Written by Tom Augspurger
Created: November 26, 2018
Last updated: June 30, 2020
The image loaded here is a cropped portion of a LANDSAT image of Walker Lake.
In addition to dask-ml, we'll use rasterio to read the data and matplotlib to plot the figures.
I'm just working on my laptop, so we could use either the threaded or distributed scheduler, but here I'll use the distributed scheduler for the diagnostics.
import rasterio
import numpy as np
import xarray as xr
import holoviews as hv
from holoviews import opts
from holoviews.operation.datashader import regrid
import cartopy.crs as ccrs
import dask.array as da
#from dask_ml.cluster import SpectralClustering
from dask.distributed import Client
hv.extension('bokeh')
import dask_ml
dask_ml.__version__
from dask_ml.cluster import SpectralClustering
client = Client(processes=False)
#client = Client(n_workers=8, threads_per_worker=1)
client
import intake
cat = intake.open_catalog('./catalog.yml')
list(cat)
landsat_5_img = cat.landsat_5.read_chunked()
landsat_5_img
crs=ccrs.epsg(32611)
x_center, y_center = crs.transform_point(-118.7081, 38.6942, ccrs.PlateCarree())
buffer = 1.7e4
xmin = x_center - buffer
xmax = x_center + buffer
ymin = y_center - buffer
ymax = y_center + buffer
ROI = landsat_5_img.sel(x=slice(xmin, xmax), y=slice(ymax, ymin))
ROI = ROI.where(ROI > ROI.nodatavals[0])
bands = ROI.astype(float)
bands = (bands - bands.mean()) / bands.std()
bands
opts.defaults(
opts.Image(invert_yaxis=True, width=250, height=250, tools=['hover'], cmap='viridis'))
hv.Layout([regrid(hv.Image(band, kdims=['x', 'y'])) for band in bands[:3]])
flat_input = bands.stack(z=('y', 'x'))
flat_input
flat_input.shape
We'll reshape the image to be how dask-ml / scikit-learn expect it: (n_samples, n_features) where n_features is 1 in this case. Then we'll persist that in memory. We still have a small dataset at this point. The large dataset, which dask helps us manage, is the intermediate n_samples x n_samples array that spectral clustering operates on. For our 2,500 x 2,500 pixel subset, that's ~50
X = flat_input.values.astype('float').T
X.shape
X = da.from_array(X, chunks=100_000)
X = client.persist(X)
And we'll fit the estimator.
clf = SpectralClustering(n_clusters=4, random_state=0,
gamma=None,
kmeans_params={'init_max_iter': 5},
persist_embedding=True)
%time clf.fit(X)
labels = clf.assign_labels_.labels_.compute()
labels.shape
labels = labels.reshape(bands[0].shape)
hv.Layout([regrid(hv.Image(band, kdims=['x', 'y'])) for band in bands])
hv.Layout([regrid(hv.Image(band, kdims=['x', 'y'])) for band in bands[3:]])
hv.Image(labels)