The Effects of Spatial Resolution on Impervious Cover Classification in Watersheds and Riparian Zones in Vermont

Impervious cover (roads, rooftops, etc.) is a known stressor on stream biota and habitat and is often used as an indicator for assessing the effects of urbanization on stream health. Understanding how spatial data resolution impacts estimates of impervious cover is important for effective modeling and management of water resources at multiple scales. However, broad scale classifications of high spatial resolution data can be time consuming and expensive. High spatial resolution data classifications in Vermont, USA were compared to nationally available impervious cover classifications in order to understand the impact of scale on impervious cover estimates. I used National Agriculture Imagery Program (NAIP) imagery, a Normalized Difference Vegetation Index, ancillary road data, and a supervised evolutionary algorithm classification program to extract and quantify impervious cover for 888 catchments in Vermont, USA. Post-classification accuracy assessments were conducted to quantify the accuracy of the data set. Impervious cover characterized from NAIP imagery ranged from 1.83-48.31% in the study catchments. Overall accuracy for the NAIP impervious cover classifications was consistently high, ranging from 93-99%. National Land Cover Database (NLCD) data showed a bias towards overestimating impervious cover in more developed catchments and underestimating impervious cover in less developed catchments. The high spatial resolution dataset characterized from NAIP data was used to develop a Bayesian classification and regression tree model to predict where the NLCD may be adequate for classifying impervious cover and where higher spatial resolution data may be needed. Data inputs included NLCD land use/land cover classifications and U.S. Census Bureau housing data. High spatial resolution impervious cover was best predicted in catchments with less than 55-65% NLCD forested land cover. For catchments with greater than 55-65% NLCD forested land cover, impervious cover was best predicted where higher levels of NLCD open space development land cover existed. In areas where a full watershed analysis may not be feasible, the condition of the riparian zone along stream channels can provide information on water quality. Impervious cover in the riparian zone was calculated using both fixed-width and elevation based buffer metrics. Percent impervious cover was obtained from the National Land Cover Database (NLCD) and compared to the high resolution imagery analysis from National Agriculture Imagery Program data within the buffer zones. Percent impervious cover ranged from 1.58-8.67% within both types of buffers. The spatial resolution of impervious cover data had less of an effect in the riparian zone than in the full catchments within the same HUC 10 and HUC 12 units. Buffer type had minimal impact on percent impervious cover, except in areas of unconfined valleys, where there were notable differences between fixed-width and elevation based buffers. The results suggest that although there is a trend toward the NLCD underestimating impervious cover at lower levels of development, it may be adequate for mapping impervious cover in the riparian zone, depending on the land use/land cover characteristics of the catchments being studied.


Introduction
Nonpoint source pollution is one of the leading causes of water pollution in the United States (US EPA 2002) and the relationship between land use/land cover, water pollution, and stream ecosystem health is well documented (Harding et al 1998;Allan 2004;Walsh et al 2005;Johnson and Host 2010;Zhang et al 2013). The extent of impervious cover (IC) in a catchment is a known stressor on stream biota and habitat and has been used as a metric for estimating the effects of urbanization on stream health (Schueler 1994(Schueler , 2003Arnold and Gibbons 1996;Brabec 2002;Morse et al 2003;Roy and Shuster 2009;Schueler et al 2009;Arnold et al 2010). Urban development drastically alters the landscape and affects stream ecosystem health by changing the magnitude and timing of stormwater runoff, leading to: higher peak stream discharges, less infiltration of stormwater, changes in channel morphology, increases in nutrient and pollutant inputs, decreases in biotic richness and subsequent increases in pollutant tolerant species, and altered thermal regimes (Wang and Kanehl 2003;Walsh et al 2005;Nelson and Palmer 2007).
The effects of combined anthropogenic stressors from urbanization on freshwater streams can be difficult to quantify. IC as an indicator metric can help to quantify these effects on stream biota and habitat (Schueler 1994;Arnold and Gibbons 1996;Schueler and Claytor 1997;Brabec 2002;Schueler et al 2009). The extent of IC in a catchment has been used to guide stormwater management practices, to develop stormwater management utility rates and credits, and to develop an impervious cover Fine-scale mapping of IC at broad spatial extents presents unique challenges.
Site-specific estimates of IC are often derived from digitized aerial imagery and the  (Homer et al 2015).
A 10-12% IC threshold historically has been associated with quantifiable impacts to water quality at the catchment scale (Schueler 2003, Schueler et al 2009 the University of Vermont. The data and methodologies described will benefit both water quality modelers and decision makers in meeting challenging water resource and policy goals, particularly at the urban-rural interface.

