TOWARDS A NEURAL NETWORK BASED MODEL FOR
TRAVEL DEMAND FORECAST: GIS AND REMOTE SENSING
APPROACH
André Dantas1, Marcus V. Lamar1, Yaeko Yamashita2, Koshi Yamamoto1, Eizo
Hideshima1
1
2
Nagoya Institute of Technology, Nagoya, Japan
Department of Civil Engineering, University of Brasilia - Brazil
E-mail: andre@keik1.ace.nitech.ac.jp
A formulation of travel forecast model based on geographical–spatial
data of urban areas is presented. It focus on land use – transportation
system interaction in order to represent displacements (trips) needs of
human activities within the urban area. The geographical–spatial data
mainly originated from satellite and aerial photograph images, are
obtained exploring Geographical Information Systems and Remote
Sensing. Neural Networks are used to establish an “artificial intelligent”
comprehension of the urban dynamic that affect travel demand process.
A case study in Boston Metropolitan Area proved the methodology’s
efficiency, given that the consideration of geographical spatial variables
resulted in a good result of trip forecast.
1. INTRODUCTION
In the process of urban transportation planning, travel demand forecast is one of the
most important analysis instruments to evaluate the capacity to attend the present and
future needs. These forecasts integrate into a planning process aiming to generate the
necessary information to decide where, when and how transportation system has to be
improved. They intend to measure how urban activities are converted in travels, by
assuming that those interactions generate the movement of people. In this sense, the
number of trips from origin to destination zones are estimated by correlating
socioeconomic data (population, income, etc) [Khisty, 1990].
Forecast models have failed in providing an efficient representation of spatial and
temporal interaction between the land use and transportation system. Models were
developed processing and correlating representative variables using traditional statistic
techniques without considering geographical – spatial reality. They do not incorporate
the complex and uncertain nature of land use and transportation system interaction
generating static and deterministic models that require too much data to be applicable in
real situations [Harris, 1991] [Rodrigue, 1997].
To overcome these limitations, this paper proposes a Remote Sensing (RS) and
Geographical Information Systems (GIS) approach to develop a model based on Neural
Network (NN). This approach intends to explore RS data as the main source to generate
information related to land use patterns and transportation system interactions using
GIS. This information is processed by NN that will establish an artificial intelligent
modelling considering the interactions in a spatial-temporal perspective. In the
sequence, it is established an RS-GIS-NN integration that provides the incorporation
and analysis of the urban dynamics in the model. Moreover, according to Morain and
Baros (1996), this integration data leads to time reduction and cost-saving benefits
against field surveys that are hardly updated by using traditional methods. This fact is
essential for the urban transportation planning process since the proposed approach
intends to reduce and/or eliminate extensive data collection as usually processed.
In this paper, we concentrate on the description of this integration, the model’s
formulation for spatial dimension (without temporal variations) and the case study for
Boston Metropolitan Area.
2. TRAVEL FORECAST MODEL: RS&GIS& INTEGRATION APPROACH
RS and GIS integration has been successfully applied to solve urban problems
generating information for decision-making activities. Many applications have being
conducted exploring this integration IN the recent years (Morain and Baros, 1996).
However, Foresman and Millete (1997) emphasize that GIS capabilities have been
concentrated in Boolean operations and conventional spatial statistics when processing
multilayered map-based information that has lead to a basic level of “display analysis”.
In order to reach more powerful spatial analytic and spatial process modelling tools NN
has been applied. Openshaw (1991) and Openshaw et al (1995) reported successful
applications and NN efficiency, showing perspectives to elaborate a new modelling
conception. In this direction, it is introduced here the integration framework specifically
to the travel forecast modelling. It can be applied to process data from different urban
areas, which would be the ideal situation providing more generalization for the model.
Its description is concentrated on the integration process and the role of its components.
