Designing a traffic monitoring program using landuse change detection

Designing a traffic monitoring program using landuse change detection

SHARE

Dr. Reginald Souleyrette
Center for Transportation Research and Education, Iowa State University
ISU Research Park, 2901 S. Loop Drive, Suite 3100, Ames, Iowa 50010-8632
Tel: +1(515)-294-5453, +1(515)-294-7188, Fax: +1(515)-294-0467, Email: [email protected]

Sitansu Pattnaik
Sitansu B Pattnaik
Center for Transportation Research and Education, Iowa State University
ISU Research Park, 2901 S. Loop Drive, Suite 3100, Ames, Iowa 50010-8632
Tel: +1(515)-294-5453, +1(515)-294-7188, Fax: +1(515)-294-0467, Email: [email protected]

Introduction
Travel patterns of a transportation system depend on driver characteristics, local/regional development and the available transportation network. These patterns are studied through the implementation of traffic monitoring programs, which measure traffic volume and characteristics. In the U.S., Traffic counts are generally classified based vehicle type and gross weight. Accurate counts are very important to transportation planning activities. Garder, 1999, says “with the movement towards design-build highway projects and warranties on performance, accurate measurement of vehicular movement is required to ascertain if the roadway has met or exceeded the design requirements”. [1]

The Federal Highway Administration (FHWA) mandates the states to collect traffic count information at specified intervals to meet the needs of the Highway Performance Monitoring System (HPMS). Each state may have its own method of implementation while satisfying the minimum federal requirements. In the state of Iowa, approximately 10,000 mechanical counts and 1,000 manual counts are collected every year.

The scope of traffic monitoring as discussed in this paper deals with the procedures for appropriate sample selection for efficient and timely monitoring. This paper proposes using multi-temporal spatial datasets as input into change detection procedures for improving design and efficiency of traffic monitoring programs. Results could enable redirecting traffic count activities, and related data management resources, to areas that are experiencing the greatest changes in land use and related traffic volume. Conversely, areas where traffic counts are static or changes are statistically insignificant over time could be counted less frequently, while other, less costly methods might be employed to generate the data needed for these locations. Due to recent study and data availability, the city of Maquoketa, located in eastern Iowa was chosen for development of the change detection procedure.

Traffic Monitoring
Traffic monitoring is undertaken to collect the volume, gross weight and classification of the vehicles in the road network. The data collected is utilized in different fields as shown in table 1.

In the absence of budgetary constraints, each road segment could be continuously monitored to determine the values of the AADT, vehicle mix by type and gross weight. “However, in practice, a few road segments are monitored continuously every day of the year to produce annual characteristics of traffic flow.”[2]

Table 1: Examples of Studies That Use Traffic Characteristics Data (TMG, May 2001)

There are two types of traffic monitoring schemes; portable short duration counts and permanent continuous counts. The first step in traffic monitoring is to select the number and location of the count locations. The number of sites is decided through statistical theory for achieving the desired precision and count station locations are based on the transportation network characteristics.

After sampling, the next step is factor generation for the traffic counts. These adjustment factors are needed to extrapolate short duration traffic counts into estimates of AADT. And they are also required for the remaining roads to determine the AADT based on permanent and short duration traffic counts at the sample locations. These factors are necessary to generate representative values for each day of the week, month and for each road type. Validation is the last step in the data analysis process wherein the results after using the adjustment factors are compared with control data.

FHWA publishes the Traffic Monitoring Guide (TMG), which is the backbone of all the traffic monitoring initiatives in different states. The guide is a set of recommendations for implementation of portable short duration counts and permanent continuous counts. The TMG also provides specific recommendations on the number, extent, and duration of monitoring efforts.

Iowa DOT’s procedure
The state of Iowa has approximately 130 permanent count locations spread across the state for traffic data collection and a manual count schedule to satisfy the federal mandates.