Study area and site selection
The study area focused on four HUC 10 catchments and eighteen HUC 12 catchments in the state of Vermont and their surrounding Municipal Separate Storm Sewer System areas (MS4s) (Figure 1-1 (Figure 1-2).

Data Preparation
We acquired 1m resolution digital imagery from 2011 NAIP data, which served as the base data for mapping IC. NAIP imagery was selected for use because it is readily available at a national scale at frequent temporal intervals, has a fine spatial resolution, supports red, green, blue (RGB), and near-infrared (NIR) spectral bands, is low or no-cost, and it has been acquired using consistent standardized United States where NIR = the value of the near-infrared band and VIR= the value of the visible infrared band.

Impervious Cover Classifications
IC data were extracted utilizing GeniePro2.4 (Observera, Inc. Los Alamos, NM). GeniePro uses genetic algorithms and supervised classification to analyze multispectral remote sensing imagery. In addition to spectral input vectors, GeniePro also incorporates spatial relationships (texture, shape, proximity) and evolutionary algorithms to classify imagery (Harvey et al 2002 However, a multiple extents accuracy assessment of NLCD land use and land cover suggests that NLCD may be accurate for spatial extents as small as 10km 2 , particularly for predominant land use classes or those with unique spectral signatures (Hollister et al 2004). Fourteen NHDPlus catchments were eliminated from the analysis due to very small catchments for which characterization was limited by the resolution of the 30m NLCD data (i.e. one or two pixels caused a large variation in the percent IC), for a total of 888 catchments (Figure 1-3). These fourteen outlier catchments were less than 1 km 2 , well below the recommended threshold for analysis of 30m NLCD data (USGS 2012).

Post-Classification Accuracy Assessment
IC characterization was checked at various stages. Initial IC characterization of preliminary data sets in GeniePro was checked visually to confirm that mapping methods were accurately delineating separate land cover features. A post-classification accuracy assessment was performed for the 1m NAIP data utilizing the methods of Stehman and Czaplewski (1998). Post-classification accuracy assessments were conducted at the HUC 10 and HUC 12 scales. Reference data included the original NAIP imagery and Google Earth historical imagery. Stratified random sampling of the reference data was performed with 100 points per class and compared to the mapped land cover classifications. Overall, user's and producer's accuracies were calculated to describe the accuracies of the IC classifications. Overall accuracy describes the number of correctly identified pixels divided by the total number of classified pixels.
Producer's accuracy measures the ratio between the number of correctly classified pixels and the total number of reference pixels for that category (probability that IC on the ground is correctly classified). User's accuracy measures the ratio of correctly classified pixels to the total number of pixels in that category (probability that a pixel labeled as IC is actually IC).

Statistical Analysis
A regression analysis was run to identify the relationship between NLCD and NAIP data. The ratio of NLCD to NAIP IC was then compared to the NAIP data within the NHDPlus catchments to identify where the NLCD data are adequate and where they may be over-or underestimating IC. Classification and Regression Tree (CART) analyses were run using Systat 13.1 (Systat Software, Inc. San Jose, CA) to quantitatively assess the level at which the NLCD under-or overestimates IC. CART analysis compares independent and dependent variables through a series of binary splits (Breiman et al 1998). CART was run for all NHD catchments in the study area using the least absolute deviation option for choosing splits with bootstrapping (sample of 800 repeated 1000 times, maximum=2 splits, p=0.05 stopping rule). CART was rerun for the 41 catchments larger than 10km 2 in the study area using the least absolute deviation option for choosing splits with bootstrapping (sample of 38 repeated 1000 times, maximum=2 splits, p=0.05 stopping rule)  The ratio of NLCD IC to NAIP IC was plotted against NAIP IC (Figure 1-5).