Five modules as shown in Figure 1 compose the framework. The integration process
starts with data obtainment. In the sequence, satellite images and aerial photographs are
stored in RS data Module (1) and Origin / destination matrix are elaborated in Trips data
Module (2) and finally maps containing transportation system information (road,
subway, train, bus, etc) are transferred to Maps data Module (3). In this obtainment
process, it is fundamental to define a common aggregation level once this definition will
direct affect all modelling process.
After this step, data from Module 1 is used in GIS Module (4) in order to process
multispectral analysis and aerial photograph interpretation generating the land use
patterns. After, they are incorporated to GIS database that previously stored data from
Module 2 and 3. Spatial-temporal queries are processed using this database and the
results are conducted to NN Module (5). In this module, query results are pre-processed
and then divided in training and test data sets. NN software performs the training
process defining a function that is tested and revised until the expected level of forecast
is reached. Finally, Module 5 returns to Module 4 (GIS), where travel forecast function
is applied to a future scenario and then visualized.
1
RS data
2
t=0
Trips data
Maps data
3
t=0
t=0
t =1
t=1
t=1
t=n
t=n
t=n
4
CLASS
Classe 3 3
Multispectral Analysis &
Aerial Photograph interpretation
Land use Patterns
GIS database
Visualization & Analysis
GIS


f X t n , Y t n ,.....,Z t n  Trips t n
Travel Forecast
Spatial-temporal Queries
f  X , Y ,....., Z   Trips
Test data
set
Pre–Processing
Data Set
NN
Forecast Function
5
Training
Process
Training data
set
Figure 1: RS and GIS and NN integration for Travel Forecast Problem
3. NN DEFINITION FOR MODELLING IN SPATIAL DIMENSION
An urban area is divided in Homogenous Aggregated Sector (HAS) as defined by Taco
et al [1999]. Trips (Tij) between a pair of HAS pairs are verified, where i and j denote the
origin and destination, respectively. Land use is represented by occupied area of each
pattern in each HAS, which is assigned to RLUHAS (Residential Land use), CLUHAS
(Commercial Land use) and SLUHAS (Service Land use). Transportation system is
characterized by the extension of each mode in each HAS, which is associated to RTSHAS
(Road Transportation System), BTSHAS (Bus Transportation System), STSHAS (Subway
Transportation System) and TTSHAS (Train Transportation System). In addition, it is
defined the distance Dij between a pair of HAS that intends to represent location
relations.
The model aims to forecast the levels of urban movements as main element to be used
in strategic planning activities. It is expected that the model’s final output is an intensity
indicator classifying the forecasted movements in categories such as high, mediumhigh, medium, medium-low and low. In this way, Tij is classified z categories according
to range variation (v) between them and the maximum and minimum trips observed in
the urban area (or study area).
In the NN modelling, these characteristics are transformed in input (X) and output (Y)
vectors that are processed as a feedforward Multilayer Perceptron (MLP) (see
Wasserman, 1989). Here, the input vector is defined such as X = (RLUi, CLUi, SLUi,
RTSi, BTSi, STSi, TTSi, RLUj, CLUj, SLUj, RTSj, BTSj, STSj, TTSj, Dij). The Y vector is
obtained from the classification of generating (y1, y2, ….yz) where y values are defined
by equation 1.
1.0 ,
k  Ti j  k
where k  (1, 2, …, z)
yk =
0.0 ,
(1)
otherwise
where k and k are the minimum and maximum value of category k, respectively.
Physically, it means that trips between two HAS (i and j) are expressed in NN by the
combination of land use, transportation systems and distances indicators for each HAS
(origin and destination).
As a typical NN, input, hidden and output layers compose it. Hidden layers (H) are
composed by are hidden neuron units H1, H2, … Hm connect input and output, expressed
by X and Y, respectively, as shown by Figure 2.