Figure 1: Map of Iowa showing the four counting zones and the level of data collection

The state is divided into four zones as shown in figure 1. Each year a quarter of the state is selected for collection of traffic counts. The counties, which are hatched, are chosen in a cycle for complete counts, which includes the secondary roads and the counts for only the primary roads are collected in the remaining counties in the zone. Thus, for primary (principal) roads, a data collection cycle is completed once every four years (one quarter of the state is counted every year). Secondary (non-principal) roads are counted only every eight years. Growth and changes in landuse as well as infrastructure development affect traffic patterns, and clearly, some areas grow more than others. To respond to this, and to make the procedure more efficient and provide the timeliest data, the present process allows out-of turn counting. The determination of change and prioritization is based on recommendations from state, county, and city officials. However, this procedure is highly subjective and only “significant” changes in landuse and network are considered.

The studies performed every year are conducted using mechanical and manual traffic counts. The mechanical counts are performed using portable automatic traffic controllers and are setup for a period of 24 to 48 hours. The manual counts are input directly into a microprocessor and are usually conducted in two time periods of four hours each or three consecutive eight hour slabs. “The traffic monitoring program of Iowa Department of Transportation is approved by the FHWA and satisfies or exceeds the TMG recommendations regarding sample size and stratification to calculate traffic within 10% with 95% confidence.”[3]

Changes at the Iowa DOT
As described in the previous section, all the roads in the state were being monitored for determining the traffic patterns, at least once every eight-years. Prior to 2000, Iowa law mandated a Quadrennial Needs Study (QNS) that required the Iowa DOT to conduct extensive field inventory of the state’s 110,000 miles of road. During this survey process, DOT field personnel could easily observe locations of development and land use change. This information formed an important part of the data needed to locate an efficient count program. The 2002 legislative session resulted in the termination of the QNS, in favor of a regional fund allocation process. This eliminated the need for DOT field survey of local roads, and has resulted in the need for a new method to direct efficient counting of traffic.

The opportunity for GIS and remote sensing
Multi temporal dataset archives of terrestrial areas were historically maintained by USGS after the launch of Landsat series of earth observing satellites. The Landsat datasets have limited utility in landuse analysis due to low spatial resolution. However, with the availability of multi-temporal datasets in a standardized format they can be very useful in understanding changes at a regional level.

A more common form of remotely sensed data is the aerial photograph, which can be acquired at the project, city, county or state level. Until recently, acquisition of aerial photos requires lower tasking time and acquisition cost than satellite imagery, Also, the spatial resolution of aerial photographs was better than available satellite imagery and they could be collected and stored by public and even private agencies for use with various transportation and environmental projects. Improvement in storage technology also made the dia-positives of the aerial photographs available in electronic formats, which aided data access and storage. Data are available at many spatial and spectral resolutions. Many state, city and county agencies have archives of remote sensing data, which can be utilized for improving planning processes without the need for significant further investment.

Change detection
Change detection for determining the change in the landuse is based on identifying the regions, which have undergone growth and development during the analysis period. This is done by determining the difference between the landuse classes in the datasets acquired at different epochs for a study area. The procedure allows an analyst to identify the areas where the landuse has changed and the manner in which it has changed.

Methods of change detection:

  • Image differencing:
  • This procedure involves pixel level operation wherein the imagery from one epoch is subtracted from that of another. The changes in the radiance values are grouped to detect the areas with appreciable change. The determination of the threshold value to determine the areas of appreciable change based on the change in radiance values is important in getting the results.

  • Post classification comparison:
  • The imagery from each dataset is independently classified by using supervised classification and then the landuse types in the earlier epoch are compared with the later epoch to determine the direction of change. This method is more descriptive as the type of landuse change can also be determined.

Data availability
The available remotely sensed datasets could be categorized with respect to their aerial extent, spatial and spectral resolution. Panchromatic aerial photos available with USGS and the yearly road vector database prepared by DOT were used in this change detection study.