Post-Classification Accuracy Assessment
An idealized relationship would yield a slope of 0 across all levels of IC (ratio=1), indicating no difference between the two classifications. The regression equation for the NLCD to NAIP ratio resulted in a non-linear relationship, with NLCD underestimating at low levels of IC and overestimating at higher levels of IC ( Figure   1-5): r 2 = 0.71 (n=22, p<0.001).
A comparison of the ratio of the UVM to NAIP data shows a ratio closer to 1 across all levels of IC, resulting in a slope closer to 0, with 95% confidence intervals between

Statistical Analysis
To further elucidate the relationship between the high resolution estimates and the NLCD data across a range of IC at a finer scale, IC classifications were compared across NHDPlus catchments. IC in NHDPlus catchments ranged from 0 -48.31% ( Figure 1-3). With more sample points, it is clear that the NLCD is underestimating IC at low levels of development, although there is a wider range of NLCD to NAIP ratios at lower levels of IC, suggesting greater uncertainty in the data. Across all NHDPlus catchments, the subset of classification trees with a single split had a median cut value of 0.77 for high resolution IC (95% confidence interval 0.18 -3.59), with average node medians of 0.10 and 0.61 for the ratio of NLCD to NAIP IC and an average 16.2% reduction in error (Figure 1-3). The median cut value shows that the data shifted at 0.77% IC, with an average NLCD to NAIP ratio of 0.10 below that level of IC and an average ratio of 0.61 above that level.
For NHDPlus catchments larger than 10km 2 , the subset of classification trees with a single split had a median cut value of 2.09 for high spatial resolution IC (95% confidence interval 1.36-7.40), with average node medians of 0.41 and 0.83 for the ratio of NLCD to NAIP IC and an average 25.3% reduction in error (Figure 1-6). The median cut value shows that the data shifted at 2.09% IC, with an average NLCD to NAIP ratio of 0.41 below 2.09% IC and 0.83 above. Even with the smaller sample size, the data follow a similar trend of underestimating IC by 20-80% at low levels of development, with less noise in the data than in the sample containing all NHD catchments (Figure 1-6). There were too few samples with greater than 10% IC to identify trends in the data at higher levels of development.

Discussion
Classification of IC from high resolution NAIP data produced a highly accurate data set for 18 HUC 12 units and 4 HUC 10 units in Vermont. Accuracy levels were quite high in all catchments and supported by a comparison to the UVM high spatial resolution dataset from the same 2011 NAIP data. Producer's accuracy was expected to be very high due to the fact that most pixels within the catchments are not impervious, thus the chance of correctly identifying a pixel as not impervious is somewhat biased. The user's accuracies were expected to be somewhat lower than the producer's accuracies but were still high, and met or exceeded the USGS standard for accuracy assessment. There are several possible sources of error in the data. First, differences in spatial resolution and methodologies may impact the classification of IC based on the format of the source data and not misclassification (Hollister et al 2004;Loveland et al 2005).
NLCD IC was characterized from 30m spatial resolution raster data while the NAIP data were characterized as rasters from 1m digital orthophotographs. The UVM dataset was classified as a vector and then rasterized for the purposes of this study. All raster data exhibit some degree of "stair stepping" effects and the comparison and conversion of data from different formats may introduce some error. Second, the reference data used in the accuracy assessments included the source data as well as historical imagery from Google Earth. Congalton and Green (2008) do not recommend using the source data as reference data as it may introduce additional error. However, the imagery was not analyzed until several years after it was taken, thus it was impossible to set up an accurate system for ground referencing in the field. Last, the analysis of the NHDPlus catchment data includes catchments that are smaller than the spatial extent recommended by the USGS (USGS 2012). However, removal of catchments less than 10km 2 results in a similar trend of underestimating IC at low levels of development by 20-80%.
We developed a cost-effective method for characterizing IC that is comparable to other high spatial resolution datasets produced with object-oriented classification and is easily reproducible with minimal GIS resources. While other methodologies require expensive and specialized software to delineate IC over broad scales, this method utilizes readily available NAIP imagery, E911 road data, GIS software, and a relatively inexpensive supervised classification program to produce highly accurate results. Through a comparison of NLCD and high spatial resolution NAIP IC classifications, we determined that NLCD data are underestimating IC in less developed catchments and overestimating IC in more urbanized catchments. The implications of this study for modeling stream response to urbanization may be significant. Future research will explore the landscape patterns and processes contributing to the discrepancy of classification accuracies across the urbanization gradient.   Figure 1-3. Ratio of medium to high spatial resolution IC classifications at the NHDPlus catchment scale over a range of % IC classified from NAIP data. Outliers removed from analysis are shown on the secondary y-axis. Solid line represents a perfect ratio of NLCD to NAIP data with slope = 0. Dashed lines represent CART average node medians of 0.10 and 0.61 for the ratio of NLCD to NAIP IC. Median cut value = 0.77% IC Figure 1-4. Relationship of NLCD to NAIP data at the HUC 10 and HUC 12 scale. Shaded area shows 95% confidence interval. Solid line through origin shows 1:1 ratio.
Figure 1-5. Ratio of medium to high spatial resolution IC classifications at the HUC 10 and HUC 12 scale over a range of % IC classified from NAIP data.
Figure 1-6. Ratio of medium to high spatial resolution IC classifications in NHDPlus catchments greater than 10 km 2 . Solid line represents a perfect ratio of NLCD to NAIP data with slope = 0. Dashed lines represent CART average node medians of 0.41 and 0.83 for the ratio of NLCD to NAIP IC.

