| Mapping and Change Analysis Study
By David E. Nagel, Scott A. Sutton, and Lee A. Smith The Lower Colorado River Authority (LCRA) is a conservation and reclamation district that was created by the Texas Legislature in 1934. The LCRA has a broad range of duties in its mission to serve central Texas including: flood management, wholesale electric and water sales, electric transmission, parks and recreation, community development, and environmental monitoring. The Authority supplies electricity to over one million Texans through 44 wholesale customers, which include 33 cities and 11 electric cooperatives. Flood control and hydroelectric power generation is achieved through the management of six dams and reservoirs along the Colorado River and additional electric generation is provided through three large fossil fuel power plants. In addition, the LCRA has extensive land holdings including over 15,000 acres of parklands and 2,300 miles of transmission lines and easements.
The LCRA monitors over 22,000 square miles of the Colorado River Basin and watershed. In order to perform comprehensive management of such a large area, the LCRA has implemented a geographic information system (GIS) with a series of applications using geo spatial and remote sensing data. As a supplement to the LCRA's digital orthophoto quarter quad (DOQQ) data, Landsat Thematic Mapper (TM) satellite imagery and derivative land cover information were identified as necessary management tools.
Earth Information Systems Corporation (EISYS) was contracted to complete the land cover classification, along with a 10-13 year change analysis study. A comprehensive cost benefit analysis showed that LCRA would receive a substantial return on its investment by using Thematic Mapper imagery along with its 1:12,000 scale, color infrared, DOQQs. In addition to the land cover and change analysis data, the numerous data layers and other information that could be extracted from the imagery for future projects, was a critical factor in justifying the original project expense. Classification
The technical procedures used in the land cover classification project were partially derived from the image processing protocol developed for the Upper Midwest GAP Analysis Project (Lillesand et al., 1997). ERDAS Imagine software was used as the primary image processing tool while ESRI, ARC/INFO and ArcView products, were used for vector operations.
Portions of five Thematic Mapper satellite scenes were required to cover the entire LCRA project area. The project area was divided into two basins, an upper and lower, that are located to the northwest and southeast of Austin. For Central Texas, spring imagery was deemed appropriate to take advantage of phenological differences among vegetation types in the study area, so the majority of the satellite data were acquired in April, 1997. The images were relatively cloud free, except for portions of two scenes that contained 1-2% cloud cover. Historical Landsat scenes acquired in April, 1985 were purchased to replace data obscured by clouds present in the 1997 data.
Aerial photography was collected for this project as the primary source of ground truth data. Two types of aerial photography were used, digital orthophoto quarter quads and National Aerial Photography Program (NAPP) photos. EISYS distributes DOQQ data for the Texas Orthoimagery Program, so these digital photos were readily accessible. DOQQ data were not yet available for approximately 1/3 of the study area, so 1995 NAPP photography was substituted. A sample of photography was chosen that covered approximately 20% of the ground area within the Lower Colorado River Basin.
Ground truth data were a critical component of the land cover classification project. More than 500 sites were visited in the field and approximately 7000 more were identified using aerial photography. Field sites were selected by viewing the satellite imagery and aerial photography simultaneously, and choosing locations that were thought to represent each land cover type present in the project area. These polygonal locations were delineated on the aerial photography, which was used as the primary navigational aid for field data collection. The sites were located adjacent to roads so that they would be easily accessible. The land cover type was recorded at each place and the data were then digitized on-screen, and attributed. The process helped to familiarize the image analysts with the vegetation of Central Texas and provided a knowledge base for interpreting the aerial photography. The analysts were then able to collect ground truth data directly from the aerial photography, without going into the field. These ground truth data provided numerous examples of each land cover type, and an attempt was made to capture the full range of variability within each cover type. One-half of the ground truth sites were randomly selected and used for classification training, while the other half were used for the post-classification accuracy assessment. The ground truth training data were next separated by land cover categories, so that individual classes could be evaluated and processed separately.
Twelve categories were identified:
High Intensity Urban
Low Intensity Urban
Golf Courses and Parks
Cultivated Land
Grassland
Deciduous Forest
Cedar
Pine Forest
Woodland/Shub Land
Barren Land
Wetland
Water
Using the satellite imagery, spectral signatures were developed from the ground truth polygons to represent each land cover type. The signature sets were then evaluated for distinguishable details and comprehensiveness. These individual signature sets were then combined to create a master set representative of all of the land cover categories defined for the project area. The final signature set was then applied to the satellite imagery using a maximum likelihood algorithm. The resultant classification file was generalized to yield a land cover layer with a minimum mapping unit of approximately 1 acre.
To avoid confusion with barren land, urban areas were classified separately from the rest of the satellite imagery. Using aerial photography as a guide, urban areas were delineated on the satellite images and extracted for independent processing. An unsupervised classification routine was used to separate high intensity from low intensity urban types.
The accuracy of the classification was tested to provide a measure of reliability of the land cover information. Half of the ground truth data, which were collected from field visits and aerial photography, were reserved for this purpose. When conducting an accuracy assessment, it is preferable if the site locations are chosen randomly, to avoid bias. In doing so however, sites often fall on the border between land cover types, in areas where the land cover class is ambiguous from aerial photography (e.g., pasture with invasive shrubs) or in areas that are inaccessible in the field. To avoid these pitfalls, accuracy assessment sites were chosen at the same time as the training data, before the classification process began. These sites were chosen deliberately and represented a diverse range of examples for each cover type. Because the sites were collected before the classification process began, the bias that sometimes results when image analysts perform accuracy assessment on a classification product that they are very familiar with, was avoided. To introduce an element of spatial randomness, the ground truth sites were randomly divided into two groups, one to be used for training and the other for accuracy assessment.
