CICEET Progress Report for the period 9/15/04 Through 2/28/05

Project Title: Integrating Technologies to Monitor and Predict Patterns of Urban Growth
Principal Investigator(s): Fay Rubin, William Salas

Figures

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Tables

Table 1

Table 2

Table 3

1. Complete project CD.
2. Complete dissemination of maps and project CD.
3. Build and apply neural network model.
4. Use neural network model to examine development pressures at the watershed scale under a range of development scenarios.
5. Apply % imperviousness algorithms to examine current (and future projected) watershed conditions.
6. Continue project outreach activities.
7. Prepare final report.

Progress on Tasks
Task 1: Complete project CD.
The last status report indicated that the project CD was being assembled, with distribution anticipated shortly. However, the project team subsequently decided to incorporate another component on the CD, presenting preliminary project results in the context of a brief overview document. This document is presently being finalized.

Task 2: Complete dissemination of maps and project CD.
All large-scale town-level maps, with associated countywide maps, have been disseminated to the towns in the study area. Although not requested, we have received a number of responses from the map recipients suggesting that the products will be very valuable in future planning initiatives.

As noted above, the distribution of the CD was delayed due to our decision to include another document on the media. We expect to distribute the final CD during the month of April.

Tasks 3 & 4: Build and apply neural network model & Assess Development Pressures An artificial neural network (ANN) is a functional mapping of d inputs into an output, c. ANN consists of hidden layers and nodes. A node is a linear transformation of inputs followed by an application of a non-linear function. The output from a layer is a vector consisting of the outputs from the non-linear function for each node.

We have created and evaluated several neural network models during the past project period. The success of the model is often dependent on presenting a representative training data set and selecting the optimum number of nodes for the hidden layers generated by the model network. We have been using a stratified, random sampling approach for collecting training data from our inputs. For the most recent model, we have used 40% and 10% sample for the new built and no change training classes, respectively. The remaining data were used for model testing. We selected different sampling percentages to avoid overfitting the model and to keep the number of training samples per class comparable. We selected a final model that used ten (d=10) spatial data layers as inputs into the model and our observed urban development laver is a binary variable indicating either no-change (c = 0) or urban expansion (c =1). We elected to use a standard backpropagation training algorithm that iterates the weights of each node until the estimation error converges below a prescribed threshold. However, to avoid overfitting, we used the technique of Bayesian regularization to limit the complexity of the model. Our final model used only 2 nodes. A list of all our final inputs to the neural network model is provided in Table 1. Accessibility and distance to other developed areas were the most important factors in the model (for example, at a scale of 2 mile grids, accessibility explained over 33% of the observed variance in development patterns). While the target values for the model are binary, either 1s (converted to urban) or 0s (not developed), the outputs of our model are continuous and represent a measure of similarity of the area (pixel) in question to those areas used to train the network. We refer to this output as our similarity index (SI). For example, areas with similar characteristics to areas that were converted to urban should have values closer to 1, and areas similar to areas that did not convert should have values closer to 0.

Calculation of Development Pressures or Probabilities
Given that our goal is to predict the likelihood of conversion across the 2 county landscape, we developed a functional mapping between the neural network similarity index (SI) values and probability of future conversion to urban. We assume that future land conversion will take place across the same distribution of similarity index values, so that the probability can be calculated using the percentages of areas that were observed to change to urban from 1962 to 1998 for a given SI value. The probability of conversion by similarity index is given in Figure 1. From this figure it is clear that areas with higher SI values have had higher conversion probabilities. For example, around 50% of areas with SI values greater than 0.6 (scaled to 60 in this figure) were converted to urban during this period, while less than 12% of areas with SI values less than 10 were converted to urban use. A power law curve fit was performed to relate SI with observed probability of conversion. This relationship between landscape SI values and conversion probabilities was used with our forecast of future demand for new built areas to create maps of development pressure. These data are presented in section E.