Resolution Impervious Cover Data Requirements for Watershed Management"
by Jessica Morgan 1,2* ; Naomi Detenbeck 3 ; Yeqiao Wang 2 is submitted to the journal Environmental Management

Introduction
Impervious cover (IC) is a known stressor on stream biota and habitat and has been used as a metric for measuring the effects of urbanization on stream health (Arnold et al. 2010;Arnold and Gibbons 1996;Brabec 2002;Morse et al. 2003;Roy and Shuster 2009;Schueler 1994;Schueler 2003;Schueler et al. 2009). As IC in a watershed increases, streams produce higher peak discharges, less infiltration of stormwater occurs, channel morphology changes, pollutant inputs increase, biotic richness decreases, pollutant tolerant species increase, and thermal regimes are altered (Nelson and Palmer 2007;Walsh et al. 2005;Wang and Kanehl 2003). Biological monitoring and assessment are often used by water quality programs to monitor changes in stream systems (Johnson and Host 2010). Indicator metrics including species richness, species composition, relative abundance, feeding relationships, body size, and others have been utilized by state agencies to assess goals set by the Federal Clean Water Act (Bellucci et al. 2013;Karr 1993). Changes to these metrics, combined with landscape information from Geographic Information Systems (GIS), can provide a detailed response of biota to urbanization impacts, track environmental changes through time, and be used to forecast the effects of future land use scenarios (Richards et al. 1997 (2010) found taxon-specific change points of declining macroinvertebrates at 0.81-3.3% developed land. Declines in community metrics have been found at 1-2% IC for macroinvertebrates and as low as 0.6% IC for diatom communities Smucker et al. 2013). Others have suggested that landscape patterns may explain varying levels of water quality in catchments with similar amounts of IC (Beck et al. 2016).
Medium resolution spatial data such as the National Land Cover Database (NLCD) have traditionally been used for regional analyses. watershed when compared to high spatial resolution land use data (Stueve et al. 2015).
A national assessment of the 2001 NLCD canopy cover and IC data sets demonstrated that the NLCD underestimated both canopy cover and IC when compared to photographic interpreted imagery (Nowak and Greenfield 2010).
High spatial resolution predictions of IC are needed for effective modeling and management of water resources at the watershed scale (Zhou et al. 2010;Zhou et al. 2014). However, broad scale classifications of high spatial resolution data can be time consuming and expensive. In addition to errors due to coarse spatial resolution, errors relating landscape characteristics to biotic responses in streams also occur due to the inability to match the temporal scale of landscape change with responses. The NCLD is updated every 5-10 years, so an exact match with the year of response measurements is rarely possible. Some responses to land-use change have built-in lags, so that simply matching the year of land-use with the year of response variables will be insufficient to describe the most precise relationship possible (Harding et al. 1998). Thus, new estimates of high spatial resolution data are needed, as well as the ability to evaluate time series data.
The ability to predict where high spatial resolution data are required based on derived relationships from readily available medium resolution data would decrease the cost and resources required for broad scale classifications of IC estimates. While many values match for both NLCD and high spatial resolution IC classifications, there are many points where there is considerable scatter along the best fit line (Figure 2-1).
A methodology is needed to enable decision makers to prioritize which areas require high spatial resolution estimates and identify areas where the NLCD are adequate for analysis. Communities would realize a cost-savings by eliminating the purchase and classification of high spatial resolution data in areas where NLCD are already adequate.
Here, we propose a Bayesian classification and regression tree (BCART) model to predict where the high spatial resolution value of IC can be predicted by NLCD data, based on the readily available medium resolution data available from the NLCD and U.S. Census Bureau. We demonstrate that high spatial resolution IC data will be more easily predicted from the NLCD estimates in less forested areas, as trees and vegetation can mask houses and roads (Claggett et al. 2013;Roy et al. 2003

