We construct scenarios of land-use change from 1991 to 2101 and analyze their implications for nutrient release in the Ipswich River Watershed, Massachusetts, USA. In the process, we create a GIS-based method that can be applied elsewhere to simulate human impacts across a landscape. The research makes important links in integrative modeling because the Ipswich River flows into Plum Island Sound, a salt marsh estuary. The estuary and its watersheds are a Long Term Ecological Research site funded by the National Science Foundation.
The land-use change component of the analysis concerns four land types: forest, residential & commercial, agriculture & open, and wetlands. We have land use maps from 1971, 1985 and 1991. We model land-use change based on spatial physical factors, legal constraints and extrapolations of quantities of change. We calibrate maps of suitability for deforestation with maps of real change between 1971 and 1985 by using multi-criteria analysis. The maps of 1971 and 1985 serve also as the basis to extrapolate the quantity of deforestation predicted for the future. The extrapolated quantities and calibrated suitability maps predict the location of deforestation between 1985 and 1991. The predicted deforestation maps are validated with the map of real forest area of 1991. Relative Operating Characteristic (ROC) and variations of kappa index of agreement measure the validation at various resolutions. In the most successful simulation run at the finest resolution Kno=84%, Klocation=15%, Kquantity=99% and Kstandard=14%, ROC=72%. To predict land-use change into the future, we use our validated simulation method to sketch a pair of scenarios. The scenarios compare a future with restrictive laws concerning land use versus a future without such laws.
The nitrogen component of the analysis is based on our study of the relation between land cover and nutrient concentrations in streams that drain small catchments within the Ipswich River watershed. This analysis is based on an empirical relationship between upstream land-use distribution and nitrate levels in small streams, which we obtained from intensive water nutrient sampling campaigns. For every decade in our two land-use change scenarios we apply this empirical nitrate loading relationship to several hundreds of sub-catchments. The results are time series of spatially distributed surfaces of nitrate loading, with corresponding transects of potential nitrate concentrations for the Ipswich River stream network. Simulation results for both development scenarios indicate a doubling in nitrate delivery to the Plum Island Sound estuary from 1991-2101. These results are preliminary, especially since they don't include any in-stream processing component. However, they demonstrate that future residential development could lead to a substantial increase in nutrient delivery, which could have serious implications for eutrophication of downstream riverine, wetland and estuarine ecosystems.
The Ipswich River watershed is in Northeastern Massachusetts, 30 miles north of the city of Boston. The watershed covers 404 square kilometers and includes parts of 21 towns. The watershed's human population is approximately 130,000 with median family income of $50,000, whereas the United States median family income is $39,000. The area is attractive to upper-middle class people who wish to commute to the city of Boston. Urbanization has occurred in the watershed following a rapid urban expansion of the city of Boston. In the last 20 years, the most important change in the Ipswich watershed is the replacement of forest by residential plots, which tends to increase nutrient loading into the water (Howarth et al. 1996).
Forest in New England underwent constant extensive land use change during the last 400 years (Foster 1995). After intense agricultural practices in the 1800’s, the first half of the 1900's has seen abandonment of agricultural land and forest re-growth. Thus, the current forest cover in the watershed is a product of recent agricultural abandonment and forest succession. However, since the middle of the 1900's, there has been conversion of forest into residential areas. Therefore, our models simulate conversion from forest to non-forest, usually as a result of new residential and commercial land.
The Ipswich River flows into Plum Island Sound, which is a relatively pristine, macrotidal, salt marsh estuary along the Gulf of Maine. The estuary and its watersheds, the Parker and Ipswich Rivers, are a Long Term Ecological Research (LTER) site supported by the Untied States National Science Foundation. A primary goal of the LTER research is to understand the effects of land use change on trophic structure and primary production of the estuary.
