The Impact of Future Land Use Scenarios on Runoff Volumes in the Muskegon River Watershed

Abstract

In this article we compared the response of surface water runoff to a storm event for different rates of urbanization, reforestation and riparian buffer setbacks across forty subwatersheds of the Muskegon River Watershed located in Michigan, USA. We also made these comparisons for several forecasted and one historical land use scenarios (over 140 years). Future land use scenarios to 2040 for forest regrowth, urbanization rates and stream setbacks were developed using the Land Transformation Model (LTM). Historical land use information, from 1900 at 5-year time step intervals, was created using a Backcast land use change model configured using artificial neural network and driven by agriculture and housing census information. We show that (1) controlling the rate of development is the most effective policy option to reduce runoff; (2) establishing setbacks along the mainstem are not as effective as controlling urban growth; (3) reforestation can abate some of the runoff effects from urban growth but not all; (4) land use patterns of the 1970s produced the least amount of runoff in most cases in the Muskegon River Watershed when compared to land use maps from 1900 to 2040; and, (5) future land use patterns here not always lead to increased (worse) runoff than the past. We found that while ten of the subwatersheds contained futures that were worse than any past land use configuration, twenty-five (62.5%) of the subwatersheds produced the greatest amount of runoff in 1900, shortly after the entire watershed was clear-cut. One third (14/40) of the subwatersheds contained the minimum amount of runoff in the 1960s and 1970s, a period when forest amounts were greatest and urban amounts relatively small.

Appendices

Appendix 1: Modeling Land Use Changes

Urban Growth into the Future

Five types of urban land classes (residential, commercial services and institutional, industrial, transportation, extractive and recreation and cemeteries) in the underlying land use vector datasets were collapsed into a single urban land use category for modeling urban changes. Urban growth was projected using six spatially explicit drivers processed using ArcGIS 9.3 and then trained using the LTM following Pijanowski and others (2002a, b, 2005, 2007). The inputs were: distance to county roads, distance to lakes, distance to rivers, distance to existing urban, distance to limited access highways (interstates), and distance to unlimited access to highways (state roads). A separate GIS layer of non-transitioning locations of areas in state and federal forests, all urban categories in 1978, and areas that are in open water, were excluded from the neural network training exercise. Urban expansion rates were based upon research performed by the Michigan Land Resource Project, which modeled the entire state of Michigan at coarse spatial resolution for the years 2020 and 2040 developing growth rate factors of slow, moderate and fast growth based upon observed past changes (Levy 2001).

Urban growth rate represents the ratio of percentage growth of urban area to that of population. The urban area percentage change U, is calculated as:

$$ U = {\frac{{u_{1} - u_{0} }}{{u_{0} }}} $$

(2)

where $u_{1}$ is urban in 1998 and $u_{0}$ is the total urban area in 1978. Population percentage change is calculated similarly as:

$$ P = {\frac{{p_{1} - p_{0} }}{{p_{0} }}} $$

(3)

where $p_{1}$ is population in 1998 and $p_{0}$ is total population in 1978. Thus, the urban growth rate index is the ratio of the two percentages U and P:

$$ rate\, = \,\frac{U}{P} $$

(4)

Urban growth rate based upon observed past growth used a factor of 4.2 for the base (i.e., business as usual) scenario and 2.1 for the slow urban growth scenario (Table 1).

Forest Regrowth Modeling for the Future

Forest regrowth was modeled separately and projected based upon training using 12 spatially-explicit drivers for forest regeneration in the LTM: distance to forests in 1978; density of agriculture at 10 km, 25 km and 100 km; density of forest in 1978 at 10 km, 25 km and 100 km; distance from urban in 1978; distance from agriculture; distance from road; size of non-forest patch; and distance from state parks. In the projections, new forests were allowed to grow where the 1998 land cover was either agricultural land or shrubs. The resultant map represented the propensity for a given cell to undergo forest regrowth. Using the output of the forest LTM we transition cells into the forest land use class at the rate of forest regrowth that was observed between 1978 and 1998 in the land use datasets. The forest regrowth scenario captured the steady abandonment of agricultural lands and their return to forest. The rate was forward projected to determine new forest cells at 5 year intervals. Since wetlands and urban areas are unlikely to change, forest transitions were allowed to only occur in areas that were agriculture or shrub in 1998.

