Keras Spatial

Keras Spatial includes data generators and tools designed to simplify the preprocessing of spatial data for deep learning applications.

Keras Spatial provides a data generator that reads samples directly from a raster data source and eliminates the need to create small, individual raster files prior to model execution. The raster data source may be local or remote service. Any necessary reprojections and scaling is handled automatically.

Central to the use of Keras Spatial is a GeoPandas GeoDataFrame which defines a virtual sample set, a list of samples that drives the data generator. The dataframe can also be used to filter samples based on different aspects such as the existance of nodata, handling imbalanced label distributions, and storing sample attributes used in normalization amoung other data augmentation functions.

Features include:

• Sample extraction from local or remote data sources -- no intermediate files
• Automatic reprojection and resampling as needed
• Sample augmentation using user-defined callback system
• Flexible structure improves organization and data management

Installation

To install the package from PyPi repository you can execute the following command:

pip install keras-spatial

or directly from GitHub

$ pip install git+https://github.com/IllinoisStateGeologicalSurvey/keras-spatial#egg=keras-spatial --process-dependency-links

Quickstart

1. Create a SpatialDataGen and set the source raster
2. Create a geodataframe with 200x200 (in projection units) samples covering the spatial extent of the raster
3. Create the generator producing arrays with shape [32, 128, 128, 1]
4. Fit model

from keras_spatial.datagen import SpatialDataGenerator

sdg = SpatialDataGenerator(source='/path/to/file.tif')
geodataframe = sdg.regular_grid(200, 200)
generator = sdg.flow_from_dataframe(geodataframe, 128, 128, batch_size=32)
model(generator, ...)

Usage

Keras Spatial provides a SpatialDataGenerator (SDG) modeled on the Keras ImageDataGenerator. The SDG allows user to work in spatial coorindates rather than pixels and easily integrate data from different coordinates systems. Reprojection and resampling is handled automatically as needed. Because Keras Spatial is based on the rasterio package, raster data source may either local files or remote resources referenced by URL.

Because the SDG reads directly from larger raster data sources rather than small, preprocessed images files, SDG makes use of a GeoDataFrame to identify each sample area. The geometry associated with the datafame is expected to be a polygon but extraction is done using a windowed read based on the bounds. As with the ImageDataGenerator, the flow_from_dataframe method returns the generator that can be passed to the Keras model.

SpatialDataGenerator class

The SDG is similar to the ImageDataGenerator albeit missing the .flow and the .flow_from_directory methods. SDG also moves more configutation and setting to the instance and with the .flow_from_dataframe having few arguments.

Arguments

• source (path or url): raster source
• width (int): array size produced by generator
• height (int): array size produced by generator
• indexes (int or tuple of ints): one or more raster bands to sampled
• interleave (str): type of interleave 'band' or 'pixel' (default='pixel')
• resampling (int): One of the values from rasterio.enums.Resampling (default=Resampling.nearest)

Raises RasterioIOError when the source is set if the file or remote resource is not available.

Examples

from keras_spatial import SpatialDataGenerator

sdg = SpatialDataGenerator(source='/path/to/file.tif')
sdg.width, sdg.height = 128,128

The source must be set prior to calling flow_from_dataframe. Width and height are also required but maybe passed as arguments to flow_from_dataframe.

The indexes argument selects bands in a multiband raster. By default all bands are read and the indexes argument is updated when the raster source is set.

In multiband situations, if interleave is set to 'band' the numpy array axes are moved to the following order [batch_size, bands, height, width]. This can lead to incompatible shapes when using multiple SDG generators -- use with care. The default interleave is 'pixel' which is compatible with Tensorflow.

# file.tif is a 5 band raster
sdg = SpatialDataGenerator('/path/to/file.tif')
gen = sdg.flow_from_dataframe(df, 128, 128, batch_size=1)
print(next(gen).shape)
> [1, 128, 128, 5]
sdg.interleave = 'band'
gen = sdg.flow_from_dataframe(df, 128, 128, batch_size=1)
print(next(gen.shape))
> [1, 5, 200, 200]

Because more than one SDG is expected to be used simultaneously and SDGs are expected to having matching spatial requirements, the SDG class provides a profile attribute that can be easily share arguments across instances as shown below. Note: source is not part of the profile.

sdg = SpatialDataGenerator(source='/path/to/file.tif')
sdg2 = SpatialDataGenerator()
sdg2.profile = sdg.profile
sdg2.source = '/path/to/file2.tif'

SpatialDataGenerator methods

flow_from_dataframe

flow_from_dataframe(geodataframe, width, height, batch_size)

Creates a generator that returns a numpy ndarray of samples read from the SDG source.

Arguments

• geodataframe (GeoDataFrame): a geodataframe with sample boundaries
• width (int): width of array
• height (int): height of array
• batch_size (int): number of samples to returned by generator

Returns

A generator of numpy ndarrays of the shape [batch_size, height, width, bands].

Example

sdg = SpatialDataGenerator(source='/path/to/file.tif')
gen = sdg.flow_from_dataframe(df, 128, 128)
arr = next(gen)

random_grid

random_grid(width, height, count, units='native')

Creates a geodataframe suitable to passing to the flow_from_dataframe method. The grid module provides a similar function using passed using spatial extents.

Arguments

• width (int): width in pixels
• height (int): height in pixels
• count (int): number of samples
• units (str): units for width and height, either native or in pixels

Returns

A GeoDataFrame defining the polygon boundary of each sample.

Example

sdg = SpatialDataGenerator(source='/path/to/file.tif')
df = sdg.random_grid(200, 200, 1000)

regular_grid

regular_grid(width, height, overlap=0.0, units='native')

Creates a geodataframe suitable to passing to the flow_from_dataframe method. The sample module provides a similar function using passed using spatial extents.

Arguments

• width (int): width in pixels
• height (int): width in pixels
• overlap (float): percentage of overlap (default=0.0)
• units (str): units for width and height, either native or in pixels

Returns

A GeoDataFrame defining the polygon boundary of each sample.

Example

sdg = SpatialDataGenerator(source='/path/to/file.tif')
df = sdg.regular_grid(200, 200)

Full Example

from keras_spatial import SpatialDataGenerator
labels = SpatialDataGenerator()
labels.source = '/path/to/labels.tif'
labels.width, labels.height = 128, 128
df = labels.regular_grid(200,200)

samples = SpatialDataGenerator()
samples.source = 'https://server.com/files/data.tif'
samples.width, samples.height = labels.width, label.height

train_gen = zip(labels.flow_from_dataframe(df), patches.flow_from_dataframe(df))
model(train_gen)