Data for the Ipswich River Watershed, Massachusetts
The Executive Office of Environmental
Affairs of Massachusetts supplies GIS data through the MassGIS program (MassGIS 1999). Land use maps are available for three points in
time: 1971, 1985 and 1991. The original classification of land-use of 21
categories is aggregated to four: 1) forest, 2) residential & commercial, 3)
agriculture & open, and 4) wetlands. An additional map of wetlands, which is
from 1:12000 scale color-infrared stereo photography, is overlaid on the
original land-use maps to revise the wetlands information. In the 1991 map, the distribution of
major land cover types in the watershed is 44% forest, 31% residential and
commercial, 17% wetlands and 8% agriculture and open. Residential and
commercial areas include multi-family units, small and large residential plots,
shopping centers and industry. Agriculture and open lands include cropland,
pastures, vacant undeveloped land, and open urban areas such institutional green
space. Maps of land cover for 1971, 1985, and 1991 show that a conversion of
forest into residential areas is the predominant land-use change.
Urban development and deforestation are partially constrained by a variety of legal constraints. A constraint is a legal characteristic of a location that prevents it from being deforested. Therefore, a constraint variable is a Boolean variable that indicates whether the location either is or is not available for deforestation. There are a variety of constraints on conversion of land to residential and commercial uses. These derive from laws relating to: the wetlands protection act, the rivers protection act, FEMA floodways, well head protection buffers, zoning, permanently protected space, and minimum lot size. We combine maps of legal constraints into one map that shows all constraints. Presently all wetlands have at least one legal constraint, and half of the forest area of 1991 has at least one legal constraint. Overall, 21% of the watershed is protected forest and 23% is unprotected forest.
We consider four factors that relate to the likelihood that any particular location is converted from forest to residential & commercial. The factors are: proximity to major roads, proximity to minor roads, proximity to rivers, and proximity to existing residential areas. All the variables are mapped in a raster based Geographic Information System with 30-meter by 30-meter grid cells.
Model calibration
We use the land maps of 1971 and 1985 to calibrate our land use
change model. Then we subject the model to validation with 1991 map data. We
describe the calibration in order of its two major components: location and
quantity.
To specify location, we use multi-criteria analysis to create a suitability map in which each grid cell contains a real value between 0 and 1, which represents its relative likelihood of conversion from forest to non-forest (Carver 1991, Jankowski 1995, Eastman et al. 1995, Eastman 1999). We create the suitability map in three stages, similar to Pontius Jr et al. (in press) and Schneider and Pontius Jr (in press).
First, we convert the four factor maps into four suitability maps. Each grid cell in the factor map is a real distance, which we place into one of several ordered categorical bins, where the bin sizes are: 0-100 meters, 100-200 meters, 200-300 meters, etc. Then we compute for each bin the proportion of the bin that experienced deforestation from 1971 to 1985. Hence, we reclassify each of the four maps of bins to four maps of proportions of deforestation.
Second, we use a weighted linear combination to merge the four maps of proportions of deforestation into a single map of suitability for deforestation. The weights are proportional to the coefficients in the first principal component, which results when we perform principal components analysis on the group of four maps of proportions of deforestation (Fung and LeDrew 1987).
Third, we overlay the map of constraints on the combined suitability map to give every constrained grid cell a minimum suitability. The suitability value is not the same as a probability. The suitability map shows the sequence in which the model chooses land for deforestation (Pereira and Duckstein 1993). For the case in which we do not consider constraints, we eliminate this third step.
To specify the total quantity of land to be converted, we extrapolate the quantity of deforestation from 1971 to 1985 using three methods: linear, logistic, and exponential decay. Extrapolation proceeds temporally until all unconstrained land is converted. The model can extrapolate deforestation either by town or by the entire watershed.
A simulation run uses the combined suitability map and a predicted quantity of deforestation to generate a map of predicted deforestation between 1985 and 1991. The model simulates deforestation for cells that have the highest relative suitability.
Model Validation
For validation, we compare the model's output map to a map of
real deforestation for the interval 1985 to 1991. Successful predictions occur
when a grid cell in the map of real deforestation matches the category in the
map of simulated deforestation. We compute several statistics that measure
overall agreement and additional statistics that separate the agreement due to
location from the agreement due to quantity.