Study Area
The study focused on 902 National Hydrography Dataset Plus (NHDPlus) catchments in Vermont in areas where high spatial resolution IC data were characterized from 1m National Agriculture Imagery Program imagery ( The USGS cautions against using NLCD data in watersheds of less than tens of square miles (USGS 2012). However, a multiple extents accuracy assessment of NLCD land use and land cover suggests that NLCD may be accurate for spatial extents as small as 10km2, particularly for predominant land use classes or those with unique spectral signatures (Hollister et al. 2004). Fourteen NHDPlus catchments (Appendix 1) were eliminated from the analysis due to very small catchments for which characterization was limited by the resolution of the 30m NLCD data (i.e. one or two pixels caused a large variation in the percent imperviousness). All outlier catchments were less than 1 km2, well below the recommended threshold for analysis of 30m NLCD data (USGS 2012). The remaining 888 NHDPlus catchments ranged in size from 0.001 to 55.2 km2.

Bayesian Classification and Regression Tree (BCART) Methods for Assessing NLCD Adequacy
Classification trees can be useful in identifying the predictive structure of a problem by determining which variables drive a phenomenon or process and have previously been utilized in characterizing urban environments (Torbick and Corbiere 2015). As data sets increase in complexity, the relationships between variables can vary across measurement space, resulting in non-homogeneity (Breiman et al. 1984).
Classification and regression trees compare independent and dependent variables through a series of binary splits and then fit a separate model within subsets of the data to identify relationships that may vary across a dataset (Chipman et al. 2002). Thus, the model structure itself, rather than just the data, is homogenous at each terminal node split (Chipman et al. 2002). Previous classification trees relied on growing large classification trees and then "pruning" them back to identify terminal node means or class probabilities. More recent Bayesian approaches utilize a prior distribution to produce smaller trees at the outset, with linear regression models at each terminal node. Prior probability distributions for a variety of models have been evaluated and default choices suggested to avoid the overfitting of models and to shrink bottom node parameters, thus stabilizing the estimation (Chipman and McCulloch 2000;Chipman et al. 2002). Chipman et al. (2002) developed an algorithm for estimating the posterior distribution that prevents the model from getting stuck on a local solution, incorporates both grow/prune steps and swap/change steps to improve tree structure, and utilizes multiple restarts to create a more diverse set of trees. The final nodes of the BCART analysis are assumed to be parametric, thus the data must meet assumptions of homoscedasticity, a linear relationship between the variables (for a linear model only), and normality of the error distribution.

Modeling differences between NLCD and NAIP data
A Bayesian CART model was developed to predict where differences are likely to occur between high spatial resolution IC data and NLCD data. The ultimate goal was to predict where NLCD data are sufficient and where higher spatial resolution data are needed. End node regression models were constructed to identify variables with good predictive ability for high spatial resolution IC data in specific landscapes determined by the classification portion of the analysis. Landscape patterns with poorer fitting end node models may require higher resolution IC data than those with well-fitting end node models. It is possible that a single predictor (or more) with a robust enough relationship with the dependent variable could yield sufficient predictive power. Factors considered in the model included the effects of historical land use/land cover, housing unit age, and current land use/land cover including measurements of canopy cover, agriculture, development intensity, and catchment size.

