pygeoprocessing 1.4.0

PyGeoprocessing: Geoprocessing routines for GIS

.. default-role:: code

About PyGeoprocessing
=====================

PyGeoprocessing is a Python/Cython based library that provides a set of commonly
used raster, vector, and hydrological operations for GIS processing. Similar
functionality can be found in ArcGIS/QGIS raster algebra, ArcGIS zonal
statistics, and ArcGIS/GRASS/TauDEM hydrological routing routines.

PyGeoprocessing is developed at the Natural Capital Project to create a
programmable, open source, and free Python based GIS processing library to support the
InVEST toolset. PyGeoprocessing's design prioritizes
computation and memory efficient runtimes, easy installation and cross
compatibility with other open source and proprietary software licenses, and a
simplified set of orthogonal GIS processing routines that interact with GIS data
via filename. Specifically the functionally provided by PyGeoprocessing includes

* a suite of raster manipulation functions (warp, align, raster calculator, reclassification, distance transform, convolution, and fast iteration)
* a suite of vector based manipulation function (zonal statistics, rasterization, interpolate points, reprojection, and disjoint polygon sets)
* a simplified hydrological routing library (D8inf/MFD flow direction, plateau drainage, weighted and unweighted flow accumulation, and weighted and unweighted flow distance)

Installing PyGeoprocessing
==========================

.. code-block:: console

$ pip install pygeoprocessing

If you `import pygeoprocessing` and see a `ValueError: numpy.dtype has the
wrong size, try recompiling`, this is the result of a version compatibility
issue with the numpy ABI in the precompiled pygeoprocessing binaries.
The solution is to recompile pygeoprocessing on your computer:

.. code-block:: console

$ pip uninstall -y pygeoprocessing
$ pip install pygeoprocessing --no-deps --no-binary :all:

Requirements
------------

Note the pip-installable requirements in `requirements.txt` are for best results, but older package versions may also work. If necessary, PyGeoprocessing can be installed without dependencies with `pip install --no-deps`.

Release History
===============

1.4.0 11/12/2018
----------------
* Adding error checking to ensure that `target_pixel_size` passed to
`warp_raster` and `align_and_resize_raster_stack` are validated to ensure
they are in the correct format. This solves an issue where an incorrect
value, such as a single numerical value, resolve into readable exception
messages.
* Added a `gdal_warp_options` parameter to `align_and_resize_raster_stack` and
`warp_raster` whose contents get passed to gdal.Warp's `warpOptions`
parameter. This was implemented to expose the CUTLINE_TOUCH_ALL
functionality but could be used for any gdal functionality.
* Modified `rasterize` API call to make `burn_values` and `option_list` both
optional parameters, along with error checking to ensure a bad input's
behavior is understood.
* Exposing GeoTIFF creation options for all the `pygeoprocessing.routing`
functions which create rasters. This is consistent with the creation
options exposed in the main `pygeoprocessing` API.
* Removing ``'mean_pixel_size'`` as a return value from ``get_raster_info``,
this is because this parameter is easily misused and easily calculated if
needed. This is a "What good programmers need, not what bad programmers
want." feature.

1.3.1 10/25/2018
----------------
* Hotfix to patch an infinite loop when aggregating upstream or downstream
with custom rasters.