The kappa index of agreement has several variations, each of which measures different characteristics of agreement (Pontius Jr 2000). Kappa equals 1 when agreement is perfect and kappa equals 0 when agreement is as expected by chance. Kno gives the overall accuracy of a simulation run. Klocation indicates the level of agreement of location, given a specified quantity. Kquantity indicates the level of agreement of quantity, given the model's ability to specify location. These variations complement the standard Kappa.
In addition, we separate the sources of success and errors across various spatial scales. We generate windows of increasingly coarse grids, and evaluate the sources of success and error at each grid resolution. This method of validation shows the extent to which errors are near to each other in space.
Finally we use the relative operating characteristic (ROC) to validate the suitability map for several scenarios of specified quantity of deforestations (Swets 1988, Pontius Jr and Schneider, in press). The ROC shows whether high suitability values are associated with real deforestation and whether low suitability values are associated with forest survival. In a perfect suitability map, the order of the suitability values would match the order in which the landscape changes, with the largest suitability values changing first and the smallest last. A perfect suitability map gives ROC equal to 1, and a suitability map in which suitability values are located at random gives ROC equal to 0.5.
Model Extrapolation
Based on the validation results, we determine the best
combination of modeling techniques to extrapolate deforestation. We use those
techniques to extrapolate deforestation from 1991 to 2101. To extrapolate the
locations, we recalibrate the suitability map using the land maps of 1971 and
1991. To extrapolate the quantities, we compute an exponential decay curve for
each town.
We examine two scenarios. First is a business-as-usual scenario in which historic patterns are extrapolated with the existing structure of legal constraints. Second is a scenario of relaxed legal constraints, in which we do not use the constraint maps. Hence for the second scenario, any forested cell is a candidate for deforestation, and there is a larger quantity of deforestation.
Interface with Nitrogen Model
Interfacing our land-use change scenarios with a nutrient loading
model yields scenarios of nutrient release into the Plum Island Sound estuary.
In order to investigate the effect of a changing, spatial arrangement of
land-use, we develop empirical nutrient loading relationships for small streams,
and apply them within a GIS framework to the entire Ipswich River watershed,
similar to Alexander et al. (2000). For preliminary purposes, we look at only the effect
of land-use change on nitrate loading and do not incorporate any changes due to
in-stream nutrient processing, uptake or release.
We consider the nutrient loadings from individual sub-catchments as a black box and route them downstream through the river network. We do not consider other forms of inorganic or organic nitrogen, since our preliminary results suggest that only nitrate loading is related to land-cover. We focus on nitrogen because it is the primary nutrient responsible for estuarine eutrophication. We do not address downstream nutrient processing, rather we simply calculate downstream mixing. Watershed level nitrogen budgets indicate that instream processing removes up to half the nitrogen entering the river network. We have no idea whether this level of processing will continue into the future or whether the system is presently approaching its nutrient reduction capacity.
The nitrate loading relationship is obtained from seasonal nutrient sampling of about 50 small streams during non-storm periods (April 1999, November 1999, February 2000 and April 2000) in which we compare the measured nitrate loading to the fractional cover and spatial arrangement of upstream land-use. Nitrate loading is expressed as a function of land-use fractional cover, where each land-use grid cell is weighted by its distance from the outlet, as follows:
NO3- = 90.97 * f2 - 6.19 * f
where: NO3- is nitrate concentration [mM] and f is the distance-weighted fractional cover of agricultural and urban land-use of the grid cell.
The entire Ipswich River watershed is segmented into several hundred subcatchments by using a hydrologically corrected DEM (MassGIS 1999) and by applying a minimum source area of 1.5 square kilometers. For each decade in our two land-use change scenarios, we calculate the distance-weighted land-use fractional cover for these subcatchments. Subsequently we apply the nitrate loading relationship to each subcatchment, resulting in decadal, spatially distributed surfaces of nitrate loading. We then route the nitrate loading downstream, producing transects of 'potential' nitrate concentrations for the entire Ipswich River stream network.