The entire project area was divided into seven classification units based on eco-regions. The units were then classified separately from one another to minimize variance across the extent of the satellite imagery. Accuracy assessment was also performed on these individual strata. The overall accuracy for each classification unit exceeded 85% and individual class figures surpassed 75%.
The final land cover layer was converted from ERDAS Imagine raster format into two different ESRI file types, to facilitate integration with existing LCRA data. The Imagine files were first converted to ESRI raster GRID format and tiled by 7.5 minute quadrangle boundaries. These files were then converted to ARC/INFO vector/polygon data. Because raster data are comprised of square cells, the resultant vectors contained only perpendicular angles and the data were stair-stepped at diagonal land cover boundaries. This block-like structure is visually distracting and gives the land cover data an unnatural appearance. To create a coverage with smooth boundaries between land cover types, an ARC Macro Language (AML) program was written. The vector data resulting from this process more closely followed the natural forms of the landscape. Both the GRID and vector data were deliverable products. Change Detection
The second phase of the project focused on changes that had occurred in the study area over a 10-13 year time period. Historical TM imagery was acquired and compared with the 1997 data to determine where land cover changes had taken place.
Many different technical approaches to change detection have been documented. A few of the most popular methods include post classification comparison, principal component analysis, and image differencing. Post classification comparison was eliminated as an option because errors in both the historical and recent classification data are generally compounded in the final change analysis product. Principal component analysis was not used because inconsistent results may occur between adjacent images. In contrast, image differencing offered a consistently repeatable step-by-step process that could be checked for accuracy at each point along the way. Consequently, this method was selected.
Image differencing involves subtracting one channel of TM data from another, where the 2 channels represent the same spectral band, but are acquired on different dates. Where the result of the subtraction is close to 0, no change has occurred. Where the values diverge from 0, spectral changes transpired between the first and second acquisition dates.
After testing numerous options, it was determined that two difference files would be created for each image to identify the areas of land cover change. Empirical results showed that a TM band 7 difference image, along with a Normalized Difference Vegetation Index difference image, was the most effective means for identifying changes in the project area. This information was used to create a binary change mask showing those areas that had changed and those that had not. At this point, the change mask only showed areas of spectral difference between the two image dates. In some cases, change was indicated, even if the land cover type hadn't changed. For example, grassland that received a great deal of rain in one image but not in the other, would appear spectrally different, even though the land cover class remained the same.
A sample of National High Altitude Photography (NHAP) aerial photos, flown between 1983 and 1985, were obtained to assist in the change classification and accuracy assessment. The photos provided approximately 60% coverage of the study area. To ensure that urban expansion was best captured, photos were obtained for each urban region classified in the 1997 imagery.
Using the binary change mask, the next step was to classify the areas of change, using the 1980's vintage TM imagery. However, rather than classify the entire change area at one time, the binary change mask was stratified using individual land cover classes from the 1997 classification. An unsupervised classification algorithm was used for this processing. Knowing the 1997 category, the image analysts could carefully scrutinize each spectral class and determine which 1980s land cover class the signature should be assigned to. This type of careful analysis could not be completed if all of the change areas were classified together.
Each individual classification was combined to create a land cover layer for the 1980s time period, but only where change had occurred. This 1980s data was overlain with the 1997 classification to produce a full land cover layer for the earlier time period. Land cover data was now available for the 1980s, and 1997 time periods. Thirty-three "from and to" classes that represented important changes were then identified by comparing the 2 land cover layers. This "from and to" GIS layer was carefully examined along with both dates of satellite imagery to determine if all of the actual change was captured, or if change was being identified where it had not occurred. Discrepancies were corrected using a variety of techniques such as GIS modeling or raster editing.
An accuracy assessment was performed on the 1980s classification, within the areas of change. In this case, a completely stratified random sample of accuracy assessment points was generated for the land cover data and these points were checked against the aerial photography. Overall accuracy figures for all of the classification units exceeded 80%. The final 1980s vintage classification and the "from and to" change analysis files were converted to GRID and vector formats, just as the 1997 land cover data had been. Results
Results of the change detection analysis showed some consistency between the lower and upper watershed basins. Conversion from grassland to shrub land was the largest change in both basins, followed by barren land to grassland. Other prominent changes indicated that cedar and shrub land were frequently cleared, giving way to grassland. Urban expansion was also an important change component in the lower basin, while shrub land gave way to cedar in the upper.
The land cover classification and change detection projects were a cost effective means of developing historical and baseline spatial data sets for further analysis and implementation into an enterprise GIS. The LCRA now uses land cover data extensively to monitor historic changes in the watershed and their impact on water quality. Other uses include soil erosion studies, impacts of impervious cover, calculations of the spread of invasive vegetation regimes, and watershed modeling. Reference Sited:
Lillesand, T., J. Chipman, D. Nagel, H. Reese, M. Bobo, R. Goldmann, 1998. Upper Midwest GAP Analysis Program Image Processing Protocol. USGS Technical Paper 98-G001., USGS Environmental Management Technical Center, Onalaska, Wisconsin. About the Authors:
David Nagel and Scott Sutton are Image Analysts at EISYS. Lee A. Smith is Technical Operations Manager at the LCRA. Back |
