This tool provides a raster of a Digital Elevation Model (DEM) over an area of interest utilizing global or continental, publicly available tile sets such as the Global Copernicus Digital Elevation Model at 30 meter resolution. See the Datasets section below for all the tiles supported and their shortnames. This tool also performs some standard transformations for processing such as:

• the conversion of the vertical datum from a reference geoid to the WGS84 ellipsoidal
• in-memory merging of various tiles preserving the source rasters
• optional half-pixel georferencing (primarily for the legacy DEMs SRTM and NASADEM, see the section on Transformations)

We rely on the GIS formats from rasterio. The API can be summarized as

from dem_stitcher import stitch_dem

# as xmin, ymin, xmax, ymax in epsg:4326
bounds = [-119.085, 33.402, -118.984, 35.435]

X, p = stitch_dem(bounds,
                  dem_name='glo_30',  # Global Copernicus 30 meter resolution DEM
                  dst_ellipsoidal_height=False)
# X is an m x n numpy array
# p is a dictionary (or a rasterio profile) including relevant GIS metadata; CRS is epsg:4326

Then, to save the DEM raster to disk:

import rasterio

with rasterio.open('dem.tif', 'w', **p) as ds:
   ds.write(X, 1)

The rasters are returned in the global lat/lon projection epsg:4326 and the API assumes that bounds are supplied in this format. We try to do the resampling and transformations all in memory to avoid unnecessary i/o and forgotten files.

Matching the NISAR DEM

The default keyword arguments of stitch_dem reproduce the NISAR DEM from glo_30, i.e.

X, p = stitch_dem(bounds,
                  dem_name='glo_30',
                  dst_ellipsoidal_height=True,   # remove EGM2008 
                  dst_area_or_point=None,        # inherit the source registration
                  dst_resolution=None)           # keep the native tile resolution

agrees with stitch_dem(bounds, dem_name='nisar_dem') pixel-for-pixel in georeferencing and to ~1 mm in height (the residual is NISAR DEM's EGM2008 grid/interpolation vs the 1 arcminute grid used here). This is verified by an integration test over randomly selected tiles and demonstrated in this notebook.

Installation

We recommend pixi. To add dem_stitcher to your own project:

pixi add dem_stitcher

To work from a clone of this repo (which installs dem_stitcher in editable mode along with the notebook and test dependencies):

git clone https://github.com/ACCESS-Cloud-Based-InSAR/dem-stitcher.git
cd dem-stitcher
pixi install
pixi run python -c "import dem_stitcher; print(dem_stitcher.__version__)"

For JupyterLab (with jupyter-collaboration for real-time collaborative editing):

pixi run jupyter lab

The default environment is the only one installed locally and uses python 3.14. Environments named py311 through py313 are declared for the CI matrix and are built on demand with pixi run -e py312 ...; remove one again with pixi clean -e py312.

Other installation methods

dem_stitcher can also be installed into a conda environment with

conda install -c conda-forge dem_stitcher

or into a virtual environment with

python -m pip install dem_stitcher

Currently, python 3.11+ is supported.

With ISCE2 or gdal

Although the thrust of using this package is for staging DEMs for InSAR (particularly ISCE2), testing and maintaining suitable environments to use with InSAR processors is beyond the scope of what we are attempting to accomplish here. We provide an example notebook here that demonstrates how to stage a DEM for ISCE2, which requires additional packages than required for the package on its own. For the notebook, we use the environment found in environment.yml of the Dockerized TopsApp repository, used to generate interferograms (GUNWs) in the cloud.

About the raster metadata

The creation metadata unrelated to georeferencing (e.g. the compress key or various other options here) returned in the dictionary profile from the stitch_dem API is copied directly from the source tiles being used if they are GeoTiff formatted (such as glo_30) else the creation metadata are copied from the GeoTiff Default Profile in rasterio (see here excluding nodata and dtype). Such metadata creation options are beyond the scope of this library.

