Extracting training data from the ODC c4dc49701d4245c6b59f22a689ab0948

Background

Training data is the most important part of any supervised machine learning workflow. The quality of the training data has a greater impact on the classification than the algorithm used. Large and accurate training data sets are preferable: increasing the training sample size results in increased classification accuracy (Maxell et al 2018). A review of training data methods in the context of Earth Observation is available here

When creating training labels, be sure to capture the spectral variability of the class, and to use imagery from the time period you want to classify (rather than relying on basemap composites). Another common problem with training data is class imbalance. This can occur when one of your classes is relatively rare and therefore the rare class will comprise a smaller proportion of the training set. When imbalanced data is used, it is common that the final classification will under-predict less abundant classes relative to their true proportion.

There are many platforms to use for gathering training labels, the best one to use depends on your application. GIS platforms are great for collection training data as they are highly flexible and mature platforms; Geo-Wiki and Collect Earth Online are two open-source websites that may also be useful depending on the reference data strategy employed. Alternatively, there are many pre-existing training datasets on the web that may be useful, e.g. Radiant Earth manages a growing number of reference datasets for use by anyone.

Description

This notebook will extract training data (feature layers, in machine learning parlance) from the open-data-cube using labelled geometries within a geojson. The default example will use the crop/non-crop labels within the 'data/crop_training_WA.geojson' file. This reference data was acquired and pre-processed from the USGS’s Global Food Security Analysis Data portal here and here.

To do this, we rely on a custom dea-notebooks function called collect_training_data, contained within the dea_tools.classification script. The principal goal of this notebook is to familarise users with this function so they can extract the appropriate data for their use-case. The default example also highlights extracting a set of useful feature layers for generating a cropland mask forWA.

  1. Preview the polygons in our training data by plotting them on a basemap

  2. Define a feature layer function to pass to collect_training_data

  3. Extract training data from the datacube using collect_training_data

  4. Export the training data to disk for use in subsequent scripts


Getting started

To run this analysis, run all the cells in the notebook, starting with the “Load packages” cell.

Load packages

[1]:
%matplotlib inline

import os
import datacube
import numpy as np
import xarray as xr
import subprocess as sp
import geopandas as gpd
from odc.io.cgroups import get_cpu_quota
from datacube.utils.geometry import assign_crs

import sys
sys.path.insert(1, '../../Tools/')
from dea_tools.plotting import map_shapefile
from dea_tools.bandindices import calculate_indices
from dea_tools.classification import collect_training_data

import warnings
warnings.filterwarnings("ignore")
/env/lib/python3.8/site-packages/geopandas/_compat.py:106: UserWarning: The Shapely GEOS version (3.8.0-CAPI-1.13.1 ) is incompatible with the GEOS version PyGEOS was compiled with (3.9.1-CAPI-1.14.2). Conversions between both will be slow.
  warnings.warn(

Analysis parameters

  • path: The path to the input vector file from which we will extract training data. A default geojson is provided.

  • field: This is the name of column in your shapefile attribute table that contains the class labels. The class labels must be integers

[2]:
path = 'data/crop_training_WA.geojson'
field = 'class'

Find the number of CPUs

[3]:
ncpus = round(get_cpu_quota())
print('ncpus = ' + str(ncpus))
ncpus = 31

Preview input data

We can load and preview our input data shapefile using geopandas. The shapefile should contain a column with class labels (e.g. ‘class’). These labels will be used to train our model.

Remember, the class labels must be represented by integers.

[4]:
# Load input data shapefile
input_data = gpd.read_file(path)

# Plot first five rows
input_data.head()
[4]:
class geometry
0 1 POINT (116.60407 -31.46883)
1 1 POINT (117.03464 -32.40830)
2 1 POINT (117.30838 -32.33747)
3 1 POINT (116.74607 -31.63750)
4 1 POINT (116.85817 -33.00131)
[5]:
# Plot training data in an interactive map
map_shapefile(input_data, attribute=field)

Extracting training data

The function collect_training_data takes our geojson containing class labels and extracts training data (features) from the datacube over the locations specified by the input geometries. The function will also pre-process our training data by stacking the arrays into a useful format and removing any NaN or inf values.

The below variables can be set within the collect_training_data function: * zonal_stats: An optional string giving the names of zonal statistics to calculate across each polygon (if the geometries in the vector file are polygons and not points). Default is None (all pixel values are returned). Supported values are ‘mean’, ‘median’, ‘max’, and ‘min’. * return_coords: If True, then the training data will contain two extra columns ‘x_coord’ and ‘y_coord’ corresponding to the x,y coordinate of each sample. This variable can be useful for handling spatial autocorrelation between samples later on in the ML workflow when we conduct k-fold cross validation. * dc_query: a datacube dictionary query object, This should not contain lat and long (x or y) variables as these are supplied by the ‘gdf’ geometries.

Note: collect_training_data also has a number of additional parameters for handling ODC I/O read failures, where polygons that return an excessive number of null values can be resubmitted to the multiprocessing queue. Check out the docs to learn more.

In addition to the parameters required for collect_training_data, we also need to set up a few parameters for the dc_query parameter, such as measurements (the bands to load from the satellite), the resolution (the cell size), and the output_crs (the output projection).

[6]:
# Set up our inputs to collect_training_data
time = ('2014')
zonal_stats = None
return_coords = True

# Set up the inputs for the ODC query
measurements = ['blue', 'green', 'red', 'nir', 'swir1', 'swir2']
resolution = (-30, 30)
output_crs = 'epsg:3577'
[7]:
# Generate a new datacube query object
query = {
    'time': time,
    'measurements': measurements,
    'resolution': resolution,
    'output_crs': output_crs,
}

Defining feature layers

To create the desired feature layers, we pass instructions to collect training data through the feature_func parameter.