Visual inspection shows modest agreement in terms of location (figure 1). This is confirmed by Klocation = 15% and Kstandard = 14%. Even at coarser resolutions, the success due to location is dwarfed by the success due to quantity, and nearly all of the error is associated with the specification of location (figure 2).When weighted over all resolutions, success due to quantity is 91%, success due to location increases that to 92%, error due to location is 7% and error due to quantity is 1%.
However, the order of the suitability values matches to a good extent actual deforestation, as indicated by ROC = 72%, which is substantially better than would be expected due to chance. ROC ranges from 47% to 53% among 10000 Monte Carlo runs where location is specified at random. There are a few points of special interest on the ROC curve (figure 3). The two points nearest (not at) the origin are the quantities that represent the correct and estimated quantities. The second to last point near (not at) the point (1,1) represents deforestation at all unconstrained locations. Thus the constraints map is one main reason for the high ROC. When we did not use the constraints map, the ROC is 68%.
In the first scenario, 44% of the 1991 forest area becomes deforested between 1991 and 2101 (figure 4). In the second scenario, which ignores legal constraints, about 49% of the 1991 forest area becomes deforested. However, in the second scenario, 44% of the deforestation occurs on land that has legal constraints. Therefore, the constraints influence the scenarios modestly in terms of quantity and substantially in terms of location.
Future Nitrogen release
due to deforestation
Comparing the surfaces of nitrate loading for 1991 (figure 5)
versus 2101 (figure 6)
indicates that nitrate loading increases significantly. In 1991, the main areas
of nitrate loading are located primarily in the upper part of the basin, while
the central part of the basin displays overall low levels. In 2101, the central
part of the basin displays high levels of nitrate loading as well.
Figure 7 shows how nitrate delivery to the estuary changes during the period 1991-2101 for both development scenarios. For the first scenario, which maintains 1991 legal constraints, the rate of change in nitrate delivery starts to level off around 2060. For the second scenario, which has no legal constraints, the decrease in rate of change in nitrate delivery is several decades later and less pronounced compared to the first scenario. Both scenarios suggest that nitrate export from the watershed and delivery to the estuary will double during the 1991-2101 simulation period (93 % increase for scenario 1; 119 % increase for scenario 2).
The transects of nitrate loading for the main stem of the Ipswich River (figure 8), show that the highest levels of nitrate concentrations are in the upper part of the basin, where there is relatively more urban development. As you move downstream, nitrate levels drop, due to dilution by water coming from less-developed areas. All three transects have similar shapes, even though the 2101 simulations predict relatively much higher nitrate loadings in the central part of the basin, compared to the 1991 simulation.
The nitrogen modeling results presented in this paper are very preliminary, especially since we look only at nitrate loading, neglecting any downstream processing, uptake or release of nitrogen. Keeping in mind the limitations of our simplified approach, we can make a general assessment with respect to the effect of a changing landscape on watershed nitrogen dynamics. First of all, our simulations results indicate that future urbanization of the Ipswich River basin will have a tremendous effect on the nitrogen budget of the Ipswich River basin; from 1991 to 2101 the nitrate loading from upland areas to small streams will double. It is unknown yet how much of these increased loadings will make it to the estuary after being subjected to downstream, riverine processing. Increased nitrogen delivery to the Plum Island Sound estuary will have serious, detrimental effect on estuarine health due to eutrophication. Our simulation results also indicate that within the next few decades, legal protection of forested land will have only a minor effect on controlling nitrate loading due to forest clearing. However, legal protection becomes more important later on, when the rate of development, and thus nitrate loading, rapidly levels off. Overall, our results indicate that we made an important step towards providing decision makers with a simple tool for developing sustainable, integrated watershed management plans by having developed the ability to predict realistic changes in future land-use, coupled with a simple nitrate loading model.