Historical Land Use Maps Generated Using the Backcast Model

The Backcast model (Ray and Pijanowski 2010) uses the land use changes between 1978 and 1998 to model urban “take away” and agricultural and forest transitions over time. Briefly, a back propagation artificial neural network was used to learn about locations of urban, agricultural and forest land use/cover change; the artificial neural network learns about spatial patterns of land use change in relationship to a variety of spatial features such as distance to roads, natural amenities (e.g., lakes) and urban infrastructure. Data from the 2000 Census for year built of the housing statistics was used to fix the quantity in the model for the amount of urban for each MCD and the National Agricultural Summary Statistics amount of land in farms statistic was used to fix the quantity in the model for amount of agriculture historically from 1900 to 1978 for each county. Values for the amount of urban and agriculture between census periods were estimated using a spline. Land use maps generated between 1900 and 1975 were used at 5 year time intervals consistent with the forward LTM projections. The Backcast Model was validated against >12,000 points of historical land use visually interpreted from aerial photographs from 1939 to 1976 to evaluate its accuracy on a decadal time scale (Ray and Pijanowski 2010). The forecast output of the Land Transformation Model was evaluated using the techniques outlined in Pijanowski (2006) and Pontius and others (2008).

Appendix 2: Modeling Surface Water Runoff

The soil conservation service runoff equation is expressed as:

$$ Q\, = \,{\frac{{\left( {P - I_{\rm a} } \right)^{2} }}{{\left( {P - I_{\rm a} } \right) + S}}} $$

(5)

where $Q$ is the runoff in inches, $P$ is the rainfall in inches, $I_{\rm a}$ is the initial abstraction, capturing all loses before runoff begins and S is the maximum retention after runoff begins. Substituting:

$$ I_{\rm a} \, = \,0.2S $$

(6)

gives:

$$ Q\, = \,{\frac{{\left( {P - 0.2S} \right)^{2} }}{{\left( {P + 0.8S} \right)}}} $$

(7)

where:

$$ S\, = \,\frac{1000}{CN} - 10 $$

(8)

and CN is the curve number. The curve number in turn is determined from four watershed characteristics: (1) hydrologic soil group; (2) land use class; (3) hydrologic surface condition of native pastures and; (4) antecedent moisture condition. For the Muskegon River Watershed, the above parameters were correspondingly derived from: (1) STATSGO Soils from the Natural Resources Conservation Service (2008); (2) 1998 land use/cover and projected (past as well as future) land cover; (3) assumed to be ‘fair’ (i.e., assumed to be class ‘C’ habitat); (4) assumed as antecedent moisture condition 2 (AMC2).

Hydrologic soil groups were obtained from the STATSGO database. STATSGO classifies all soils into four hydrologic soil groupings (HSG). HSGs are classified based upon their infiltration rates into A, B, C and D groups. For this watershed area, we observed all four soil groups as present in the soil survey map, with 58% of the watershed in class A, 15% in class B, 13% in class C and 12% in class D. The vector polygons for soil regions were rasterized as the resolution of the land cover timesteps as an integer raster grid with hydrologic soil groups mapped as follows: A = 1, B = 10, C = 100, D = 1000 creating a soil classification grid. The land cover projection for a given scenario was multiplied by the soil classification grid to generate a map with unique combinations of both hydrologic soil type and land cover.

This combined map was reclassified based on land cover/soil combinations shown in Table 4 to create a raster grid of curve number values for each raster location. Table 4 is based on Table 2-2a in Technical Release 55 (USDA 1986) that presents simplified procedures for estimating runoff and peak discharges in small watersheds, with the exception of curve number for the urban land cover type. We fixed urban curve number for all soil types at that of soil type D, (pers. comm. B. Engel, who has studied soil compaction in urban areas and estimates them to be of type D for runoff modeling). For each of the 40 subwatersheds in the Muskegon River Watershed an area weighted average curve number value was computed. The area weighted curve number for each of the 40 subwatersheds was calculated across 100 policy maps, resulting in 400 curve number summarizations for the forward projections and 600 curve number summarizations for the 15 backcast maps across 40 subwatersheds. Area weighted curve numbers were input into Eq. 7 with a rainfall amount of five inches to simulate the runoff for such a storm event.

Table 4 Curve number (CN) values for seven land use classes, across four hydrologic soil types for the MRW, based on TR-55