Analyzing Urban Land Use Changes in Urban Environments

Analyzing Urban Land Use Changes in Urban Environments or Using SLEUTH for Transportation Planning [GIS-T Workshop] Keith C. Clarke Department of Geography UC Santa Barbara

Summary l History/Background on SLEUTH l Model theory and operation l Data requirements l Calibration l Outputs l Model use forecasting l The Role of transportation

History/Background of SLEUTH

Complex systems theory l Non-linear dynamics l Behavior states: chaos, self-organization, stable l Phase transitions l Emergence l Cellular automata are simple examples of complex systems

Cellular Automata l Gridded world l Cells have finite states l Rules define state transitions l Time is incremental l Cells are autonomous, act as agents l Self-replicating machines: Von Neumann l Classic example is Conway’s LIFE

Urban Cellular Automata l Cells are pixels l States are land uses l Time is “units”, e. g. years l Rules determine growth and change l Different models have different rule sets l Many models now developed, few tested l Requiem for large scale models (Lee)

Background q Urban Dynamics Regional geographic analysis of temporal spatial data bases q About a dozen applications of three model versions q Retrospective and Analysis q q Cellular Automaton Model of Change § § Initial set of state conditions (binary to Anderson II) Change rules Independent agents Urban model drives deltatron model

Cellular urban modeling l l l Clarke cellular automaton urban growth model (UGM) Multiple applications (e. g. San Francisco, Washington/Baltimore) Project Gigalopolis New applications under way: Chicago, New York, Portland, Philadelphia, MAIA, Mexico City, Santa Barbara 1998/9 funding made model portable and web-based (USGS: EROS Data Center, EPA Collaboration) 1999 -01 work extends and integrates model with other efforts (LANL and USGS collaboration, NSF Urban Research Initiative, SBECP) EPA has provided significant input (Ron Matheney and Tom Cathey)

Gigalopolis: Project Goal l Use historical data for urban areas to simulate present day urbanization l Simulate using a Cellular Automaton Model (SLEUTH) l Run the model into the future l Simulate alternative futures l Compare across scale and cities l Apply to Urban Dynamics cities

Cellular land transition models l Increasingly important methods l Many different models l Link to CA and complex systems theory l Transition probability based l Deltatron changes weighted methods by forcing autocorrelation in change space l Allow modeling, visualization and experiments

Model handles land use l l l So far works at crudest level (Anderson Level 1) Calibration under way in MAIA and Lower 48 States (GIRAS, MRLC, Loveland) Needs two LULC layers Based on the concept of deltatrons Generates synthetic LU change based on transition matrix and enfored spatial/temporal autocorrelation Applies CA in change space

Land Use Change properties l Driver=urbanization l State probabilities (static, dynamic) l Class magnitudes l Spatial autocorrelation(s)

Project Web Site l Set of background materials, e. g. publications Documentation as web pages in HTML Source Code for model in C Version 3. 0 now on web for download Uses utilities and GD GIF libraries Parallel version requires MPI Set of sample calibration data demo_city l http: //www. ncgia. ucsb. edu/projects/gig/ncgia. html l l l

Project Web Site: Shareware C code and Documentation

Model Theory and Operation

Background q Urban Retrospectives q Geographic analysis q Urban dynamics q Model comparison q Model integration

The rules

Spontaneous Growth n urban settlements may occur anywhere on a landscape n f (diffusion coefficient, slope resistance)

Creation of new spreading centers § § § Some new urban settlements will become centers of further growth. Others will remain isolated. f (spontaneous growth, breed coefficient, slope resistance)

Organic Growth § The most common type of development occurs at urban edges and as in-filling § f (spread coefficient, slope resistance) §

Road Influenced Growth § Urbanization has a tendency to follow lines of transportation f (breed coefficient, road_gravity coefficient, slope resistance, diffusion coefficient) §

Behavior Rules T 0 spontaneous spreading center organic road influenced deltatron T 1

Behavior Rules T 0 T 1 spontaneous spreading center organic road influenced f (slope resistance, diffusion coefficient) f (slope resistance, breed coefficient) f (slope resistance, spread coefficient) f (slope resistance, diffusion coefficient, breed coefficient, road gravity) deltatron For i time periods (years)

