = 20, (1013, 26), 25 Locs.	c = 15 (671, 76), 25 Locs.

Datasets are named, e.g., train_A_v0.csv, according to this table. To get an adequate statistical description of each site requires a minimum number of samples depending on the application and user tolerance. We recommend using v3s for applications where a smaller number of well-sampled site is appropriate, while applications which require more sites but can tolerate fewer examples may benefit more from v0 or v1.

In the figure above we also report the performance of random forest trained on an 80:20 stratified train:test split of the data. The trends suggest that the more locations you have the harder it is to tell them apart from each other.

Usage

One way to download the data is by using the datasets library.

from datasets import load_dataset

df = load_dataset(
    "mahynski/crassostrea-virginica-chem",
    data_files="train_A_v0.csv",
)['train'].to_pandas()