1.3.0 10/25/2018
-----------------
* Fixed a handful of docstring errors.
* Improved runtime of ``zonal_statistics`` by a couple of orders of magnitude
for large vectors by using spatial indexes when calculating disjoint polygon
overlap sets, using database transactions, and memory buffers.
* Improved runtime performance of ``reproject_vector`` by using database
transactions.
* Improved logging for long runtimes in ``zonal_statistics``.
* Changed ``zonal_statistics`` API and functionality to aggregate across the
FIDs of the aggregate vector. This is to be consistent with QGIS and other
zonal statistics functionality. Additionally, fixed a bug where very small
polygons might not get aggregated if they lie in the same pixel as another
polygon that does not intersect it. The algorithm now runs in two passes:
* aggregate pixels whose centers intersect the aggregate polygons
* any polygons that were not aggregated are geometrically intersected
with pixels to determine coverage.
* Removed the ``calculate_raster_stats`` function since it duplicates GDAL
functionality, but with a slower runtime, and now functions in
``pygeoprocessing`` that create rasters also calculate stats on the fly if
desired.
* Fixes an issue in `get_raster_info` and `get_vector_info` where the path to
the raster/vector includes non-standard OS pathing (such as a NETCDF), info
will still calculate info.
* Added functionality to `align_raster_stack` and `warp_raster` to define a
base spatial reference system for rasters if not is not defined or one wishes
to override the existing one. This functionality is useful when reprojecting
a rasters that does not have a spatial reference defined in the dataset but
is otherwise known.
* Added a `weight_raster_path_band` parameter to both `flow_accumulation_d8`
and `flow_accumulation_mfd` that allows the caller to use per-pixel weights
from a parallel raster as opposed to assuming a weight of 1 per pixel.
* Added a `weight_raster_path_band` parameter to both
`distance_to_channel_mfd` and `distance_to_channel_d8` that allows the
caller to use per-pixel weights from a parallel raster as opposed to
assuming a distance of 1 between neighboring pixels or sqrt(2) between
diagonal ones.
* Added an option to `reproject_vector` that allows a caller to specify which
fields, if any, to copy to the target vector after reprojection.
* Adding a check in `align_and_resize_raster_stack` for duplicate target
output paths to avoid problems where multiple rasters are being warped to
the same path.
* Created a public `merge_bounding_box_list` function that's useful for union
or intersection of bounding boxes consistent with the format in
PyGeoprocessing.
* Added functionality in `align_and_resize_raster_stack` and `warp_raster`
to use a vector to mask out pixel values that lie outside of the polygon
coverage area. This parameter is called `vector_mask_options` and is
fully documented in both functions. It is similar to the cutline
functionality provided in `gdal.Warp`.
* Fixed an issue in the `flow_accumulation_*` functions where a weight raster
whose values were equal to the nodata value of the flow accumulation raster
OR simply nodata would cause infinite loops.

1.2.3 (7/25/2018)
-----------------
* Exposing a parameter and setting reasonable defaults for the number of
processes to allocate to ``convolve_2d`` and ``warp_raster``. Fixes an issue
where the number of processes could exponentiate if many processes were
calling these functions.
* Fixing an issue on ``zonal_statistics`` and ``convolve_2d`` that would
attempt to both read and write to the target raster with two different GDAL
objects. This caused an issue on Linux where the read file was not caught up
with the written one. Refactored to use only one handle.
* Fixing a rare race condition where an exception could occur in
``raster_calculator`` that would be obscured by an access to an object that
had not yet been assigned.
* ``align_and_resize_raster_stack`` now terminates its process pool.
* Increased the timeout in joining ``raster_calculator``'s stats worker.
On a slow system 5 seconds was not quite enough time.

1.2.2 (7/25/2018)
-----------------
* Hotfixed a bug that would cause numpy arrays to be treated as broadcastable
even if they were passed in "raw".

1.2.1 (7/22/2018)
-----------------
* Fixing an issue with `warp_raster` that would round off bounding boxes
for rasters that did not fit perfectly into the target raster's provided
pixel size.
* Cautiously `join`ing all process pools to avoid a potential bug where a
deamonized subprocess in a process pool may still have access to a raster
but another process may require write access to it.

1.2.0 (7/19/2018)
-----------------

* Several PyGeoprocessing functions now take advantage of multiple CPU cores:
* `raster_calculator` uses a separate thread to calculate raster statistics
in a `nogil` section of Cython code. In timing with a big rasters we
saw performance improvements of about 35%.
* `align_and_resize_raster_stack` uses as many CPU cores, up to the number
of CPUs reported by multiprocessing.cpu_count (but no less than 1), to
process each raster warp while also accounting for the fact that
`gdal.Warp` uses 2 cores on its own.
* `warp_raster` now directly uses `gdal.Warp`'s multithreading directly.
In practice it seems to utilize two cores.
* `convolve_2d` attempts to use `multiprocessing.cpu_count` cpus to
calculate separable convolutions per block while using the main thread to
aggregate and write the result to the target raster. In practice we saw
this improve runtimes by about 50% for large rasters.
* Fixed a bug that caused some nodata values to not be treated as nodata
if there was a numerical roundoff.
* A recent GDAL upgrade (might have been 2.0?) changed the reference to
nearest neighbor interpolation from 'nearest' to 'near'. This PR changes
PyGeoprocessing to be consistent with that change.
* ``raster_calculator`` can now also take "raw" arguments in the form of a
(value, "raw") tuple. The parameter `value` will be passed directly to
`local_op`. Scalars are no longer a special case and need to be passed as
"raw" parameters.
* Raising `ValueError` in `get_raster_info` and `get_vector_info` in cases
where non-filepath non-GIS values are passed as parameters. Previously
such an error would result in an unhelpful error in the GDAL library.