Effects of historical land use/land cover
To explore the effects of historical land use/land cover on the accuracy of IC data, land use/land cover data were allocated to the NHDPlusV1 catchments, using the While the total developed for the two datasets matched, the developed "from-to" categories did not match in total developed from forested and agricultural patches in extent or spatial distribution, above and beyond what might be expected from the six year difference in the data sets. A NLCD 1992/2001 retrofit land cover change program is available for comparison to the 1992 NLCD classifications, but it is only recommended for use at a regional scale, thus it was not included in the analysis (Fry et al. 2009). Due to such a major inconsistency between the NLCD and CCAP datasets, we decided to use the background NLCD landscape as a classification variable instead, assuming that in areas of agriculture, development was from agricultural parcels and in areas of forest, development was from a forested landscape. hypothesis testing, and reliance on a single best model (Freckleton 2011;Whittingham et al. 2006). To reduce the level of multiple hypothesis testing, multicollinearity between variables was identified using a variance inflation factor, with a factor of less than 10 indicating independence (Philippi 1993  which includes an extra parameter to account for more variance in the data. However, the BCART program cannot accommodate a quasibinomial fit, so the arcsine transformed data was used for the BCART analysis (Chipman and McCulloch 2002).

BCART Analysis
Data analyses were run for both the full data set (n=888 catchments) and the catchments greater than 10 km 2 (n=41). BCART was first run for the full data set with 100,000 iterations and one restart to identify the s 2 of the initial pooled estimate of the model (s 2 = 0.001903) and where the log likelihood stabilized (8,000 iterations).
BCART was then run once each with the prior probabilities suggested by Chipman and McCulloch (2000) with 20 restarts to make the convergence more efficient and to prevent the model from getting stuck at a local solution (Chipman et al. 2002). The tree size with the largest log likelihood (least negative) value for the most visited solution was chosen (Chipman and McCulloch 2000). All features were standardized to have zero mean and unit variance. The procedure was repeated for the catchments greater than 10km 2 , resulting in an s 2 of 0.002562 and a stabilized log likelihood at 1,000 iterations. Cross validations were performed ten times with a random sampling of 10% of the observed data. Residuals were plotted against the predicted data to check for homogeneous variance and against original predictors to check for potential non-linearities. The residuals were homogeneous, suggesting that the assumption of homoscedasticity was met and there were no non-linearities with the original predictors, indicating that second and third order terms were not necessary. Final regression models for each node were re-run using the MASS package in R (Venables and Ripley 2002) to determine the significant regression coefficients. The R code for all analyses is listed in Appendix 2.

Study Area Catchments
The best BCART model for the full data set was found with prior probability parameters of 0.5, 2, 1, and 0.404*s 2 =0.000769, with a stabilized log likelihood at 8,000 iterations and 20 restarts. The classification variables that best partitioned the data included NLCD11Forest (split point between 55-65% throughout the crossvalidations) and open space development (OSD) (split point of ~1.2%) (Figure 2-3).
These classification variables were consistent through the ten cross-validation runs, with catchment size (split point 0.005-0.01 km2) occasionally occurring as an additional classification variable. A generalized best fit model for the full data set is shown in Figure 2-3 and the specific model for cross validation number 5 is shown in Final regression models explained 92% of the variance for Node 1, 69% of the variance for Node 2, and 79% of the variance for Node 3 ( Table 2-2). Because the data were standardized, the magnitude of regression coefficients can be compared to determine the relative importance of each (Chipman et al. 2002). The most important predictor for all three nodes was the NLCD11 Imp ( Figure 2-7). For Node 1, other moderating variables, in descending order of magnitude, included PtCanAnaly and NLCD11Ag. Node 2 had one moderating variable -NLCD11Ag. Moderating variables for Node 3 included PtCanAnaly, WtAvgPre40, and WtAvgHouseAge in descending order of importance. Test data for each cross validation run showed a good fit with the model with an r 2 of 0.93 for 890 test fits (Figure 2-8).
The root mean square error (RMSE) provides a measure of the predictive power of a model, with smaller numbers reflecting a better model fit. The RMSE for Nodes 1-3 was 1.01, 1.02, and 1.01 respectively. While it appears that all three nodes performed equally well, the RMSE represents a much larger proportion of the predicted value for Node 2 (<1.2% OSD), suggesting that it has less predictive power than the other two nodes.

Catchments Larger than Minimum Recommended Areas
The best BCART model for the 41 larger catchments was with prior probability parameters of 0.5, 2, 1, and 0.404*s 2 =0.001035, with a stabilized log likelihood at 1,000 iterations and 20 restarts. One classification variable (NLCD11Ag) best partitioned the data with a split point of 58.24%, splitting off only 1 catchment from the rest (Figure 2-9). This was consistent through the ten cross-validation runs. A generalized description of Node 1 and an example catchment is shown in Figure 2 Node 2 could not be evaluated due to the small sample size. Test data for each cross validation run showed a moderate fit with the model with an r 2 of 0.79 for 40 test fits ( Figure 2-12).