Hidden Layer
Input Layer
x1
W11
H1
V11
y1
y2
H2
x2
W1m
xn
Output Layer
Wnm
Hm
Vmz
Figure 2: Typical NN structure
V1z
Yz
The output of the neurons are calculated by:
 n

H r  f   Wrs X s 
 s 1

(2)
m

y g  f   Vu H ug 
 u 1

(3)
where f is the activation function assumed in this study a sigmoid function. W and V are
weight matrixes that are obtained by applying a backpropagation based training
algorithm [Wasserman, 1989].
4. CASE STUDY
This case study intended to evaluate the integration process and the proposed NN
model. Tests were conducted in Boston Metropolitan Area (Massachusetts State –
USA), which covers about 1400 square miles (3580 square kilometers) with nearly three
million people live. It was selected a study area in Boston South area, near to Boston
Medical Center that involves seventeen Traffic Zones (TZ) as shown in Figure 2. It was
assumed that the TZ`s were equivalent to HAS definition (see item 3). The study area is
mainly occupied by residential land use and it is located very near to downtown.
TZ limits
Figure 3: Study area – selected TZ
RS and Transportation system data were obtained from MassGIS database
(http://www.magnet.state.ma.us/mgis/). In this database, data was projected to
Massachusetts State plane Mainland Zone (FIPSZONE 2001) coordinate system, Datum
NAD83, units meters. It was used black and white digital orthophotos produced in 1992
in 1:5000 scale. Additionally, bus route maps from Metropolitan Boston Transportation
Authority (MBTA) were incorporated to Transportation data. Finally, Central
Transportation Planning Staff (CTPS) provided access to Travel data related to 1990
survey that involves all purposes trips as well as HAS definition.
In the sequence, GIS database construction activities are described. Next, NN
experiments and results are introduced.
4.1 GIS database
Geoconcept GIS software was used to develop these tests. First, Land Use patterns were
obtained following United States Geological Service (USGS) classification system
[Avery and Berlin, 1990] and Taco’s methodology [Taco et al, 1999]. It was analyzed
up to Level II (Residential, Commercial and Services) according to model’s
requirements (see item 3). Figure 3 presents the Land Use patterns for the study area. In
the sequence, Transportation System was digitalized in the GIS database. Figure 4
shows Road, Bus, Subway and Train systems for study area.
4.2 NN experiments and results
Once GIS database was elaborated, the X and Y vectors were established from this
database using spatial queries to conduct NN experiments. The input vector X was
normalized using equation 4.


xncTZ  1  2 xcTZ  xcmin / xcmax  xcmin

(4)
where xcTZ is the value from X vector for characteristic c (land use or transportation
system or distance) in TZ and xncZT is the normalized value. xcmin and xcmax are the
minimum and maximum value for the related vector.
In the sequence, Y vectors were generated through the classification of Tij matrix
according to equation 1. Considering that the maximum and minimum trips were 67 and
0, respectively. Then, z (categories) and v (range) were defined 5 and 14, respectively.
In order to create training and test data, X and Y vectors were divided following a
random distribution. Test data were selected according to a standard criterion, which
establishes that it has to be 25% of the total. So, 216 and 73 vectors were assigned to
training and test sets, respectively.
Five NN structure types were simulated in T-learn software. From the basic structure
described in item 3, it was established variations in hidden layer. Structures A, B and C
were defined with 15, 30 and 8 nodes in hidden layer, respectively. In structure D and
E, it was inserted one and two hidden layers, receptively. The NN were trained until a
minimum MSE error was reached in the test data through the application of equation 5.
1 p   
MSE    Tw  Yw
p w1




2
(5)
where MSE is the mean square error considering forecasted (Y) and expected (T)
classified trips for p testing data vectors.
