def load_crophealth_data():
"""
Loads Sentinel-2 analysis-ready data (ARD) product for the crop health
case-study area. The ARD product is provided for the last year.
Last modified: January 2020
outputs
ds - data set containing combined, masked data from Sentinel-2a and -2b.
Masked values are set to 'nan'
"""
# Suppress warnings
warnings.filterwarnings('ignore')
# Initialise the data cube. 'app' argument is used to identify this app
dc = datacube.Datacube(app='Crophealth-app')
# Specify latitude and longitude ranges
latitude = (-24.974997, -24.995971)
longitude = (152.429994, 152.395805)
# Specify the date range
# Calculated as today's date, subtract 90 days to match NRT availability
# Dates are converted to strings as required by loading function below
end_date = dt.date.today()
start_date = end_date - dt.timedelta(days=365)
time = (start_date.strftime("%Y-%m-%d"), end_date.strftime("%Y-%m-%d"))
# Construct the data cube query
products = ["s2a_ard_granule", "s2b_ard_granule"]
query = {
'x':
longitude,
'y':
latitude,
'time':
time,
'measurements': [
'nbar_red', 'nbar_green', 'nbar_blue', 'nbar_nir_1', 'nbar_swir_2',
'nbar_swir_3'
],
'output_crs':
'EPSG:3577',
'resolution': (-10, 10)
}
# Load the data and mask out bad quality pixels
ds_s2 = load_ard(dc, products=products, min_gooddata=0.5, **query)
# Calculate the normalised difference vegetation index (NDVI) across
# all pixels for each image.
# This is stored as an attribute of the data
ds_s2 = calculate_indices(ds_s2, index='NDVI', collection='ga_s2_1')
# Return the data
return (ds_s2)
"output_crs": "EPSG:32755",
"resolution": (-10, 10)
}
prefire_data = load_ard(
dc=dc,
products=['s2a_ard_granule', 's2b_ard_granule'],
measurements=['nbart_nir_1', 'nbart_swir_3'],
min_gooddata=0,
# dask_chunks={'x': 'auto', 'y': 'auto'},
group_by='solar_day',
**query_1)
prefire_image = prefire_data.median(dim='time')
prefire_image = calculate_indices(prefire_image,
index='NBR',
collection='ga_s2_1',
drop=False)
prefire_burnratio = prefire_image.NBR
prefire_burnratio.data
query_2 = {
"x": (central_lon - buffer, central_lon + buffer),
"y": (central_lat - buffer, central_lat + buffer),
"time": (postfire_start, postfire_end),
"output_crs": "EPSG:32755",
"resolution": (-10, 10)
}
postfire_data = load_ard(
dc=dc,
products=['s2a_ard_granule', 's2b_ard_granule'],
def get_training_data_for_shp(gdf,
index,
row,
out_arrs,
out_vars,
products,
dc_query,
custom_func=None,
field=None,
calc_indices=None,
reduce_func=None,
drop=True,
zonal_stats=None):
"""
Function to extract data from the ODC for training a machine learning classifier
using a geopandas geodataframe of labelled geometries.
This function provides a number of pre-defined methods for producing training data,
including: calcuating band indices, reducing time series using summary statistics,
and/or generating zonal statistics across polygons. The 'custom_func' parameter provides
a method for the user to supply a custom function for generating features rather than using the
pre-defined methods.
Parameters
----------
gdf : geopandas geodataframe
geometry data in the form of a geopandas geodataframe
products : list
a list of products to load from the datacube.
e.g. ['ga_ls7e_ard_3', 'ga_ls8c_ard_3']
dc_query : dictionary
Datacube query object, should not contain lat and long (x or y)
variables as these are supplied by the 'gdf' variable
field : string
A string containing the name of column with class labels.
Field must contain numeric values.
out_arrs : list
An empty list into which the training data arrays are stored.
out_vars : list
An empty list into which the data varaible names are stored.
custom_func : function, optional
A custom function for generating feature layers. If this parameter
is set, all other options (excluding 'zonal_stats'), will be ignored.