1.1.0 (7/6/2018)
----------------
* PyGeoprocessing now supports Python 2 and 3, and is tested on python 2.7
and 3.6 Testing across multiple versions is configured to be run via
``tox``.
* After testing (tox configuration included under ``tox-libcompat.ini``), numpy
requirement has been dropped to ``numpy>=1.10.0`` and scipy has been modified
to be ``scipy>=0.14.1,!=0.19.1``.
* A dependency on ``future`` has been added for compatibility between python
versions.
* Fixed a crash in ``pygeoprocessing.routing.flow_dir_mfd`` and
``flow_dir_d8`` if a base raster was passed in that did not have a power of
two blocksize.
* ``raster_calculator`` can now take numpy arrays and scalar values along with
raster path band tuples. Arrays and scalars are broadcast to the raster size
according to numpy array broadcasting rules.
* ``align_and_resize_raster_stack`` can now take a desired target projection
which causes all input rasters to be warped to that projection on output.

1.0.1 (5/16/2018)
-----------------
* Hotfix patch to remove upper bound on required numpy version. This was
causing a conflict with InVEST's looser requirement. Requirement is now
set to >=1.13.0.

1.0.0 (4/29/2018)
-----------------
* This release marks a feature-complete version of PyGeoprocessing with a
full suite of routing and geoprocessing capabilities.
* `pygeoprocessing.routing` module has a `flow_dir_mfd` function that
calculates a 32 bit multiple flow direction raster.
* `pygeoprocessing.routing` module has a `flow_accumulation_mfd` function that
uses the flow direction raster from `pygeoprocessing.routing.flow_dir_mfd`
to calculate a per-pixel continuous flow accumulation raster.
* `pygeoprocessing.routing` module has a `distance_to_channel_mfd` function
that calculates distance to a channel raster given a pygeoprocessing MFD
raster.
* `pygeoprocessing.routing` module has a `distance_to_channel_d8` function
that calculates distance to a channel raster given a pygeoprocessing D8
raster.

0.7.0 (4/18/2018)
-----------------
* Versioning is now handled by ``setuptools_scm`` rather than
``natcap.versioner``. ``pygeoprocessing.__version__`` is now fetched from
the package metadata.
* Raster creation defaults now set "COMPRESS=LZW" for all rasters created in
PyGeoprocessing, including internal temporary rasters. This option was
chosen after profiling large raster creation runs on platter hard drives.
In many cases processing time was dominated by several orders of magnitude
as a write-to-disk. When compression is turned on overall runtime of very
large rasters is significantly reduced. Note this otherwise increases the
runtime small raster creation and processing by a small amount.
* `pygeoprocessing.routing` module now has a `fill_pits`, function which
fills hydrological pits with a focus on runtime efficiency, memory space
efficiency, and cache locality.
* `pygeoprocessing.routing` module has a `flow_dir_d8` that uses largest
slope to determine the downhill flow direction.
* `pygeoprocessing.routing` module has a `flow_accumulation_d8` that uses
a pygeoprocessing D8 flow direction raster to calculate per-pixel flow
accumulation.
* Added a `merge_rasters` function to `pygeoprocessing` that will mosaic a
set of rasters in the same projection, pixel size, and band count.

0.6.0 (1/10/2017)
-----------------
* Added an optional parameter to `iterblocks` to allow the `largest_block` to
be set something other than the PyGeoprocessing default. This in turn
allows the `largest_block` parameter in `raster_calculator` to be passed
through to `iterblocks`.
* Upgraded PyGeoprocessing GDAL dependency to >=2.0.
* Added a `working_dir` optional parameter to `zonal_statistics`,
`distance_transform_edt`, and `convolve_2d` which specifies a directory in
which temporary files will be created during execution of the function.
If set to `None` files are created in the default system temporary
directory.

0.5.0 (9/14/2017)
-----------------
* Fixed an issue where NETCDF files incorrectly raised Exceptions in
`raster_calculator` and `rasterize` because they aren't filepaths.
* Added a NullHandler so that users wouldn't get an error that a logger
handler was undefined.
* Added `ignore_nodata`, `mask_nodata`, and `normalize_kernel` options to
`convolve_2d` which make this function capable of adapting the nodata
overlap with the kernel rather than zero out the result, as well as on
the fly normalization of the kernel for weighted averaging purposes. This
is in part to make this functionality more consistent with ArcGIS's
spatial filters.

