Home  Projects  Members  Publications  Committee  Data  RSPA  Meetings  Contact 
We utilize archived remote sensing data for regionalscale assessment of historical changes in land cover that has occurred over a 91,000 km2 area since 1980. More specifically, we develop multitemporal land use/land cover maps of the region and compare the observed trends in development and environmental change to changes in watershed hydrology. Whereas "development" is often quantified in terms of population or housing statistics in a geospatial context of counties, zip codes, or other "districts," remote sensing offers an alternative means of quantifying development in a pixel or segmentbased geospatial context. We analyze the longterm hydrologic trends of 18 watersheds of various sizes and comprised of different land cover types. Two watersheds, the Flint River, Alabama and Big Creek, Georgia are analyzed in more detail. Whereas these watersheds were principally rural in 1980, they have succumbed to urban sprawl over the past two decades, which has had a significant impact on the watershed's hydrology.

The study area measures approximately 350 x 260 km comprising 60 counties of the southern Appalachian region of northeastern Alabama, northwestern Georgia, and southcentral Tennessee (Figure 1). This region includes the metropolitan regions of Atlanta, Georgia, Chattanooga, Tennessee, Birmingham, Alabama and Huntsville, Alabama.

The land cover classification methodology is discussed elsewhere:
http://wwwghcc.msfc.nasa.gov/land/ncrst/raclassmeth.html

Results of the land cover classification is discussed elsewhere:
http://wwwghcc.msfc.nasa.gov/land/ncrst/raclass.html

The U.S. Geological Survey operates stream gages on rivers in each of these watersheds. Watersheds were chosen that were of an appropriate size such that changes in hydrologic response due to transportation related land cover change would be evident. The criteria for selection included relatively long streamflow records, less than 900 square miles in area, wide distribution around the region, and representative of a variety of land use/land cover conditions from very rural forested and agricultural to predominantly urban watersheds. Nineteen watersheds that range in area from 72 to 885 sq. mi. met these criteria. Using the 1:250,000 scale DLG files available from the USGS Eros Data Center, the drainage areas of these watersheds were delineated in the ArcView GIS environment. They are shown in Figure 2.
Figure 3 is a collection of plots showing annual peak streamflow in each of the watersheds from the beginning of the record or 1930 to 2002. It was initially determined to characterize the hydrologic response within the region in terms of three variables. These variables were mean annual streamflow, frequency of inundation, and duration of inundation. These variables were selected due to their perceived relationship to ecosystem and environmental conditions. Mean water level, frequency of inundation above specified levels, and duration of these inundations could be related to various environmental concerns including wetlands identification, endangered species habitat and flood plain analysis. The streamflow records of the selected basins were examined for statistically significant timedependent trends in all three of these variables. Mean monthly streamflow for each watershed streamgage was obtained from the permanent archives of the USGS. These data were used to determine mean annual flow for each year of record for each site. Mean monthly rainfall data was obtained for a large number of raingages throughout the study area. Mean annual streamflow (cfs) was converted to water depth and then normalize for total annual rainfall to remove possible climatological trends in rainfall that might bias streamflow records. Rainfallnormalized mean annual streamflow was plotted for the period 1980 to 2000. Trends in these time series are reflected in least squares regression (Figure 3). The slope of the observed trends is low; eight of the 19 are negative trends. Due to the short duration of the record (20 years) and high degree of temporal variability, none of the trends is statistically significant at the 0.05 level of significance. Nevertheless, these data suggest that for the period between 1980 and 2000, there may be a positive relationship between the slope of the trend in rainfallnormalized mean annual streamflow and the percentage change (increase) in developed land within each river basin (Table 1, Figure 4).
Table 1: Percent change in developed land in 19 watersheds and the associated trend in mean annual streamflow.
The frequency of inundation above a given threshold value is merely the number of times each year that the streamflow exceeds the threshold limit. However, because this "frequency" is a discrete random variable (a tally of occurrence), its sampling distribution is unknown and therefore cannot be examined directly for statistical significance. In order to analyze these data by Poisson regression, the data must be transformed using a technique described by Keim and Cruise (1997). This technique is based on the assumption that a Poisson counting process can represent the frequency of occurrence of events above a certain threshold level, then the time intervals between recurrences of the process are exponentially distributed. The data series (intervals between events) can then be summed in groups of two or four (depending on the length of the record and the desired power of the test) in order to transform the data to lognormally distributed random variables. Next, the data are further transformed to a normal distribution by merely taking the logarithm of each summed group. This data series is then regressed against the cumulative midpoints of the summation of the group times in days and the regression holds exactly. The first step in the application of Poisson regression is to find the threshold limit at which the counts of occurrences are Poisson distributed. Figure 5 shows the peak annual streamflow for each watershed. A test first devised by Cunnane (1973) and later expanded by Cruise and Arora (1990) can be used to make this determination. The test relies on the fact that the mean and variance of the Poisson distribution are equal. Thus, the ratio, R = Var(n)/E(n) (where n is the number of events per year) should approach unity as the threshold level is increased if the process is indeed Poisson admissible (Cruise and Arora, 1990). The R values are tested using the test statistic R(N1) which is known to be c2 distributed (Cunnane, 1973) and where N = total number of years of record. Thus, the critical R value would be given by Rc = c2 (N1),a/N1. Keim and Cruise (1998) recommend a significance level of 0.1 for this onetailed test. The daily streamflow data were obtained for each of the 10 watersheds for which annual trends were identified and the test described above was applied to each series. In all ten cases, a threshold level for Poisson admissibility was successfully identified. Once the lowest Poisson admissible threshold is identified, the series will theoretically remain Poisson distributed for all thresholds above this value. The next step is to determine the number of days between recurrences of streamflow events above the threshold level. This was accomplished for each data series and then the steps described above were completed to determine if significant trends in frequency of inundation above the thresholds were evident. In the duration analysis, the same threshold values were used as in the frequency analysis for consistency purposes. The analysis of durations is made more easily since the annual duration of flooding events has traditionally been assumed to be a lognormally distributed variable due to the fact that the series arises through a summation process. Thus, the procedure merely consisted of determining the number of days each year that each stream remained above the determined threshold and taking the logarithms of these values. The logarithms were then regressed against cumulative time in days to determine if significant trends exists.