Patterns/process of land cover change Introduction of new land cover type (invasion, diffusion) l Land cover class extension from edges (spread, contagion) l Perpetuation of change (lagged autocorrelation) l

Deltatron Dynamics: Land cover Delta-space To/From Transition matrix l Table of land cover class average slopes l Urbanization drives change within the model l Urban (and others) invariant class l

Deltatrons at work

A Deltatron is: “Bringer of change” (semi-independent agent) l Placeholder of where and what type of land cover transition took place during its lifetime l Tracks how much time has passed since a change has occurred (Lifetime) l Enforces spatial and temporal auto-correlation of land cover transitions by its life cycle l

Deltatron Land Cover Model Phase 1: Create change select random pixel For n new urban cells Select two land classes at random Create delta space Of the two: Find the land class most similar to current slope Average slope Transition Probability Matrix spread change land cover Check the transition probability

Deltatron Land Cover Model Phase 2: Perpetuate change search for change in the neighborhood find associated land cover transitions delta space Transition Probability Matrix create deltatrons n Age or kill deltatrons impose change in land cover n

Data Requirements

1900 q Slope q Land Cover q Excluded q Urban q Transportation q Hillshade 1925 1950 1975 2000

Thematic Data Input Consistency Between Data Layers l Hierarchy l More scope vs. Definition problematic with increase of temporal

Thematic Data Input Consistency Between Data Layers (cont. ) l Consistent data source l Project Specific Documentation

Thematic Data Input Data resolution l Optimum resolution of data layers is unknown l The SLEUTH can “work” for any data resolution l Tested at 30 m to 1 km l Roads least realistic at coarse scale

Thematic Data Input: Issues l Vertical Integration of Temporal Data Layers l Misregistration produced artificial change l Deurbanization particularly upsetting to model l Road breaks should be avoided

UGM Process Flow Data Set Preparation Create Geographic Temporal Database l l l Source data – historical maps, areal photographs, remotely sensed data, GIS vector/grid data Select by attribute – urban – transportation – landuse – excluded – slope Geo-registration – extent (lat, long) Data type standardization – vector to raster – Arc. Info vector data: LINEGRID or POLYGRID resolution (rows, columns)

UGM Process Flow Data Set Preparation Image Format Specifics Urban Values: 0 = not urban, 0 < n < 255 = urban Roads Values: 0 = not road, 0 < n < 255 = road

UGM Process Flow Data Set Preparation Image Format Specifics l l Landuse: any method can be used Values: Each value matches a given classification value. l l – 1 = urban, 2 = agriculture, 3 = rangeland, etc. Slope: the average percent slope of the terrain is derived from a DEM Values: 0 - 100

UGM Process Flow Data Set Preparation Image Format Specifics l Excluded Areas: water bodies and land where urbanization cannot occur. – – This layer may contain binary data (0 and 99) or ranged values indicating probabilities of exclusion. Values: 0 -99. 0 = not excluded, 99 = excluded l Background: hillshaded image of region (used only with the graphic version of the model)

UGM Process Flow Data Set Preparation l Final data format must be as a GIF image. – Arc. Info: GRIDIMAGE -> TIF l l xv: TIF -> GIF Create Schedule Files urban. dates – roads. dates – landuse. classes – l Naming convention Contents of urban. dates 1930. urban 1950. urban 1970. urban 1990. urban Contents of landuse. classes 0 Unclass UNC 1 Urban URB 2 Agric 3 Range 4 Forest 5 Water EXC 6 Wetland 7 Barren

Thematic Data Input Exclusion Feature Hierarchy and Probability l The exclusion layer – Previously: Binary static possibility of growth occurring – The Latest: a range (0 - 100) l l Enables the exploration of zoning scenarios – e. g. ; green zones and urban corridors

Calibration

Calibration l Most essential element l Ensures realism l Ensures accountability and repeatability l Tests sensitivity l Required for complex systems models l Conducted in Monte Carlo mode

The Method l “Brute force calibration” l Phased exploration of parameter space l Start with coarse parameter steps and coarsened spatial data l Step to finer and finer data as calibration proceeds l Good rather than best solution l 5 parameters 0 -100 = 101^5 permutations

The Problem l. Model calibration for a medium sized data set and minimal data layers requires about 1200 CPU hours on a typical workstation l. CS calls problem tractability