Residential
Services
Commercial
Figure 3: Land Use patterns
Road
Bus
Train
Subway
Figure 4: Transportation System
The Table 1 summarizes the results for these five structures. It can be observed that
MSE variation was very small indicating that the NN structure did not affect the final
results. It can be verified that structure E presented the worst result comparing to the
others that are on the same level considering the number of interactions to reach the
minimum. However, this difference does not compromise the time execution since just
few seconds are necessary to perform the complete optimization process. Finally, the
recognition rates show that all the structures reached the same level of optimization and
indicating that the network structure is not a fundamental parameter to obtain a good
modelling for this formulation.
Analyzing in detail the forecasted results, it is verified that the errors occurred in five
testing vectors. These vectors were related to categories 2, 3 and 4. On the other hand,
all the test vectors related to category 5 (0  Tij  14) were correctly forecasted. This NN
performance can be explained by the composition of X and Y vectors that were obtained
from study area. It is observed that there is a great concentration of data on category 5
(252 vectors – 87% of the total). Meanwhile, categories 1, 2, 3 and 4 have only 1, 3, 5
and 28 vectors, respectively. This composition led to the construction of training and
testing data sets that resulted in NN efficiency just for those vectors in category 5.
Table 1 –Simulation results
Interactions
Training
Medium Error
Test
Recognition
Rate (%)
MSE
A
52920
B
55944
Structures
C
52920
D
64584
E
138456
0.099626
0.099721
0.099698
0.099692
0.099267
94.5
94.5
94.5
94.5
945
0.150688
0.15633
0.15993
0.157894
0.155705
Despite the high recognition rate, the final results were compromised by data
composition. In this way, it is suggested that a special treatment should be done to
provide a better NN modelling. In this sense, it is proposed the creation of a “balanced”
data set generating training and test vectors smoothly distributed. It was necessary to
“create” data for those categories with few vectors (1,2,3,4) and “eliminate” some from
category 5.
It was defined that 40 and 10 vectors of each category will compose the training and
testing data, respectively. First, it was “created” data for those vectors related to
categories 1, 2, 3 and 4 by adding a gaussian noise with average zero and variance
0.001 to those existing vectors, generating new “noisy” version of the original ones.
This operation was performed in order to generate to supply the NN with vectors highly
correlated with original ones. Next, it was reduced the number of vectors from category
5 applying the LBG algorithm [Linde et al., 1980]. The LBG aims to design quasioptimum codebooks in vector quantization based on coding systems. It was applied to
create 50 template vectors of category 5, which minimize the MSE of the representation
of all vectors of category 5 by that 50 templates.
Using the “balanced” data set, X and Y vectors were defined for training and testing.
Again, it was simulated the structures A, B, C, D and E as previously defined. In Table
2, it is verified that structure E reached the best result for testing data generating almost
the same recognition rate obtained before.
Table 2 – Simulation results for “balanced” data
Interactions
Training
Medium Error
Test
Recognition
Rate (%)
MSE
A
79600
B
1031000
Structures
C
4554800
0.024952
0.000195
0.015074
0.024986
0.024920
90.0
90.0
88.0
90.0
94.0
0.191919
0.189542
0.194383
0.180819
0.101601
D
96400
E
144400
As it was expected, using “balanced” data the recognition rates are worse than the
previous experiment. However, the “balanced” data provided a better modelling for
those categories that have only few samples in training set, improving the modelling
capability of the NN. In all categories, the NN was able to forecast correctly the most
part of testing vectors. For instance, in structure E only three vectors presented an
incorrect forecast. In these vectors, it is noticed that one or more characteristics of the
input vector (X) is zero. Specifically, one of the errors occurred for trips from TZ 48 to
58, which do not contain train and subway transportation systems. Similarly, it is
observed the same behavior for TZ 56. Physically, these errors express how difficult it is
to the NN to reach the expected result without the information about the transportation
system.
5. CONCLUSIONS
This paper presented a model that provides the representation of urban dynamic in
travel forecast process, which the traditional modelling did not contemplate. We can not
criticize the traditional modelling, since it was developed in the generation where the
tools for simulation were related to that time. In this way, the traditional modelling has
to be upgraded to the new era where new techniques and technology such as GIS, RS,
NN become to fulfill the deficiency of traditional modelling.