Task 5: Apply % imperviousness algorithms to examine current (and future projected) watershed conditions.
We have completed the calculation of percent imperviousness by subwatershed for the historical land use estimates. For all project years, transportation features were assumed to be 100% impervious. For the remaining candidate land uses (residential, industrial/commercial, mixed urban, and auxiliary transportation), the data were derived by applying the appropriate area-weighted coefficient to each polygon area. (Note that railroads were assumed to be pervious, and were not included in the calculation.)

The coefficients were generated by overlaying pre-existing, digitized impervious surface data (1998) for a small region in the study area with the 1998 land use data for the 4 classes, and deriving the percent of each class mapped as impervious (see Table 2). The 1998 factors were then applied to the remaining project years.

Task 6: Continue project outreach activities.
Presentations on the project goals, and preliminary results, were delivered at NOAA Headquarters in Silver Springs, MD in November 2004 as part of their on-going seminar series and at a workshop for Local Decision Makers in New Hampshire’s Coastal Watersheds held at UNH in October 2004. We have continued to share project methodologies and project results with constituencies at a variety of governmental, NGO and private sector levels. We have presented a project overview to staff from Metcalf & Eddy, in support of their work on the “New Hampshire Seacoast Region Wastewater Management Study”. Customized products have been prepared and delivered to staff affiliated with the NH NROC (Natural Resources Outreach Coalition) for use in their community outreach efforts. We have also had exploratory discussions with representatives from the Society for the Protection of NH Forests and the Merrimac Valley (MA) Planning Commission (Gaylord Burke, Executive Director). And finally, we have entertained preliminary inquiries and anticipate further interaction with the NH Department of Transportation as they undertake planning for the Newington/Dover (Spaulding Turnpike) project.

Task 7 Prepare final report.
We did not address this task during the past project period since our project will continue through August 2005.

Additional tasks completed during this project period:
We updated our forecast models once we identified some inconsistencies in our TAZ level population and employment data. These new models had greater explanatory power than our previous models. Based on population and employment information for Rockingham, Strafford and Southern New Hampshire Regional Planning Commission regions, our models explained over 78%, 52% and 72% of the observed variance in the magnitude of development at the TAZ level. Results for the Rockingham model are illustrated here in Figure 2.

Difficulties Encountered
No major difficulties were encountered during the past 6 month project period.

Anticipated Success in Meeting Project Objectives in Scheduled Project Period
We continue to anticipate meeting all our data development and modeling objectives, as well as developing a strong outreach component of interested users from regional and state level stakeholders.

Preliminary Data
Table 3 contains the results from our revised economic, demographic and new built area forecasts out to 2020 and 2035. Figure 3 presents the TAZ-level mean similarity index (SI) values. Mean SI values were calculated based on removing areas that cannot, or should not, be developed (e.g. wetland, conservation lands). Figure 4 and Figure 5 present our preliminary TAZ level development forecasts for 2020 and 2035. Note that the total development by 2020 and 2035 is constrained by our forecasts, given in table 2, and allocated to each TAZ based on SI values, the relationship between SI and development probability (Figure 1) and the availability of developable lands.

Figures 6-8 present the percent imperviousness data for the subwatersheds that are fully contained within the two-county study area for the years 1962, 1974, and 1998.

Tasks and Activities for Next Reporting Period

Tasks for the Next Reporting Period
1 Continue to refine and test neural network model.
2 Complete and disseminate project CD.
3 Continue project outreach activities.
4 Prepare final report.

Work Plan to Accomplish Tasks
1. We will continue to refine and select a final model based on our continued training and testing of candidate neural network models.

2. As previously noted, the project CD will be finalized as soon as the brief document presenting preliminary results is available. Other components are already finalized, and the mailing list has been assembled. Thus, distribution of the CD should not be problematic.

3. We will continue to seek opportunities to disseminate the project results. As noted previously in this report, a number of opportunities have been identified over the next few months, including in national, state and regional venues.

4. A final project report will be prepared.

Concerns or Difficulties
We anticipate no significant concerns or difficulties during the next project period.

Expenditures
As interest in this project and project results continues to grow, we are experiencing more requests for project information and project coordination/collaboration. A dedicated technology transfer is needed, as our project expenditures will soon exceed our planned budget.