/gval

A high-level Python framework to evaluate the skill of geospatial datasets by comparing candidates to benchmark maps producing agreement maps and metrics.

Primary LanguageJupyter NotebookOtherNOASSERTION

alt text alt text

Build and TestCoveragePyPI version

GVAL (pronounced "g-val") is a high-level Python framework to evaluate the skill of geospatial datasets by comparing candidates to benchmark maps producing agreement maps and metrics.

GVAL is intended to work on raster and vector files as xarray and geopandas objects, respectively. Abilities to prepare or homogenize maps for comparison are included. The comparisons are based on scoring philosophies for three statistical data types including categorical, continuous, and probabilistic.

See the full documentation.

WARNING:

  • Our current public API and output formats are likely to change in the future.
  • Software is provided "AS-IS" without any guarantees. Please QA/QC your metrics carefully until this project matures.

Installation

General Use

To use this package:

pip install gval

Or for bleeding edge updates install from the repository:

pip install 'git+https://github.com/NOAA-OWP/gval'

Using GVAL

Categorical Example

An example of running the entire process for two-class categorical rasters with one function using minimal arguments is demonstrated below:

import gval
import rioxarray as rxr

candidate = rxr.open_rasterio('candidate_map_two_class_categorical.tif', mask_and_scale=True)
benchmark = rxr.open_rasterio('benchmark_map_two_class_categorical.tif', mask_and_scale=True)

(agreement_map,
 crosstab_table,
 metric_table) = candidate.gval.categorical_compare(benchmark,
                                                   positive_categories=[2],
                                                   negative_categories=[0, 1])

Categorical Outputs

agreement_map

alt text

crosstab_table

alt text

metric_table

alt text

Continuous Example

The same can be done for rasters with continuous valued statistical data types as shown below (in this case only a subset of the default statistics will be run):

import gval
import rioxarray as rxr

candidate = rxr.open_rasterio('livneh_2011_precip.tif', mask_and_scale=True) # VIC
benchmark = rxr.open_rasterio('prism_2011_precip.tif', mask_and_scale=True) # PRISM

agreement, metric_table = candidate.gval.continuous_compare(
    benchmark,
    metrics=[
        "coefficient_of_determination",
        "mean_percentage_error",
        "mean_absolute_percentage_error",
        "mean_normalized_mean_absolute_error"
    ]
)

Continuous Outputs

agreement_map

alt text

metric_table

alt text

Catalog Example

Entire catalogs can be compared in GVAL, which are represented by dataframes of maps. The following is a candidate and benchmark catalog for continuous datasets:

candidate_catalog

alt text

benchmark_catalog

alt text

With the following code a comparison of each pair of maps can be run with the following code. Since the parameter agreement_map_field is provided the column agreement_maps found in the candidate catalog will be used to export the agreement map to that location. (Note the first pair of candidate and benchmark maps are single band rasters while the second pair are multiband rasters):

import pandas as pd

from gval.catalogs.catalogs import catalog_compare

candidate_continuous_catalog = pd.read_csv('candidate_catalog_0.csv')
benchmark_continuous_catalog = pd.read_csv('benchmark_catalog_0.csv')

arguments = {
    "candidate_catalog": candidate_continuous_catalog,
    "benchmark_catalog": benchmark_continuous_catalog,
    "on": "compare_id",
    "agreement_map_field": "agreement_maps",
    "map_ids": "map_id",
    "how": "inner",
    "compare_type": "continuous",
    "compare_kwargs": {
        "metrics": (
            "coefficient_of_determination",
            "mean_absolute_error",
            "mean_absolute_percentage_error",
        ),
        "encode_nodata": True,
        "nodata": -9999,
    },
    "open_kwargs": {
        "mask_and_scale": True,
        "masked": True
    }
}

agreement_continuous_catalog = catalog_compare(**arguments)

Catalog Outputs

agreement_map

alt text alt text

catalog_metrics

alt text

(Note that both catalog level attributes in the candidate and benchmark catalogs are present in the catalog metrics table.)

For more detailed examples of how to use this software, check out these notebook tutorials.

Contributing

Guidelines for contributing to this repository can be found at CONTRIBUTING.

Citation

Please cite our work if using this package. See 'cite this repository' in the about section on GitHub or refer to CITATION.cff