0.4.4 (8/18/2017)
-----------------
* When testing for raster alignment `raster_calculator` no longer checks the
string equality for projections or geotransforms. Instead it only checks
raster size equality. This fixes issues where users rasters DO align, but
have a slightly different text format of the WKT of projection. It also
abstracts the problem of georeferencing away from raster_calculator that is
only a grid based operation.

0.4.3 (8/16/2017)
-----------------
* Changed the error message in `reclassify_raster` so it's more informative
about how many values are missing and the values in the input lookup table.
* Added an optional parameter `target_nodata` to `convolve_2d` to set the
desired target nodata value.

0.4.2 (6/20/2017)
-----------------
* Hotfix to fix an issue with `iterblocks` that would return signed values on
unsigned raster types.
* Hotfix to correctly cite Natural Capital Project partners in license and
update the copyright year.
* Hotfix to patch an issue that gave incorrect results in many PyGeoprocessing
functions when a raster was passed with an NoData value. In these cases the
internal raster block masks would blindly pass through on the first row
since a test for `numpy.ndarray == None` is `False` and later `x[False]`
is the equivalent of indexing the first row of the array.

0.4.1 (6/19/2017)
-----------------
* Non-backwards compatible refactor of core PyGeoprocessing geoprocessing
pipeline. This is to in part expose only orthogonal functionality, address
runtime complexity issues, and follow more conventional GIS naming
conventions. Changes include:
* Full test coverage for `pygeoprocessing.geoprocessing` module
* Dropping "uri" moniker in lieu of "path".
* If a raster path is specified and operation requires a single band,
argument is passed as a "(path, band)" tuple where the band index starts
at 1 as convention for raster bands.
* Shapefile paths are assumed to operate on the first layer. It is so
rare for a shapefile to have more than one layer, functions that would
be confused by multiple layers have a layer_index that defaults to 0
that can be overridden in the call.
* Be careful, many of the parameter orders have been changed and renamed.
Generally inputs come first, outputs last. Input parameters are
often prefixed with "base_" while output parameters are prefixed with
"target_".
* Functions that take rasters as inputs must have their rasters aligned
before the call to that function. The function
`align_and_resize_raster_stack` can handle this.
* `vectorize_datasets` refactored to `raster_calculator` since that name
is often used as a convention when referring to raster calculations.
* `vectorize_points` refactored to meaningful `interpolate_points`.
* `aggregate_by_shapefile` refactored to `zonal_statistics` and now
returns a dictionary rather than a named tuple.
* All functions that create rasters expose the underlying GeoTIFF options
through a default parameter `gtiff_creation_options` which default to
"('TILED=YES', 'BIGTIFF=IF_SAFER')".
* Individual functions for raster and vector properties have been
aggregated into `get_raster_info` and `get_vector_info` respectively.
* Introducing `warp_raster` to wrap GDAL's `ReprojectImage` functionality
that also works on bounding box clips.
* Removed the `temporary_filename()` paradigm. Users should manage
temporary filenames directly.
* Numerous API changes from the 0.3.x version of PyGeoprocessing.
* Fixing an issue with aggregate_raster_values that caused a crash if feature
IDs were not in increasing order starting with 0.
* Removed "create_rat/create_rat_uri" and migrated it to
natcap.invest.wind_energy; the only InVEST model that uses that function.
* Fixing an issue with aggregate_raster_values that caused a crash if feature IDs were not in increasing order starting with 0.
* Removed "create_rat/create_rat_uri" and migrated it to natcap.invest.wind_energy; the only InVEST model that uses that function.

0.3.3 (2/9/2017)
----------------
* Fixing a memory leak with large polygons when calculating disjoint set.

0.3.2 (1/24/2017)
-----------------
* Hotfix to patch an issue with watershed delineation packing that causes some field values to lose precision due to default field widths being set.

0.3.1 (1/18/2017)
-----------------
* Hotfix patch to address an issue in watershed delineation that doesn't pack the target watershed output file. Half the shapefile consists of features polygonalized around nodata values that are flagged for deletion, but not removed from the file. This patch packs those features and returns a clean watershed.