Credentials

The accessing of NASADEM, SRTM, and the NISAR DEM require earthdata login credentials to be put into the ~/.netrc file. If these are not present, the stitcher will fail with ValueError asking you to update the ~/.netrc. The appropriate entry appears as:

machine urs.earthdata.nasa.gov
    login <username>
    password <password>

For nisar_dem, the tiles are read directly with rasterio/GDAL, so the stitcher opens and reads them within a rasterio.Env that sets GDAL_HTTP_NETRC, GDAL_HTTP_COOKIEFILE, and GDAL_HTTP_COOKIEJAR so GDAL can authenticate through the Earthdata cloud redirect. If you get nisar_dem urls from get_dem_tile_paths and open them yourself, use the same environment:

import rasterio
from dem_stitcher.credentials import earthdata_gdal_env

with earthdata_gdal_env():
    with rasterio.open(url) as ds:
        dem_arr = ds.read(1)

All the reads this library performs (tiles and geoids alike) happen within a rasterio.Env that sets GDAL_DISABLE_READDIR_ON_OPEN='EMPTY_DIR', which stops GDAL from probing for sidecar files (*.aux.xml, *.msk, ...) next to remote rasters - those probes are 403s that GDAL logs as CPLE_AppDefined warnings. dem_stitcher.rio_tools.gdal_read_env() (included in earthdata_gdal_env() above) builds that environment if you are opening urls yourself.

Notebooks

We have notebooks to demonstrate common usage:

• Basic Demo
• Comparing DEMs
• Generating a VRT from source DEM tiles
• Staging a DEM for ISCE2 - this notebook requires the installation of a few extra libraries including ISCE2 via conda-forge

We also demonstrate how the tiles used to organize the urls for the DEMs were generated for this tool were generated in this notebook.

DEMs Supported

The DEMs that are currently supported are:

In [1]: from dem_stitcher.datasets import DATASETS; DATASETS
Out[1]: ['3dep', 'glo_30', 'glo_90', 'glo_90_missing', 'nasadem', 'nisar_dem', 'srtm_v3']

The shortnames aboves are the strings required to use stitch_dem. Below, we expound upon these DEM shortnames and link to their respective data repositories.

1. glo_30/glo_90: Copernicus GLO-30/GLO-90 DEM. The tile sets are the 30 and 90 meter resolution, respectively [link].
2. The USGS DEM 3dep: 3Dep 1/3 arc-second over North America - we are storing the ~10 meter resolution dataset. There are many more as noted here. The files for these DEMs are here
3. srtm_v3: SRTM v3 [link] - tiles are downloaded from the LP DAAC Earthdata Cloud archive and require Earthdata credentials in ~/.netrc (see Credentials)
4. nasadem: Nasadem [link] - tiles are downloaded from the LP DAAC Earthdata Cloud archive and require Earthdata credentials in ~/.netrc (see Credentials)
5. glo_90_missing: these are tiles that are in glo_90 but not in glo_30. They are over the countries Armenia and Azerbaijan. Used internally to help fill in gaps in coverage of glo_30.
6. nisar_dem: the NISAR mission DEM v1.2 [link] - the Copernicus GLO-30 (2023_1) re-referenced to the WGS84 ellipsoid by JPL NISAR team. Only the epsg:4326 tile set is cataloged (global coverage including ocean tiles). Since the EGM2008 geoid has already been removed from this DEM, only dst_ellipsoidal_height=True is supported and no geoid is applied by the stitcher. Requires Earthdata credentials in ~/.netrc (see Credentials).

All the tiles are given in lat/lon CRS (i.e. epsg:4326 for global tiles or epsg:4269 for USGS tiles in North America). A notable omission to the tile sets is the Artic DEM here, which is suitable for DEMs merged at the north pole of the globe due to lat/lon distortion.

If there are issues with obtaining dem tiles from urls embedded within the geoparquet tile tables (e.g. a 404 error as here), please see the Development section below and/or open an issue ticket.

DEM Transformations

Wherever possible, we do not resample the original DEMs unless specified by the user to do so. When extents are specified, we obtain the the minimum pixel extent within the merged tile DEMs that contain that extent. Any required resampling (e.g. updating the CRS or updating the resolution because the tiles have non-square resolution at high latitudes) is done after these required translations. We importantly note that order in which these transformations are done is crucial as affine transformations are not commutative. Here are some notes/discussions:

1. All DEMs are resampled to epsg:4326. Most DEMs are already in this CRS except the USGS DEMs over North America, which are in epsg:4269, whose xy projection is also lon/lat but has different vertical data. For our purposes, these two CRSs are almost identical. The nuanced differences between these CRS's is noted here.

