Project Title: Remotely Sensed Indices of Land Use Intensity for Watershed-level Monitoring
Principal Investigator(s): Richard G. Lathrop Jr., Robert A. Zampella

Accomplishments
Scheduled Tasks:
The original timeline assumed a start date of late spring 2000 with our first milestone scheduled for August 2000. This first task was to acquire imagery and ground truth for the study site. Our second milestone was scheduled for January of 2001 and consisted of completing the IKONOS land use mapping. As the actual start date of the project (in terms of funding received at Rutgers) was not until 11/15/2000, this schedule has been pushed back.

Once the imagery was in hand, our next scheduled task was to investigate the comparative utility of various image processing techniques for the estimation of impervious surface cover.

Progress on Tasks
We were successful in acquire satellite remotely sensed imagery for the Mullica River study region. Landsat 7 Enhanced Thematic Mapper (ETM+) imagery was acquired on April 5, 2001 during the leaf-off season. Two days prior on April 3, 2001, we acquired Space Imaging IKONOS scenes over two 5km x 5 km study areas in the Mullica River basin. Due to the comparatively high cost of the IKONOS imagery, imagery for only a comparatively small subset of the basin could be acquired. These two study areas included a mix of suburban and rural land use/land covers.

Based on our preliminary results, we felt it was necessary to collect additional IKONOS imagery over an area that is more intensely urban with a higher percent of impervious cover, as well as under some different conditions. We successfully acquired additional IKONOS imagery over our study area during November, 2001. We may acquire additional imagery if available.

Our objective is to investigate the utility of satellite remotely sensing techniques to map indicators of urban land use intensity: impervious surface and the managed lawn. Using medium scale imagery (e.g. Landsat Thematic Mapper), individual pixels generally represent a mixture of urban land covers. Thus we are concentrating our efforts at developing un-mixing techniques that estimate subpixel proportions of the land covers of interest. Three different supervised classification methods are being evaluated: traditional maximum-likelihood hard classification, supervised fuzzy c-means (FCM) and Self-Organizing Map (SOM)-Learning Vector Quantization (LVQ) neural network.

Linear mixture modeling
We have been experimenting with linear mixture modeling of Landsat ETM+ imagery for estimating sub-pixel level of three different land surface cover, namely impervious surface, managed lawn and forest vegetated surfaces. Various linear mixture models with different end-member sets have been examined:
one impervious surface end-member set (using the median value of all impervious surface training set),
two impervious end-member set (using the extreme value of bright and dark impervious surface end-member set).
three impervious end-member set (using the median value of bright, medium-dark, dark impervious surface training set), and

Among these three different LMMs, the LMM with two extreme impervious surface end-member set gave the best results.

We have been comparing our modeled results with several different validation data sets. Validation of the end-member images with high-resolution (1m pixel) GIS maps of land cover visually interpreted from rectified aerial photographs, revealed that the total estimation of three classes are quite accurate, even though impervious surface is slightly over-estimated.

Supervised-classified IKONOS data with three land cover classes is also used as a reference data. Compared to IKONOS, LMM slightly overestimates impervious surface and lawn and underestimates forest.

Currently, we suspect that this over- and under-estimation of LMM is caused by the linear mixture assumption of LMM. In reality, except impervious surface, the reflectance of some land cover classes (e.g., forests) are not mixed linearly but rather nonlinearly. Additionally the result reveals that the confined-LMM we used is not strictly confined. So the sum of the estimated percentage of all land cover classes is somewhat over 1.

Fuzzy Classification
As a non-parametric method, Fuzzy c-means clustering shows good result validated to NJDEP LU/LC data (reported in last report). Preliminary validation of the end-member images with high-resolution (1m pixel) GIS maps of land cover visually interpreted from rectified aerial photographs, reveals that impervious surface and lawn are slightly under-estimated and forest is over-estimated.

Validation with classified IKONOS is still on-going.

Neural Network Techniques
Until now, one of the greatest difficulties of using SOM-LVQ is its sensitivity to different initialization parameters. We therefore, tested the algorithm using many different trials to find the best SOM-LVQ model. In previous report, SOM-LVQ neural network provided a slight better result compared to FCM in the validation with NJDEP LU/LC. But with more detailed as aerial photograph based vector map reference data set, such the preliminary result provided worse performance compared to LMM or FCM. Even though this result is based on only partial results, it contradicts to our original expectations. As the most computer-intensive and complicated method, SOM-LVQ neural network was expected to provide the most accurate estimation, but the results highly over-estimate forest and the highly under-estimate lawn area.

However, these results are still preliminary as they are based on one photo plot and we will be investigating other photoplots and IKONOS imagery.

Figure 1 graphically shows the results of the three algorithms for one of the test areas.

We are continuing to refine the various techniques. With the additional IKONOS imagery that we have recently purchased we plan on finalizing our validation assessment during the next study period and then move onto the watershed based comparison with water quality data.

Difficulties Encountered
Our original schedule ( as laid out in the proposal) had been pushed back by over six months as an official University account was not established until 11/15/2000. Due to the significant cash outlay required to purchase the IKONOS imagery, we had to wait until our account was established to buy data. While we have successfully acquired IKONOS imagery over the study area during the Spring 2001 time period, we felt it is necessary to purchase additional IKONOS imagery over areas with higher impervious surface cover and under different vegetation conditions. We have successfully done so.

Anticipated Success in Meeting Project Objectives in Scheduled Project Period
We have requested and received a 12 month no-cost extension for the project due to delays in getting the project funding in place. We are now making steady progress and do not anticipate any difficulties in meeting the revised schedule.

Preliminary Results
Preliminary impervious surface cover maps for the developed areas of the Mullica River basin using Landsat Thematic imagery and based on the three different methodologies.

Tasks and activities for next reporting period

Tasks for the next reporting period
Continue work on refining the estimation of impervious surface cover and lawn area based on the Landsat ETM+ imagery using the various algorithms described above.

Investigate the above methods using the higher spatial resolution IKONOS imagery, as compared to Landsat ETM+.

Work plan to accomplish tasks
S. Lee, the Graduate Assistant assigned to this project, J. Bognar, Project Coordinator, and R. Lathrop, the PI will continue to work on the above tasks.

Concerns or difficulties
None.

Expenditures
At this point, are in the range anticipated for the work accomplished to date.

List of Presentations on Project:

A. Faiz Rahman and Richard Lathrop. 2002. Estimating Urban Land Use Intensity by Combined Use of Aerial Photography, GIS and Spectral Un-mixing of Satellite Imagery. American Association of Geographers Annual Conference, Los Angeles, CA, April, 2002.

Sangbum Lee and Richard Lathrop. 2002. Comparison of three different supervised classification methods in estimating land cover proportions of urban sprawl. American Association of Geographers Annual Conference, Los Angeles, CA, April, 2002.

Sangbum Lee and Richard Lathrop. 2002. Sub-pixel estimation of urban land cover intensity using fuzzy c-means clustering. American Society of Photogrammetry and Remote Sensing Annual Meeting, Washington, DC. April 2002.