The result of the 'custom_func' must be a single xarray dataset
containing 2D coordinates (i.e x, y - no time dimension). The custom function
has access to the datacube dataset extracted using the 'dc_query' params.
Example custom function to return multiple products:
`def custom_function(ds):
dc = datacube.Datacube(app='custom_function')
mad = dc.load(product='ls8_nbart_tmad_annual', like=ds.geobox)
output = xr.merge([ds, mad])
return output`
calc_indices: list, optional
If not using a custom func, then this parameter provides a method for
calculating a number of remote sensing indices (e.g. `['NDWI', 'NDVI']`).
reduce_func : string, optional
Function to reduce the data from multiple time steps to
a single timestep. Options are 'mean', 'median', 'std',
'max', 'min', 'geomedian'. Ignored if 'custom_func' is provided.
drop : boolean, optional
If this variable is set to True, and 'calc_indices' are supplied, the
spectral bands will be dropped from the dataset leaving only the
band indices as data variables in the dataset. Default is True.
zonal_stats : string, optional
An optional string giving the names of zonal statistics to calculate
for each polygon. Default is None (all pixel values are returned). Supported
values are 'mean', 'median', 'max', 'min', and 'std'. Will work in
conjunction with a 'custom_func'.
Returns
--------
Two lists, a list of numpy.arrays containing classes and extracted data for
each pixel or polygon, and another containing the data variable names.
"""
# prevent function altering dictionary kwargs
dc_query = deepcopy(dc_query)
# remove dask chunks if supplied as using
# mulitprocessing for parallelization
if 'dask_chunks' in dc_query.keys():
dc_query.pop('dask_chunks', None)
# connect to datacube
dc = datacube.Datacube(app='training_data')
# set up query based on polygon (convert to albers)
geom = geometry.Geometry(gdf.geometry.values[index].__geo_interface__,
geometry.CRS('epsg:3577'))
q = {"geopolygon": geom}
# merge polygon query with user supplied query params
dc_query.update(q)
# load_ard doesn't handle derivative products, so check
# products aren't one of those below
others = [
'ls5_nbart_geomedian_annual', 'ls7_nbart_geomedian_annual',
'ls8_nbart_geomedian_annual', 'ls5_nbart_tmad_annual',
'ls7_nbart_tmad_annual', 'ls8_nbart_tmad_annual',
'landsat_barest_earth', 'ls8_barest_earth_albers'
]
if products[0] in others:
ds = dc.load(product=products[0], **dc_query)
ds = ds.where(ds != 0, np.nan)
else:
# load data
with HiddenPrints():
ds = load_ard(dc=dc,
products=products,
output_crs='EPSG:3577',
**dc_query)
# create polygon mask
with HiddenPrints():
mask = xr_rasterize(gdf.iloc[[index]], ds)
# Use custom function for training data if it exists
if custom_func is not None:
with HiddenPrints():
data = custom_func(ds)
# Mask dataset
data = data.where(mask)
else:
# Mask dataset
ds = ds.where(mask)
# first check enough variables are set to run functions
if (len(ds.time.values) > 1) and (reduce_func == None):
raise ValueError(
"You're dataset has " + str(len(ds.time.values)) +
" time-steps, please provide a reduction function," +
" e.g. reduce_func='mean'")
if calc_indices is not None:
# determine which collection is being loaded
if products[0] in others:
collection = 'ga_ls_2'
elif '3' in products[0]:
collection = 'ga_ls_3'
elif 's2' in products[0]:
collection = 'ga_s2_1'
if len(ds.time.values) > 1:
if reduce_func in ['mean', 'median', 'std', 'max', 'min']:
with HiddenPrints():
data = calculate_indices(ds,
index=calc_indices,
drop=drop,
collection=collection)
# getattr is equivalent to calling data.reduce_func
method_to_call = getattr(data, reduce_func)
data = method_to_call(dim='time')
elif reduce_func == 'geomedian':
data = GeoMedian().compute(ds)
with HiddenPrints():
data = calculate_indices(data,
index=calc_indices,
drop=drop,
collection=collection)
else:
raise Exception(