2. All DEM outputs will have origin and pixel spacing aligning with the original DEM tiles unless a resolution for the final product is specified, which will alter the pixel spacing.

3. The 'AREA_OR_POINT' gdal tag should not have any impact on georeferencing as noted in the data model page and quoted below:

   AREA_OR_POINT: May be either "Area" (the default) or "Point". Indicates whether a pixel value should be assumed to represent a sampling over the region of the pixel or a point sample at the center of the pixel. This is not intended to influence interpretation of georeferencing which remains area oriented.

   As such, the default here is to assume the DEMs source georeferencing is correct and no transformation is required on the DEM tiles nor geoid. Specifically, dst_area_or_point=None) and the output inherits the source registration: 'Point' for all supported DEMs except 3dep, which is 'Area'. If the geoid is removed, the geoid is removed on the native grid beforehand (no pixel translations). However, as noted on the USGS DEM page, the SRTM and NED DEMs are expected to have a half pixel shift, the image from the link is shared below:

   

   In this interpretation, georeferenced rasters can be tied to map coordintaes using either (a) upper-left corners of pixels or (b) the pixel centers i.e. Point and Area tags in gdal, respectively, and seen as {'AREA_OR_POINT: 'Point'}. Note that tying a pixel to the upper-left cortner (i.e. Area tag) is the default pixel reference for gdal as indicated here. Passing 'Area' or 'Point' explicitly in dst_area_or_point will translate the output transform by half a pixel with this intrepretation. Here is quite a long discussion about pixel vs. area tag within gdal: SRTM v3, NASADEM, and GLO-30 are Pixel-centered, i.e. {'AREA_OR_POINT: 'Point'}.

4. Transform geoid heights to WGS84 Ellipsoidal height. This is done using the rasters here. We:

   • Interpolate the geoid (with cubic resampling) at the native DEM sample locations, i.e. before any Area/Point relabeling of the output grid, so that dst_area_or_point only shifts the output transform by half a pixel and never changes the height samples (see #151). Stitching glo_30 with dst_ellipsoidal_height=True agrees with the independently produced NISAR DEM (Copernicus GLO-30 with EGM2008 removed at the source by the NISAR team at JPL) at the millimeter level; the residual is the difference between NISAR DEM's EGM2008 grid/interpolation and the 1 arcminute grid used here. See this notebook.
   • For time-series analysis built on products stitched with versions before 3.0.0 (e.g. existing ARIA products), stitch_dem(..., geoid_correction_mode='aria-legacy') reproduces the pre-3.0.0 geoid correction bit-for-bit for DEMs delivered in epsg:4326: the geoid is sampled on the Area/Point-relabeled grid with bilinear resampling and, for 'Point', translated by half a geoid pixel — i.e. the #151 bias is reproduced intentionally, and a UserWarning is emitted on every call. Pre-3.0.0 versions also defaulted dst_area_or_point to 'Area', so pass it explicitly ('Point' for ARIA products) for full call-for-call parity.

5. All DEMs are converted to float32 and have nodata np.nan. Although this can increase data size of certain rasters (SRTM has integer heights), this ensures (a) easy comparison across DEMs and (b) no side-effects of the stitcher due to dtypes and/or nodata values. There is one caveat: the user can ensure that DEM nodata pixels are set to 0 using merge_nodata_value in stitch_dem, in which case 0 is filled in where np.nan was. We note specifying this "fill value" via merge_nodata_value does not change the nodata value of output DEM dataset (i.e. nodata in the rasterio profile will remain np.nan). When transforming to ellipsoidal heights and setting 0 as merge_nodata_value, the geoid values are filled in the DEMs nodata areas; if the geoid has nodata in the bounding box, this will be the source of subsequent no data. For reference, this datatype and nodata is specified in merge_tile_datasets in merge.py. Other nodata values can be specified outside the stitcher for the application of choice (e.g. ISCE2 requires nodata to be filled as 0).