Implementations to date l DEC Alpha l Silicon Graphics (Indy 10000 and O 2) l Silicon Graphics Origin 2000 cluster 32 processors: 2 GB RAM l Rolla, MS MCMC Beowulf Linux Cluster l Supercomputers (NESC EPA: NC) – Cray C-90 and T 3 D – Cray T 3 E-1200

past Calibration Predicting the present from the past For n Monte Carlo iterations For n coefficient sets “present”

UGM Process Flow UGM Compilation l Download Programs and Data (into a new directory) l contents of downloaded UGM. tar. gz – Clarke Urban Growth Model – Land Cover Deltatron Model – gd libraries – schedule files and calibrate file set to accept demo_city – demo_city data set

UGM Process Flow UGM Compilation l Set Up Model and Utilities – l Compile the gd libraries – l l enter: "make" to compile the model Type: "grow" – l by entering “make” in the GD subdirectory In the Model Directory – l gunzip and untar the UGM file this will begin the program The user will be prompted for what type of run, output and coeffecient values are desired. – These values are entered into the calibrate file. – test mode – animation vs calibrate Verify results – compare stats from demo_city with documented results

UGM Process Flow UGM Calibration l Phases of calibration – Coarse l l Iterations in large increments spanning coefficients’ full range images 1/4 full size – Fine l l Increments are smaller with a more focused coefficient range images 1/2 full size – Final l l The coefficient range should be narrowed to single increments images are full size

UGM Process Flow UGM Calibration Set constants and verify l Update working directory – move project data (your GIF image files) into the model directory – edit *. dates and landuse. classes to reflect your datasets l enter: "grow" – at the prompt: use ‘old’ calibrate file, ECHO 1, 2 number_of_times, test l Examine the numbers computed to standard output – Make sure they make sense for your data. – Values to examine are in the stats file, and are echoed by the program. l Use a viewing tool such as xv to examine the file cumulate. final. gif which should show a map of the result.

UGM Process Flow UGM Calibration Coarse Calibration Run Set constants "4 number_of_timesn” Run calibration enter: ”grow" Monitor results as they are written to the file control. stats until the script completes. Select "best" results

UGM Process Flow UGM Calibration Fine Calibration Run l Change calibrate to "bracket" results l CONTROLFILE "5 number_of_times” l Run calibration – enter: ”grow" l monitor results l Select "best" results

UGM Process Flow UGM Calibration Final Calibration Run l Change calibrate to "bracket" results – CONTROLFILE "10 number_of_times l Run calibration – enter: “grow" l monitor results l Select "BEST" results – these final parameters are used as input to model the data set's predicted growth

Version 3. 0 innovations l l l l Recoded into modular flow ANSI C Dynamic memory allocation returned to flat memory Optimized for Cray memory model Parallelized Built MPI (message passing interface) link Several code speed-ups and fixes Code tested and verified against Version 2. 1

Parallelization: Definition l Partitioning of the serial computation task into tasks which can be performed on different processors independently l Ideally, no necessary communication between processes. l Often quasi-independent, requiring communication or synchronization.

Issues for parallelization l Verifiable correctness l Concurrency l Locality vs communications l Scalability (across processors) l Load balancing l Deterministic (repeatable)

Types of Parallelism l Functional parallelism – Each processor performs different operations on own data l Domain decomposition – Data are partitioned between processors e. g. quad tree – Problems at edges, and with random numbers l Perfectly (embarrassingly) parallel – Processors work completely independently

Parallelization: Goal l Process larger data sets. l Process higher resolution. l Process more scenarios in less time. l Make model calibration possible! (=geocomputation)

SLEUTH Calibration l Brute force methodolgy l Partitions and explores parameter space l Scales across spatial resolutions l Works in phases with increasing parametric and spatial detail l Is embarrassingly parallel! l Massive speed-up attained

The Cost: MPI l MPI is a library of Fortran and C callable routines l Handles inter-process communication l Standard since 1993 l Queries environment for number of available processors l If processors=1, runs serially

Geocomputation See: Geocomputation: A Primer (Longley et al. 1999: J. Wiley) l Solution of close to untractable geographic problems using the power of the computer: recursion, speed, parallelization, automation, visualization. l Goal is to seek pattern and form unavailable by traditional means, e. g. stable state modeling, statistical analysis. l Ideal for complex systems and CA l