0.3.0 (10/21/2016)
------------------
* Added `rel_tol` and `abs_tol` parameters to `testing.assertions` to be
consistent with PEP485 and deal with real world testing situations that
required an absolute tolerance.
* Removed calls to ``logging.basicConfig`` throughout pygeoprocessing. Client
applications may need to adjust their logging if pygeoprocessing's log
messages are desired.
* Added a flag to `aggregate_raster_values_uri` that can be used to indicate
incoming polygons do not overlap, or the user does not care about overlap.
This can be used in cases where there is a computational or memory
bottleneck in calculating the polygon disjoint sets that would ultimately be
unnecessary if it is known a priori that such a check is unnecessary.
* Fixed an issue where in some cases different nodata values for 'signal' and
'kernel' would cause incorrect convolution results in `convolve_2d_uri`.
* Added functionality to `pygeoprocessing.iterblocks` to iterate over largest
memory aligned block that fits into the number of elements provided by the
parameter. With default parameters, this uses a ceiling around 16MB of
memory per band.
* Added functionality to `pygeoprocessing.iterblocks` to return only the
offset dictionary. This functionality would be used in cases where memory
aligned writes are desired without first reading arrays from the band.
* Refactored `pygeoprocessing.convolve_2d_uri` to use `iterblocks` to take
advantage of large block sizes for FFT summing window method.
* Refactoring source side to migrate source files from [REPO]/pygeoprocessing
to [REPO]/src/pygeoprocessing.
* Adding a pavement script with routines to fetch SVN test data, build a
virtual environment, and clean the environment in a Windows based operating
system.
* Adding `transform_bounding_box` to calculate the largest projected bounding
box given the four corners on a local coordinate system.
* Removing GDAL, Shapely from the hard requirements in setup.py. This will
allow pygeoprocessing to be built by package managers like pip without these
two packages being installed. GDAL and Shapely will still need to be
installed for pygeoprocessing to run as expected.
* Fixed a defect in ``pygeoprocessing.testing.assert_checksums_equal``
preventing BSD-style checksum files from being analyzed correctly.
* Fixed an issue in reclassify_dataset_uri that would cause an exception if
the incoming raster didn't have a nodata value defined.
* Fixed a defect in ``pygeoprocessing.geoprocessing.get_lookup_from_csv``
where the dialect was unable to be detected when analyzing a CSV that was
larger than 1K in size. This fix enables the correct detection of comma or
semicolon delimited CSV files, so long as the header row by itself is not
larger than 1K.
* Intra-package imports are now relative. Addresses an import issue for users
with multiple copies of pygeoprocessing installed across multiple Python
installations.
* Exposed cython routing functions so they may be imported from C modules.
* `get_lookup_from_csv` attempts to determine the dialect of the CSV instead
of assuming comma delimited.
* Added relative numerical tolerance parameters to the PyGeoprocessing raster
and csv tests with in the same API style as `numpy.testing.allclose`.
* Fixed an incomparability with GDAL 1.11.3 bindings that expects a boolean
type in `band.ComputeStatistics`. Before this fix PyGeoprocessing would
crash with a TypeError on many operations.
* Fixed a defect in pygeoprocessing.routing.calculate_transport where the
nodata types were cast as int even though the base type of the routing
rasters were floats. In extreme cases this could cause a crash on a type
that could not be converted to an int, like an `inf`, and in subtle cases
this would result in nodata values in the raster being ignored during
routing.
* Added functions to construct raster and vectors on disk from reasonable
datatypes (numpy matrices for rasters, lists of Shapely geometries for
vectors).
* Fixed an issue where reproject_datasource_uri would add geometry that
couldn't be projected directly into the output datasource. Function now
only adds geometries that transformed without error and reports if any
features failed to transform.
* Added file flushing and dataset swig deletion in reproject_datasource_uri to
handle a race condition that might have been occurring.
* Fixed an issue when "None" was passed in on new raster creation that would
attempt to directly set that value as the nodata value in the raster.
* Added basic filetype-specific assertions for many geospatial filetypes, and
tests for these assertions. These assertions are exposed in
`pygeoprocessing.testing`.
* Pygeoprocessing package tests can be run by invoking
`python setup.py nosetests`. A subset of tests may also be run from an
installed pygeoprocessing distribution by calling `pygeoprocessing.test()`.
* Fixed an issue with reclassify dataset that would occur when small rasters
whose first memory block would extend beyond the size of the raster thus
passing in "0" values in the out of bounds area. Reclassify dataset
identified these as valid pixels, even though vectorize_datsets would mask
them out later. Now vectorize_datasets only passes memory blocks that
contain valid pixel data to its kernel op.
* Added support for very small AOIs that result in rasters less than a pixel
wide. Additionally an `all_touched` flag was added to allow the
ALL_TOUCHED=TRUE option to be passed to RasterizeLayer in the AOI mask
calculation.
* Added watershed delineation routine to
pygeoprocessing.routing.delineate_watershed. Operates on a DEM and point
shapefile, optionally snaps outlet points to nearest stream as defined by a
thresholded flow accumulation raster and copies the outlet point fields into
the constructed watershed shapefile.
* Fixing a memory leak in block caches that held on to dataset, band, and
block references even after the object was destroyed.
* Add an option to route_flux that lets the current pixel's source be included
in the flux, or not. Previous version would include on the source no matter
what.
* Now using natcap.versioner for versioning instead of local versioning logic.