There are some notebooks that illustrate how tiles are merged by comparing the output of our stitcher with the original tiles.

As a performance note, when merging DEM tiles, we merge the needed tiles within the extent in memory and this process has an associated overhead. The benefit is there is no unnecessary files saved locally and later releases optimize the in-memory transformations well.

Dateline support

We assume that the supplied bounds overlap the standard lat/lon CRS grid i.e. longitudes between -/+ 180 longitude and are within -/+ 90 latitude. If there is a single dateline crossing by the supplied bounds, then the tiles are wrapped the dateline and individually translated to a particular hemisphere dicated by the bounds provided to generate a continuous raster over the area provided. We assume a maximum of one dateline crossing in the bounds you specified (if you have multiple dateline crossings, then stitch_dem will run out of memory). Similar wrapping tiles around the North and South poles (i.e. at -/+ 90 latitude) is not supported (a different CRS is what's required) and an exception will be raised.

For Development

1. Clone this repo git clone https://github.com/ACCESS-Cloud-Based-InSAR/dem-stitcher.git
2. Navigate with your terminal to the repo.
3. Run pixi install - this creates the environment and installs dem_stitcher in editable mode.

Linting and formatting are handled by ruff and exposed as pixi tasks:

pixi run lint
pixi run format
pixi run fix

DEM Urls

If urls or readers need to be updated (they consistently do) or you want to add a new global or large DEM, then there are two points of contact:

1. The notebooks that format the geoparquet tile tables used for this library are here
2. The readers are here

The former is the more likely. When re-generating tiles, make sure to run all tests including integration tests (i.e. pytest tests). For example, if regenerating glo tiles, glo-30 requires both resolution parameters (30 meters and 90 meters) and an additional notebook for filling in missing 30 meter tiles. These should be clearly spelled out in the notebook linked above.

Testing

For the test suite, run pixi run pytest tests (or pixi run test, which skips the notebook tests).

There are two category of tests: unit tests and integration tests. The former can be run using pytest tests -m 'not integration' and similarly the latter with pytest tests -m 'integration'. Our unit tests are those marked without the integration tag (via pytest) that use synthetic data or data within the library to verify correct outputs of the library (e.g. that a small input raster is modified correctly). Integration tests ensure the dem-stitcher API works as expected, downloading the DEM tiles from their respective servers to ensure the stitcher runs to completion - the integration tests only make very basic checks to ensure the format of the ouptut data is correct (e.g. checking the output raster has a particular shape or that nodata is np.nan). Our integration tests also include tests that run the notebooks that serve as documentation via papermill (such tests have an additional tag notebook). Integration tests will require the ~/.netrc setup above and working internet. Our testing workflow via Github actions currently runs the entire test suite except those tagged with notebook, as these tests take considerably longer to run.

Contributing

We welcome contributions to this open-source package. To do so:

1. Create an GitHub issue ticket desrcribing what changes you need (e.g. issue-1)
2. Fork this repo
3. Make your modifications in your own fork
4. Make a pull-request (PR) in this repo with the code in your fork and tag the repo owner or a relevant contributor.

We use ruff to ensure some basic code quality (configured in pyproject.toml). These will be checked for each commit in a PR. Try to write tests wherever possible.

Support

1. Create an GitHub issue ticket desrcribing what changes you would like to see or to report a bug.
2. We will work on solving this issue (hopefully with you).

Acknowledgements

This tool was developed to support cloud SAR processing using ISCE2 and various research projects at JPL. The early work of this repository was done by Charlie Marshak, David Bekaert, Michael Denbina, and Marc Simard. Since the utilization of this package for GUNW generation (see this repo), a subset of the ACCESS team, including Joseph (Joe) H. Kennedy, Simran Sangha, Grace Bato, Andrew Johnston, and Charlie Marshak, have improved this repository greatly. In particular, Joe Kennedy has lead the inclusion/development of actions, tests, packaging, distribution (including PyPI and conda-forge) and all the things to make this package more reliable, accessible, readable, etc. Simran Sangha has helped make sure output rasters are compatible with ISCE2 and other important bug-fixes.