  • feature_func: A function for generating feature layers that is applied to the data within the bounds of the input geometry. The ‘feature_func’ must accept a ‘dc_query’ object, and return a single xarray.Dataset or xarray.DataArray containing 2D coordinates (i.e x, y - no time dimension). e.g.

    def feature_function(query):
        dc = datacube.Datacube(app='feature_layers')
        ds = dc.load(**query)
        ds = ds.mean('time')
        return ds
    

Below, we will define a more complicated feature layer function than the brief example shown above. We will calculate some band indices on the Landsat 8 geomedian, append the ternary median aboslute deviation dataset from the same year: ls8_nbart_tmad_annual, and append fractional cover percentiles for the photosynthetic vegetation band, also from the same year: fc_percentile_albers_annual.

[8]:
def feature_layers(query):
    #connect to the datacube
    dc = datacube.Datacube(app='custom_feature_layers')

    #load ls8 geomedian
    ds = dc.load(product='ls8_nbart_geomedian_annual',
                 **query)

    # Calculate some band indices
    da = calculate_indices(ds,
                           index=['NDVI', 'LAI', 'MNDWI'],
                           drop=False,
                           collection='ga_ls_2')

    # Add TMADs dataset
    tmad = dc.load(product='ls8_nbart_tmad_annual',
                   measurements=['sdev','edev','bcdev'],
                   like=ds.geobox, #will match geomedian extent
                   time='2014' #same as geomedian
                  )

    # Add Fractional cover percentiles
    fc = dc.load(product='fc_percentile_albers_annual',
                   measurements=['PV_PC_10','PV_PC_50','PV_PC_90'], #only the PV band
                   like=ds.geobox, #will match geomedian extent
                   time='2014' #same as geomedian
                  )

    # Merge results into single dataset
    result = xr.merge([da, tmad, fc],compat='override')

    return result

Now, we can pass this function to collect_training_data. This will take a few minutes to run all 430 samples on the default sandbox as it only has two cpus.

[9]:
%%time
column_names, model_input = collect_training_data(
    gdf=input_data,
    dc_query=query,
    ncpus=ncpus,
    return_coords=return_coords,
    field=field,
    zonal_stats=zonal_stats,
    feature_func=feature_layers)
Collecting training data in parallel mode
Percentage of possible fails after run 1 = 0.0 %
Removed 0 rows wth NaNs &/or Infs
Output shape:  (430, 18)
CPU times: user 1.47 s, sys: 388 ms, total: 1.86 s
Wall time: 33.6 s
[10]:
print(column_names)
print('')
print(np.array_str(model_input, precision=2, suppress_small=True))
['class', 'blue', 'green', 'red', 'nir', 'swir1', 'swir2', 'NDVI', 'LAI', 'MNDWI', 'sdev', 'edev', 'bcdev', 'PV_PC_10', 'PV_PC_50', 'PV_PC_90', 'x_coord', 'y_coord']

[[       1.     1005.     1464. ...       68. -1393035. -3614685.]
 [       1.      900.     1290. ...       71. -1387965. -3625995.]
 [       1.      916.     1374. ...       72. -1391205. -3622155.]
 ...
 [       0.      384.      584. ...       55.  -632505. -1746885.]
 [       0.      232.      344. ...       71. -1468095. -3805095.]
 [       0.      526.      928. ...       15.  -865575. -2818005.]]

Separate coordinate data

By setting return_coords=True in the collect_training_data function, our training data now has two extra columns called x_coord and y_coord. We need to separate these from our training dataset as they will not be used to train the machine learning model. Instead, these variables will be used to help conduct Spatial K-fold Cross validation (SKVC) in the notebook 3_Evaluate_optimize_fit_classifier. For more information on why this is important, see this article.

[11]:
# Select the variables we want to use to train our model
coord_variables = ['x_coord', 'y_coord']

# Extract relevant indices from the processed shapefile
model_col_indices = [column_names.index(var_name) for var_name in coord_variables]

# Export to coordinates to file
np.savetxt("results/training_data_coordinates.txt", model_input[:, model_col_indices])

Export training data

Once we’ve collected all the training data we require, we can write the data to disk. This will allow us to import the data in the next step(s) of the workflow.

[12]:
# Set the name and location of the output file
output_file = "results/test_training_data.txt"
[13]:
# Grab all columns except the x-y coords
model_col_indices = [column_names.index(var_name) for var_name in column_names[0:-2]]

# Export files to disk
np.savetxt(output_file, model_input[:, model_col_indices], header=" ".join(column_names[0:-2]), fmt="%4f")

Additional information

License: The code in this notebook is licensed under the Apache License, Version 2.0. Digital Earth Australia data is licensed under the Creative Commons by Attribution 4.0 license.

Contact: If you need assistance, please post a question on the Open Data Cube Slack channel or on the GIS Stack Exchange using the open-data-cube tag (you can view previously asked questions here). If you would like to report an issue with this notebook, you can file one on Github.

Last modified: September 2021

Compatible datacube version:

[14]:
print(datacube.__version__)
1.8.5

Tags

Browse all available tags on the DEA User Guide’s Tags Index

Tags Landsat 8 geomedian, Landsat 8 TMAD, machine learning, collect_training_data, Fractional Cover