reduce_func + " is not one of the supported" +
" reduce functions ('mean','median','std','max','min', 'geomedian')"
)
else:
with HiddenPrints():
data = calculate_indices(ds,
index=calc_indices,
drop=drop,
collection=collection)
# when band indices are not required, reduce the
# dataset to a 2d array through reduce function
if calc_indices is None:
if len(ds.time.values) > 1:
if reduce_func == 'geomedian':
data = GeoMedian().compute(ds)
elif reduce_func in ['mean', 'median', 'std', 'max', 'min']:
method_to_call = getattr(ds, reduce_func)
data = method_to_call('time')
else:
data = ds.squeeze()
if zonal_stats is None:
# If no zonal stats were requested then extract all pixel values
flat_train = sklearn_flatten(data)
# Make a labelled array of identical size
flat_val = np.repeat(row[field], flat_train.shape[0])
stacked = np.hstack((np.expand_dims(flat_val, axis=1), flat_train))
elif zonal_stats in ['mean', 'median', 'std', 'max', 'min']:
method_to_call = getattr(data, zonal_stats)
flat_train = method_to_call()
flat_train = flat_train.to_array()
stacked = np.hstack((row[field], flat_train))
else:
raise Exception(
zonal_stats + " is not one of the supported" +
" reduce functions ('mean','median','std','max','min')")
# Append training data and labels to list
out_arrs.append(stacked)
out_vars.append([field] + list(data.data_vars))
def get_training_data_for_shp(path,
out,
product,
time,
crs='EPSG:3577',
field='classnum',
calc_indices=None,
feature_stats=None,
collection='ga_ls_2'):
"""
Function to extract data for training classifier using a shapefile
of labelled polygons. Currently works for single time steps.
Parameters
----------
path : string
Path to shapefile containing labelled polygons.
out : list
Empty list to contain output data.
product : string
String of product name from which to load and extract datacube
data e.g. 'ls8_nbart_tmad_annual'
time : tuple
A tuple containing the time period from which to extract
training data e.g. ('2015-01-01', '2015-12-31').
crs : string
A string containing desired crs e.g. 'EPSG:3577'
field : string
A string containing name of column with labels in shapefile
attribute table. Field must contain numeric values.
calc_indices: list, optional
An optional list giving the names of any remote sensing indices
to be calculated on the loaded data (e.g. `['NDWI', 'NDVI']`.
This step will be skipped if any of the indices cannot be
computed on the input product.
feature_stats: string, optional
An optional string giving the names of statistics to calculate
for the polygon. Default is None (all pixel values). Supported
values are 'mean' or 'geomedian' (from the `hdstats` module).
Returns
--------
A list of numpy.arrays containing classes and extracted data for
each pixel or polygon.
"""
# Import hdstats as only needed for this function
if feature_stats == 'geomedian':
try:
import hdstats
except ImportError as err:
raise
raise ImportError(
'Can not import hdstats module needed to calculate'
' geomedian.\n{}'.format(err))
dc = datacube.Datacube(app='training_data')
query = {'time': time}
query['crs'] = crs
shp = gp.read_file(path)
bounds = shp.total_bounds
minx = bounds[0]
maxx = bounds[2]
miny = bounds[1]
maxy = bounds[3]
query['x'] = (minx, maxx)
query['y'] = (miny, maxy)
print("Loading data...")
data = dc.load(product=product, group_by='solar_day', **query)
# Check if geomedian is in the product and if indices are wanted
if calc_indices is not None:
try:
print("Calculating indices...")
# Calculate indices - will use for all features
for index in calc_indices:
data = dea_bandindices.calculate_indices(data,
index,
collection=collection)
except ValueError:
print(
"Input dataset not suitable for selected indices, just extracting product data"
)
pass
# Remove time step if present
try:
data = data.isel(time=0)
# Don't worry if it isn't
except ValueError:
pass
print("Rasterizing features and extracting data...")