Table 2: Data typology for Iowa

The older imagery from 1992 had a spatial resolution of 1 meter and the recent imagery from 2002 had a spatial resolution of 0.3 meter. Table 2 shows the data typology for the state of Iowa. In table 2 “X” marks the datasets used for change detection the city of Maquoketa.

Implementation of a conventional change detection procedure:

An image differencing method for change detection was applied to identify the changes in the city of Maquoketa during the analysis period. These preliminary results were not very useful, as the aerial photo from the two epochs did not match accurately. To address this issue, a post classification comparison procedure was applied. This was done by first classifying the aerial photographs from two different time periods into five landuse classes: thick vegetation, cultivated land, dense residential, sparse residential and water bodies. However, the results of classification were again disappointing.

Image differencing was then applied to find the regions with landuse change. The end result continued to lack utility due to the low spectral resolution (8 bit panchromatic) of the available imagery. This resulted in the selection of impure training pixels for supervised classification. While high spectral resolution imagery could have improved the results of supervised classification, it is likely that there would be problems using the data for automated change detection. For example, color infrared imagery will soon be available for the study area. However, older imagery would still have low spatial and spectral resolution and there would continue to be problems with image-to-image registration. Finally, the analyst has to manually identify the change in landuse in the corresponding images after comparing the regions, making the process difficult when large areas are considered. Therefore, we propose the procedure for use with available datasets of low spatial and spectral information in the following section.

Recommended procedure for change detection
The proposed process specifically deals with datasets of low spectral resolution available for large areas. Available vector datasets describing the road network in different years are used along with the raster imagery to derive changes in road network and landuse.

Step 1 Identifying area for analysis: –
Most development occurs within a certain distance of roads. At the least, new or larger driveways will be created for access to the development. In some cases, new roads are built (see step 2). To reduce computation requirements, we can take advantage of the proximity of development to existing roads, and reduce the areal extent of data required. Prior to change detection, we can buffer a vector roads database and clip the aerial photos. The size of the buffer depends on the registration of the two images and the vector database. (In the Maquoketa databases, we estimated the registration of the three databases to generally be within 10 meters.) We suggest that the buffer size be set to include the maximum registration error plus 30 meters, as most development will show significant differences within 30 meters of the nearest road. By ignoring change outside the buffered area, the computations are much faster and an analyst conducting manual identification of change can proceed through the database at a much faster rate.


Figure 2: Results of Change Detection. [Blue (new), Red (unchanged)]

Step 2 Change detection: –
Aerial photographs from 2002 and 1992 were used as the input for identifying landuse change. The imagery is first manually classified to identify built-up areas, by assigning breakpoints derived from the spectral grouping observed in their respective histograms. Image differencing is then performed for post-classification comparison of these classified images. The results are shown in figure 2.The regions in blue are locations of potential new development and the red pixels denote built up areas, which are unchanged in the analysis time frame.

Step 3 Identifying new roads: –
In steps 1 and 2, areas indicating the potential for new development were identified. To assist the analyst in determining areas of actual and significant development, GIS and remote sensing can help to identify the spatial location of new road segments. New roads proximate to areas identified as potentially built -up and changed, would indicate a high likelihood of development of interest to traffic planners. The process for identifying new road segments could range from simple attribute queries in a roads database, to spatial subtracting of two vector road databases from different dates, to complex centerline extraction from aerial or satellite imagery. Of course, if a recent roads database and an older roads database include unique attribute identifiers, one can simply query out the new roads. In Maquoketa, two vector road databases were available. These were compared spatially to find differences. Because the two databases did not overlay, the older database was rubber-sheeted to the new one before differencing. Resulting “existing roads” were then buffered to create “existing road areas” (see figure 3). The “new roads” were identified after subtracting the buffer of existing roads from the latest road database.