The Flint River, with and area of 568 square miles, is a tributary to the Tennessee River. Much of the watershed (342 sq. mi.) is contained in Madison County, Alabama (Figure 6). The land within this watershed is predominantly agricultural and has experienced significant recent residential growth from the City of Huntsville. The U.S. Geological Survey National WaterQuality Assessment Program is currently investigating water quality in the lower Tennessee River basin with several monitoring activities targeted in the Flint River Basin (Hoos et al., 2002).
Table 2 shows the area of Developed and Undeveloped Land in 1984, 1990 and 2000 as determined by classification of satellite remote sensing imagery. These data are shown graphically in Figure 7. In 1984, 5.5% of the basin in Madison County was developed. By 1990, the area of Developed Land had increased by 121% to 12.1% of the area of the basin in Madison County. By 2000, the area of Developed Land had increased by an additional 67% to 20.2%. Over the 16year period, the area of Developed Land increased by an average of almost 1% per year. From 1980 to 1994, rainfallnormalized mean annual streamflow increased by 17% (Figure 8). (In 1994, the gaging station was moved down river several miles. The new streamgage did not begin acquiring data until 1996, however, these data are not directly comparable). If the trend in streamflow was projected to the present year, we could estimate that streamflow in the Flint River Basin to be 29% greater than it was in 1984. The reason for the increase in streamflow is not known with certainty, but these data suggest that it is related to the increase in the percentage of Developed Land.
Table 2: Area of Developed and Undeveloped land in Flint River watershed and associated statistics.

The Big Creek watershed is a small basin on the northern fringe of Atlanta, Georgia's vast urban sprawl (Figure 9). The area of the basin is approximately 72 square miles upstream from USGS Gage # 02335700 (Big Creek near Alphareta, GA) and is contained in three counties: Fulton, Forsyth and Cherokee. From 1980 to 1998, the population of the basin increased by 200% and the number of new businesses increased 15fold (Figure 10). Increases in developed land of this magnitude are associated with concomitant increases in Developed Land and impervious surfaces, which have a significant effect on the hydrologic regime. Because of the magnitude of changes within Big Creek, it was selected for study using remote sensing imagery to classify land use and land cover of Big Creek.
Changes in the percentage of developed land within the watershed are correlated to changes in streamflow at decadal intervals (Figure 11). In 1980, 12% of the area of the watershed was Developed Land. In 1990, Developed Land area had increased by about 40% to 17% of the area of the basin. From 1990 to 2000, the area of Developed Land had nearly doubled to 32% or nearly onethird of the area of the basin. Figure 12 shows rainfallnormalized mean annual streamflow for Big Creek from 1960 to 2000. From 1960 to 1985, streamflow was fairly consistent with a slight positive trend. From 1985 to 2000, however, mean annual streamflow was highly variable. Although there is a significant positive trend during this latter period, there is not a significant increase in streamflow because of a general decrease in streamflow during the latter half of the 1980's. Causality cannot be determined with any certainty, but clearly the increase in variability since about 1985 is highly correlation with an increase in Developed Land area.

Cruise, J.F. and Arora, K., 1990. A hydroclimatic application strategy for the Poisson partial duration model, Wat. Res. Bull., 26(3): 431442. Cunnane, C., 1973. A note on the Poisson assumption in partial duration series models, Water Resources Research, 15(2): 489494. Hoos, A.B., Garrett, J.W., an Knight, R.R., 2002. Water Quality of the Flint River Basin, Alabama and Tennessee, 19992000., WaterResources Investigations Report 014185, U.S. Geological Survey, Nashville, TN, 37p. Keim, B.D. and Cruise, J.F., 1998. A technique to measure trends in the frequency of discrete random events, Journal of Climate, 11(5): 848855.