# Initialize counter for status messages.
i = 0
# Go through each feature
for poly_geom, poly_class_id in zip(shp.geometry, shp[field]):
print(" Feature {:04}/{:04}\r".format(i + 1, len(shp.geometry)),
end='')
# Rasterise the feature
mask = rasterize([(poly_geom, poly_class_id)],
out_shape=(data.y.size, data.x.size),
transform=data.affine)
# Convert mask from numpy to DataArray
mask = xr.DataArray(mask, coords=(data.y, data.x))
# Mask out areas that were not within the labelled feature
data_masked = data.where(mask == poly_class_id, np.nan)
if feature_stats is None:
# If no summary stats were requested then
# extract all pixel values
flat_train = sklearn_flatten(data_masked)
# Make a labelled array of identical size
flat_val = np.repeat(poly_class_id, flat_train.shape[0])
stacked = np.hstack((np.expand_dims(flat_val, axis=1), flat_train))
elif feature_stats == 'mean':
# For the mean of each polygon take the mean over all
# axis, ignoring masked out values (nan).
# This gives a single pixel value for each band
flat_train = data_masked.mean(axis=None, skipna=True)
flat_train = flat_train.to_array()
stacked = np.hstack((poly_class_id, flat_train))
elif feature_stats == 'geomedian':
# For the geomedian flatten so have a 2D array with
# bands and pixel values. Then use hdstats
# to calculate the geomedian
flat_train = sklearn_flatten(data_masked)
flat_train_median = hdstats.geomedian(flat_train, axis=0)
# Geomedian will return a single value for each band so join
# this with class id to create a single row in output
stacked = np.hstack((poly_class_id, flat_train_median))
# Append training data and label to list
out.append(stacked)
# Update status counter (feature number)
i = i + 1
# Return a list of labels for columns in output array
return [field] + list(data.data_vars)
def dNBR_processing(coordinates):
# Load all data in baseline period available from s2a/b_ard_granule datasets
prefire_ard = load_ard(
dc=dc,
products=['s2a_ard_granule', 's2b_ard_granule'],
x=(coordinates.x - 0.1, coordinates.x + 0.1),
y=(coordinates.y - 0.1, coordinates.y + 0.1),
time=(prefire_start, prefire_end),
measurements=['nbart_nir_1', 'nbart_swir_3'],
min_gooddata=0.1,
output_crs='EPSG:32755', # UTM Zone 55S
resolution=(-10, 10),
group_by='solar_day')
prefire_ard = calculate_indices(prefire_ard,
index='NBR',
collection='ga_s2_1',
drop=False)
# Compute median using all observations in the dataset along the time axis
prefire_image = prefire_ard.median(dim='time')
# Delete baseline_combined
del prefire_ard
# Select NBR
prefire_NBR = prefire_image.NBR
del prefire_image
# Load all data in post-fire period available from s2a/b_ard_granule datasets
postfire_ard = load_ard(
dc=dc,
products=['s2a_ard_granule', 's2b_ard_granule'],
x=(coordinates.x - 0.1, coordinates.x + 0.1),
y=(coordinates.y - 0.1, coordinates.y + 0.1),
time=(postfire_start, postfire_end),
measurements=['nbart_nir_1', 'nbart_swir_3'],
min_gooddata=0.1,
output_crs='EPSG:32755', # UTM Zone 55S
resolution=(-10, 10),
group_by='solar_day')
# Calculate NBR on all post-fire images
postfire_ard = calculate_indices(postfire_ard,
index='NBR',
collection='ga_s2_1',
drop=False)
# Calculate the median post-fire image
postfire_image = postfire_ard.median(dim='time')
del postfire_ard
# Select NBR
postfire_NBR = postfire_image.NBR
del postfire_image
# Calculate delta
delta_NBR = prefire_NBR - postfire_NBR
del prefire_NBR
del postfire_NBR
x = np.round_(coordinates.x, decimals=4)
y = np.round_(coordinates.y, decimals=4)
# Turn dNBR into a x-array dataset for export to GeoTIFF
dnbr_dataset = delta_NBR.to_dataset(name='delta_NBR')
# cog.write_cog(dnbr_dataset, './NBR_geotiffs/{x}_{y}_dNBR.tif')
write_geotiff(f'/scratch/wj97/ab4513/dNBR_geotiffs/{x}_{y}_dNBR.tif',
dnbr_dataset)
del delta_NBR
del dnbr_dataset