In order to establish the representation of the urban dynamic, it was explored the
integration of RS, GIS and NN that was described and discussed. This integration was
essential to concept a framework that combines raster data from RS, GIS capabilities
and a non-linear mathematical formulation (NN) to express complex interactions
between the land use and transportation systems.
After the definition of NN formulation for modelling in spatial dimension, the case
study for Boston Metropolitan Area proved model’s efficiency. The proposed
formulation was capable to process the travel forecast modelling reaching a
considerable recognition rate. It was applied a special treatment on the obtained data in
order to generate a better generalization for NN modelling. Such treatment proved to be
efficient to forecast trips in all the categories. The observed errors were related to cases
where there was not sufficient information for NN establish the suitable forecast. One
possible direction to solve this limitation is related to the aggregation level of the data.
In these experiments, TZ definition was used considering that there was not any other
suitable definition of HAS. In Boston case, TZ definition was designed for microanalysis
generating units with a very limited area and specific characteristics of land use and
transportation system. It is expected that using a macro aggregation level these kind of
errors would be eliminated.
Acknowledgments: The authors would like to express gratitude by the scholarship
supported from Japanese Ministry of Education (Monbushou) for the development of
this research. We also would like to thank Central Transportation Planning Staff (CTPS)
of Boston for all support and travel data used in this research.
REFERENCES
Avery, T. E., Berlin, G. L. Fundamental of Remote Sensing and Airphoto Intrepretation. Maxwell
Macmillan International, New York, USA [1990].
Fischer, M.M. From Conventional to Knowledge-based Geographical Information Systems. Computer,
Environmental and Urban Systems. Vol. 18, no. 4, pp-233-242. USA [1994].
Fisher, M.M Spatial analysis: retrospect and prospect. In Geographical Information Systems: Principles
and Technical issues. Eds Longely, P.A., Goodchild, M.F, Maguirre, D.J., Rhind, D.W. v. 1, Second
Edition, USA [1999].
Foresman , T.W., Millete, T.L.; Integration of remote sensing and GIS technologies for planning. In
Integration of Geographical Information Systems and Remote Sensing [1997].
Harris, B.; Land use Models in Transportation Planning: a review of past developments and current best
practice;
Delawae
Valley
Regional
Planning
Commision.
(http://www.bts.gov/tmip/papers/landuse/compendium/dvrpe-appb.htm) [1996].
Khisty, C. J.; Transportation Engineering: an introduction; pp. 388; Prentice-Hall [1990].
Linde, Y., Buzo, A, Gray, R.M., “An Algorithm for vector Quantizers design”, IEEE Trans. On
Communications, v.com-28, pp.84-95, January. [1980].
Morain, S.; Baros, S. L., Raster imagery in Geographical Information systems. On Word Press, p. 495;
USA[1996] .
Openshaw, S.; A concepts-rich approach to spatial analysis, theory generation, and scientific discorvery
in GIS using massively parallel computing. In Innovations in GIS 1 [1991]
Openshaw S; Blake, M; Wymer, C ; Using neurocomputing methods to classify Britain’s residential areas. In
Innovations in GIS 2 [1995]
Rodrigue, J. P. Parallel modelling and Neural Networks: an overview for transportation / land use
systems. Transportation Research C, v. 5, n. 5, p. 259-271 [1997].
Taco, P.W., Yamashita, Y, Moreira, Souza, N, Dantas, A. Trip Generation Model: A New Conception
Using Remote Sensing and Geographic Information Systems. Photogrammetrie Fernerkundung
Geoinformation, Germany (to be published) [1999].
Wasserman, P. D. (1989) Neural Computing: theory and practice, p. 210; Van Nostrand Reinhold, New
York, USA [1989].