Cray T 3 E at NESC (Hickory) l Type: Massively Parallel Processor l CPU: 64 600 MHz Processor elements l Peak: 1. 2 Gflops/sec/PE l Memory: 256 MB/PE, 16 GB total l Topology: 3 -D Torus l OS: Unicos/mk

Model Outputs

UGM Process Flow UGM Products l Numeric – l The numerical output consists of goodness-of-fit calculations contained in the stats file. Graphic – single images l l – animations l l single run: a snapshot of a particular year Monte Carlo: a cumulative Monte Carlo image that results from multiple runs. These Monte Carlo images will show a probability of urbanization for a given year. The model can merge these images together to produce an animated gif of urban growth over time. Integration – – The images can also be introduced back into a GIS environment and used as data layers for further analysis in their spatial context. Arc. Info (for example) l l l Transform images into Arc acceptable format (e. g. : TIFF) Transform images into grids with Arc: GRIDIMAGE Georeference grids with Grid: CONTROLPOINTS

SLEUTH Outputs l Statistics l Logs l Images l Uncertainty l Animations maps

Land cover predictions and model calibration

Model Use for Forecasting

UGM Process Flow Prediction Run Predictions should not exceed amount of data known for the past e. g. 60 years of historical data sets can provide reasonable predictions up to 60 years beyond last data layer. 4. 1. Set control file 4. 1. 1. Use "BEST" performing results from calibration runs 4. 1. 1. 1 run the program "PARAM_AVG" usage: PARAM_AVG <last_parameter_filename> e. g. ; if the last date you used was 1990, enter"param_avg param 1990. log" PARAM_AVG will return the averaged, best final values of the five control parameters i. e. What the parameters were calibrated to, not what they started as. 4. 1. 1. 2 Place these values into calibrate update schedule files Change number_of_times should be very high (e. g. ; 100) ***** cumulate files? also - animate images every ? years. . . 4. 2 Remove old log files ( if any exist ) enter: "rm *. log" 4. 3 Update schedule files Unless you have any "future" layers ( e. g. roads in 2025 ) just include the present or last images' dates. 4. 5 Update excluded layer ( if desired ) Planned or proposed "non-urban" areas such as greenbelts or wetlands can be included in the excluded layer in order to explore the possible effects of landuse planning. Run model predictions enter: "grow" 4. 5 View Results 4. 5. 1 cumulate files 4. 5. 2 animated gifs 4. 5. 3 stats

Prediction (the future from the present) l Probability Images n Alternate Scenarios n Land Cover Uncertainty

2030 Growth predictions for Carpenteria, CA Urban growth boundary enforced Urban growth boundary 50% enforced Urban growth boundary not enforced

The Role of Transportation

Santa Barbara Study Area q Coastal California, 90 miles from Los Angeles q Intense physical constraints to new growth q SLEUTH is one of a set of models being used to envision Santa Barbara’s urban future. q A voluntary urban growth boundary is currently in place $1, 000 question. . . q How does the growth boundary affect urbanization?

Current Work NSF: Urban Change – Integrated Modeling Environment § Integration with UGROW and other models § Incorporate Level II land cover modeling § Fire or other natural hazard research UCSB/USGS § Formalization of SLEUTH § Simple GIS version of SLEUTH Urban Dynamics Workgroup § Beta test and release of v 3. 0 § Test for Monte Carlo Significance § Test for Scalability of Model § Compare SLEUTH with other Urban/Land Cover models

INITIAL WEIGHT 50 Clarke Urban Growth Model Road Weight Algorithm AADT Average Annual Daily Traffic (AADT) <5000 -10000 -15000 -20000 >20000 -5 +10 -10 -20 Local Functional Class Location Collector Arterial Interstate +20 +15 -10 -20 Urban +5 Rural -5 Distance to Interstate Junctions (Non Interstate) -----------0 -1 mile +10 1 -2 miles +5 2 -3 miles -5 >3 miles -10 Distance to Interstate ROAD WEIGHT 0 -1 Mile +10 1 -2 miles +5 2 -3 miles -5 >3 miles…-10 Toby N. Carlson, Dept. of Meteorology; John T. Marker, Kostas Goulias, Pennsylvania Transportation Institute Penn State University