Figure 3: The newer roads built in the analysis period (RED) * includes errors due to large mismatch in the road networks

Step 4 Identifying areas of development: –
Potential sites with landuse change were identified as shown in figure 2.These sites are potential regions to focus traffic monitoring efforts. Potential developing areas that are co-located with new roads are especially promising and deserving of more focused attention. The regions with verified landuse changed can be flagged and then the traffic count locations can be established. Recent aerial photos of the area can then be efficiently examined manually by focusing only on these areas classified as potentially developing.

Step 5 Traffic count location: –
Figure 4 shows a 3-dimensional visualization that may be useful in identifying traffic monitoring stations by change detection.


Figure 4: 3-D visualization of landuse change

The 3-D scene can be easily rotated to obtain a better viewing perspective and can be modified as required for improved visual interpretation.

Limitations and recommendations
This methodology may miss large (or small) development that does not occur within the suggested buffer distance. Most small development occurs in close proximity to roads. However, some larger developments (with high traffic generation) occur on large tracts of land and may be well set-back from the adjoining highway. This development may be connected to the highway by small, two-lane, private service roads. High traffic generators that may be set-back from the highway may include factories, hotels or office parks, which will be large in size and mode therefore detectable on lower resolution imagery using either change detection or manual techniques. We suggest that the change detection process be repeated for the areas that were originally omitted (outside the buffers) using resampled imagery (larger pixel size). While the area to be examined would be much larger, smaller viewing scales could be used to move through large areas of imagery in an efficient manner without missing the large sized developments. We suggest that an adequate viewing scale for identifying these facilities would be 1:24,000 (1:12,000 is better), with aerial photos resampled to 10-meter pixel (5 meter is better.)

Another limitation of the proposed methodology is that traffic count location cannot solely be determined based on landuse change. An understanding of traffic flow patterns in a region and knowledge of existing and historical traffic count locations and procedures must be coupled with the new information on land use change to design an effective counting program.

Conclusions and Recommendations
It is interesting to note that much of the land excluded from change detection analysis is agricultural in nature and difficult to automatically process with 8-bit imagery, due to variation in reflectance values. This was due to the different crops and the state of the field in the cultivation cycle. Because only 8-bit imagery was available, these changes could not be successfully classified into a unique landuse class. However, using multi-spectral imagery, it may be possible to automatically process these areas for change, reducing or even eliminating the need for preprocessing the data with proximity buffers.

This study was conducted using only a very small portion of the State’s land area (less than 100 square kilometers out of more than 150,000). For this procedure to be feasible on a large scale, simple GIS tools could be designed to expedite the human machine interface. These may include tools for flagging areas already visited that do not have development, flagging areas for further study, and flagging areas of definite development. Display of traffic count locations would also be advantageous. In the state of Iowa, the total number of counts in an eight-year period (the cycle for counting all the roads) exceeds 100,000 (nearly one count per mile of the 110,000 miles of roads in Iowa). A city the size of Maquoketa could have 50 count locations to be defined.

Inclusion of change detection in collecting data for modeling the traffic patterns would allow the location of the 50 count stations to be focused in the regions of significant change. Moreover, the total number of count locations can also be reduced to optimize the resources while collecting enough data for traffic modeling.

Recently, Color Infra red (CIR) imagery has been flown for the state of Iowa. Once these data become available, the increased spectral resolution will likely facilitate accurate segmentation of the image into landuse classes, improving the change detection process.

The objective of this research was to utilize available resources for better traffic monitoring in an era of decreasing resources. Ideally, for implementing a change detection study, an analyst would have multi-spectral imagery of high spatial resolution and accurate vector datasets. However, it is recognized that remotely sensed data are available in varied resolutions and formats. Making use of spectrally sparse datasets requires a creative pre-processing phase to be suitable for analysis.

References:

  • Gardner, M.P., Highway traffic monitoring, Committee on highway traffic monitoring; TRB, 1999.
  • Iowa DOT, Iowa’s monitoring program, Systems monitoring section, 2002.
  • FHWA, Variability in traffic monitoring data, Final Summary Report, 1997.
  • FHWA, Traffic monitoring Guide, May 2001.