Title of Manuscript: Priority-Flood: An Optimal Depression-Filling and Watershed-Labeling Algorithm for Digital Elevation Models
Authors: Richard Barnes, Clarence Lehman, David Mulla
Corresponding Author: Richard Barnes (rbarnes@umn.edu)
DOI Number of Manuscript 10.1016/j.cageo.2013.04.024
Code Repositories
This repository contains a reference implementation of the algorithms presented in the manuscript above. These implementations were used in performing the tests described in the manuscript. The manuscript contains pseudocode for (most of) the implementations included here.
Don't. If you want to use these algorithms as an out-of-the-box terrain analysis system, please download RichDEM.
This repo references the RichDEM terrain analysis softare, of which these
algorithms are all a part. main.cpp will run all of the algorithms mentioned
above. The #include directives in main.cpp identify the necessary RichDEM
libraries for using these implementations for your own work.
After cloning this repo you must acquire RichDEM by running:
git submodule init
git submodule update
To compile the programs run:
make
The result is a program called priority_flood.exe.
The program is run by typing:
./priority_flood.exe <ALGORITHM NUMBER> <INPUT DEM> <OUTPUT FILE>
./priority_flood.exe 3 input-elevations.tif output.tif
The algorithms available are described briefly below and in greater detail in the manuscript.
This program reads in a digital elevation model (DEM) file specified on the command line. The file may be of any type recognised by GDAL. The program will run one of the algorithms described in the manuscript (and below), store the result in an output file, and report how long this took.
The program outputs a digital elevaiton model without any internally-draining depressions/pits or digital dams. The output is in GeoTIFF format.
-
Algorithm 1: Priority-Flood This algorithm alters the input DEM to produce an output with no depressions or digital dams. Every cell which would have been in a depression is increased to the level of that depression's outlet, leaving a flat region in its place. It runs slower than Algorithm 2, but is otherwise the same.
-
Algorithm 2: Improved Priority-Flood This algorithm alters the input DEM to produce an output with no depressions or digital dams. Every cell which would have been in a depression is increased to the level of that depression's outlet, leaving a flat region in its place. It runs slower than Algorithm 2, but is otherwise the same.
-
Algorithm 3: Priority-Flood+Epsilon This algorithm alters the input DEM to produce an output with no depressions or digital dams. Every cell which would have been in a depression is increased to the level of that depression's output, plus an additional increment which is sufficient to direct flow to the periphery of the DEM.
-
Algorithm 4: Priority-Flood+FlowDirs This algorithm determines a D8 flow direction for every cell in the DEM by implicitly filling depressions and eliminating digital dams. Though all depressions are guaranteed to drain, local elevation information is still used to determine flow directions within a depression. It is, in essence, a depression-carving algorithm.
-
Algorithm 5: Priority-Flood+Watershed Labels For each cell c in a DEM, this algorithm determines which cell on the DEM's periphery c will drain to. c is then given a label which corresponds to the peripheral cell. All cells bearing a common label belong to the same watershed.
-
Algorithm 6: Zhou et al. (2016) Priority-Flood Zhou et al. have developed an even more efficient variant of the Priority-Flood. It provides an output identical to Algorithms 1 and 2, but runs in approximately half the time.
Algorithm 4: Priority-Flood+FlowDirs and its output use the D8 neighbour system to indicate flow directions. In this system all the flow from a central cell is directed to a single neighbour which is represented by a number according to the following system where 0 indicates the central cell.
234
105
876
All of the algorithms assume that cells marked as having NoData will have extremely negative numerical values: less than the value of any of the actual data. NaN is considered to be less than all values, including negative infinity.
Work by Cris Luengo on the speed of various priority queue algorithms is discussed in the manuscript. His website providing code for his implementatations is here.
Commit 31c4e31a6e765f9a (02016-09-14) updated this repo to rely explicitly on the RichDEM codebase.
Commit 51f9a7838d3e88628ef6c74846edd0cb18e7ffe6 (02015-09-25) introduced a number of changes to the code versus what was originally published with the manuscript. The old codebase uses ASCII-formatted data for input and output; the new codebase uses GDAL to handle many kinds of data.
The old codebase had the advantage of not relying on external libraries and being readily accessible to all parties. It had the disadvantage of being a slow, clumsy, and limited way to work with the data. As of 02015-09-25, the code requires the use of the GDAL library, greatly expanding the data formats and data types which can be worked with, as well as greatly speeding up I/O.
Note that using the aforementioned 51f9a7838d directly will result in silent
casting of your data to the float type; commit
8b11f535af23368d3bd26609cc88df3dbb7111f1 (02015-09-28) fixes this issue.