Study Area Catchments
Urbanized areas generally have less canopy cover than suburban or rural areas.
Thus, it was suspected that high spatial resolution IC data should be more easily predicted from the NLCD estimates in urban areas than in less developed areas, where trees and vegetation can mask houses and roads (Claggett et al. 2013;Roy et al. 2003 (Figure 2-7). This is further supported by the adjusted r 2 values for the end node regressions ( Table 2-2). Node 1 (less forested) with an adjusted r 2 = 0.92, suggests a well-fit model. Node 2 (more forested, low OSD) shows much less predictive power (r 2 =0.69) and Node 3 (more forested, higher OSD) shows moderate predictive power (r 2 =0.79). It is likely that NLCD best predicts the high resolution impervious data in less forested watersheds because there is less canopy cover to complicate the classification process. In forested watersheds with low levels of OSD, it may be that since NLCD "burns in" road networks, which account for the majority of IC in rural areas, 30m pixels may over-represent smaller rural roads, leading to poor predictive power for the model. Medium and larger roads may be adequately represented by the 30m NLCD, leading to better predictive power for the model.
Overall, the data suggest that NLCD IC data may be adequate in catchments with less than 55-65% forested land cover but may require moderating variables in more heavily forested catchments, and likely requires higher resolution IC estimates in heavily forested catchments with low levels of development.
The catchments with less than 55-65% forested cover best predicted high resolution IC values using canopy cover, agricultural land use, and percent medium resolution IC (Node 1, Figure 2-4). Canopy cover, agriculture, and NLCD IC all had positive predictive relationships with high spatial resolution IC estimates in Node 1 ( Figure 2-7). This further supports the hypothesis that the high resolution IC classifications are more easily predicted in catchments with less forested land cover.
In contrast, catchments with greater than 55-65% forested cover were further split  (Figure 2-7). This makes sense as developments from agricultural land use are less likely to contain canopy cover that can obscure IC.
However, as noted above, overall the end node model for Node 2 had less predictive power (r 2 =0.69), thus it is likely that higher resolution estimates of IC would be required in these catchments.
Catchments with greater than 1.2% open space development best predicted the high spatial resolution IC values using canopy cover, medium resolution IC data, and housing age (Node 3, Figure 2-4). Both heavily forested areas and residential development age had a negative relationship with the ability to predict high spatial resolution IC data, with the proportion of developments built before 1940 being slightly more important than the average housing age (Figure 2-7

Catchments Larger than Minimum Recommended Area
The 41 catchments larger than the minimum recommended area for analysis (Hollister et al. 2004;USGS 2012) (Figure 2-11). HRImp was also weakly negatively correlated with both housing variables and NLCD11Ag. Thus, as percent agriculture and housing age increased, predictive power for the model decreased. This is consistent with the full data set for the housing metrics, but the opposite of the results for agriculture. This may be because there were only 9 catchments with less than 55% NLCD forested in the 41 catchment data set, which is the primary classification variable for the full data set analysis. This essentially eliminated the sample size for less forested areas in the 41 catchment analysis. It may be that the negative relationship to agriculture is simply a response to a lack of less forested catchments available for analysis.

Limitations of the Approach
There are several limitations to the approach utilized in this study. First, the high resolution data were classified from NAIP imagery taken during the growing season.
While the benefits to NAIP imagery include that it is free/low cost, available at regular intervals, and contains a near-infrared band for improving IC classifications, the cost is that it is "leaf on" imagery can make classifying IC more difficult in forested areas.
This may have contributed to the importance of the NLCD forested land use class in the classification portion of the analysis. However, the cost-effective benefits of using NAIP imagery, combined with using ancillary data such as road datasets, likely outweigh the cost of using a "leaf on" data set.
Second, the analysis included catchments smaller than those recommended by the USGS (Hollister et al. 2004;USGS 2012). However, out of 888 catchments available for analysis, only 41 catchments were greater than the 10km2 recommended by Hollister et al. (2004) and only 1 was larger than the "tens of square miles" (20mi2=51.8km2) threshold suggested by the USGS (2012). Further, tools such as the NHDPlus Ca3T were developed for allocating land use attributes to NHDPlus Catchments and are routinely used for regional analyses, even when the majority of catchments do not meet the minimum mapping unit. The USGS recommends that any use of the data in watersheds of less than tens of square miles should be examined at the local extent to determine its appropriateness for analysis. Our BCART model provides a method for determining where the NLCD may be adequate and where higher resolution data may be needed.
Last, the small sample size for 41 catchments that met the minimum mapping unit essentially removed NLCD11 Forest from the analysis. This, combined with using only 10% (4 catchments) per cross-validation analysis, may have impacted the accuracy of the minimum mapping unit analysis. Further study is required to determine the impact of catchment size on the adequacy of NLCD IC classifications.