0.2.2 (2015-05-07)
------------------

* Adding MinGW-specific compiler flags for statically linking pygeoprocessing
binaries against libstdc++ and libgcc. Fixes an issue on many user's
computers when installing from a wheel on the Python Package Index without
having two needed DLLs on the PATH, resulting in an ImportError on pygeoprocessing.geoprocessing_core.pyd.
* Fixing an issue with versioning where 'dev' was displayed instead of the
version recorded in pygeoprocessing/__init__.py.
* Adding all pygeoprocessing.geoprocessing functions to
pygeoprocessing.__all__, which allows those functions to appear when
calling help(pygeoprocessing).
* Adding routing_core.pxd to the manifest. This fixes an issue where some
users were unable to compiler pygeoprocessing from source.

0.2.1 (2015-04-23)
------------------

* Fixed a bug on the test that determines if a raster should be memory
blocked. Rasters were not getting square blocked if the memory block was
row aligned. Now creates 256x256 blocks on rasters larger than 256x256.
* Updates to reclassify_dataset_uri to use numpy.digitize rather than Python
loops across the number of keys.
* More informative error messages raised on incorrect bounding box mode.
* Updated docstring on get_lookup_from_table to indicate the headers are case
insensitive.
* Added updates to align dataset list that report which dataset is being
aligned. This is helpful for logging feedback when many datasets are passed
in that don't take long enough to get a report from the underlying reproject
dataset function.
* pygeoprocessing.routing.routing_core includes pxd to be \`cimport`able from
a Cython module.

0.2.0 (2015-04-14)
------------------

* Fixed a library wide issue relating to the underlying numpy types of
GDT_Byte Datasets. Now correctly identify the signed and unsigned versions
and removed all instances where code used to mod byte data to unsigned data
and correctly creates signed/unsigned byte datasets during resampling.
* Removed extract_band_and_nodata function since it exposes the underlying
GDAL types.
* Removed reclassify_by_dictionary since reclassify_dataset_uri provided
almost the same functionality and was widely used.
* Removed the class OrderedDict that was not used.
* Removed the function calculate_value_not_in_dataset since it loaded the
entire dataset into memory and was not useful.

0.1.8 (2015-04-13)
------------------

* Fixed an issue on reclassifying signed byte rasters that had negative nodata
values but the internal type stored for vectorize datasets was unsigned.

0.1.7 (2015-04-02)
------------------

* Package logger objects are now identified by python hierarchical package
paths (e.g. pygeoprocessing.routing)
* Fixed an issue where rasters that had undefined nodata values caused
striping in the reclassify_dataset_uri function.

0.1.6 (2015-03-24)
------------------

* Fixing LICENSE.TXT to .txt issue that keeps reoccurring.

0.1.5 (2015-03-16)
------------------

* Fixed an issue where int32 dems with INT_MIN as the nodata value were being
treated as real DEM values because of an internal cast to a float for the
nodata type, but a cast to double for the DEM values.
* Fixed an issue where flat regions, such as reservoirs, that could only drain
off the edge of the DEM now correctly drain as opposed to having undefined
flow directions.

0.1.4 (2015-03-13)
------------------

* Fixed a memory issue for DEMs on the order of 25k X 25k, still may have
issues with larger DEMs.

0.1.3 (2015-03-08)
------------------

* Fixed an issue so tox correctly executes on the repository.
* Created a history file to document current and previous releases.
* Created an informative README.rst.

0.1.2 (2015-03-04)
------------------

* Fixing issue that caused "LICENSE.TXT not found" during pip install.

0.1.1 (2015-03-04)
------------------

* Fixing issue with automatic versioning scheme.

0.1.0 (2015-02-26)
------------------

* First release on PyPI.