Management Implications
The ability to determine where NLCD data are adequate and where higher Our model provides a quick and cost-effective method for assessing the utility of NLCD IC data over broad spatial extents. The data and methodologies described will benefit municipalities and watershed managers by providing information on where NLCD predicts high resolution IC data accurately and where higher resolution data may be needed. This information can then be used to determine where limited monetary and geospatial resources can be spent to best mitigate the impacts of urbanization on water quality.   BCART example best fit model output for catchments greater than 10 km 2 . Numbers indicate number of catchments, listed with significant regression variables for each node and adjusted r 2 for node final regressions. Node numbers correspond to Figures 10-11 and Table 2-3. Variables are defined in Table 2

"Impervious Cover in the Riparian Zone -An Analysis of Method and Scale"
by Jessica Morgan 1,2* ; Yeqiao Wang 2 ; Naomi Detenbeck 3 will be submitted to the journal Northeastern Naturalist

Introduction
Riparian areas provide a number of functions for maintaining the integrity of stream ecosystems threatened by encroaching urbanization, including: water temperature regulation, bank stability, sources of organic and inorganic material, energy dissipation, and nutrient, sediment, and contaminant retention (Allan 2004, Chen et al. 1998, Gregory et al. 1991, Naiman and Decamps 1997, Naiman et al. 2005). Further, forested and wetland riparian areas have been shown to help to mitigate the effects of urbanization, even when the natural functions of these zones are altered by stormwater drainage infrastructure (Smucker et al. 2013). Recognizing the key role that buffer areas play in preserving water quality, municipalities across the Northeast have enacted regulations to protect and restore riparian areas using a variety of methods for determining optimal buffer widths.
Riparian buffer zone regulations vary by state and municipality and many analyses for informing stormwater regulations utilize fixed-width buffers. Most commonly, these analyses create 30m-120m wide buffers, based on the 30m pixel width of the National Land Cover Database (NLCD) (Goetz 2006). Previous analyses have defined 120m fixed-width buffer distances along flow paths and summarized land use types (urban, agriculture, etc.) at the riparian and catchment scale , Smucker et al. 2013. Several studies have found that buffer distances smaller than 100m were not accurate predictors of fish or macroinvertebrate populations (Lammert and Allan 1999, Roth et al. 1996, Van Sickle et al. 2004, Vølstad et al. 2003. In another study, a 30m buffer was found to be a weak secondary predictor of stream health, with regional land use undermining the ability of vegetation to maintain high quality habitat (Roth et al. 1996). Fixed-width buffers may underestimate the actual riparian boundary by up to 2.5 times the distance from the stream (Skally and Sagor 2001). Further, it has been suggested that medium spatial resolution data such as the NLCD may not be adequate for mapping riparian zone features (Baker et al. 2007, Fernández et al. 2014, Hollenhorst et al. 2006. Accurate mapping of riparian zone condition from remote sensing data presents several challenges, including the determination of functional buffer width (Baker et al. 2006) and the spatial resolution of elevation (Abood et al. 2012), land use/land cover (Hollenhorst et al. 2006), and stream map data (Baker et al. 2007) .
Fixed-width buffers may not correspond to relevant ecotones and can include areas that are irrelevant to the functions of buffers, causing the misinterpretation of landscape patterns (Baker et al. 2006). Functional riparian metrics have been developed based on the connectivity of source land cover area to the stream channel (Baker et al. 2006) and distance weighting has been used to examine land use/land cover influence on stream health (Van Sickle and Johnson 2008). More recent studies have used elevation and flood height data to delineate riparian buffer zones based on readily available digital elevation models (DEMs) and the 50 year flood plain (Abood et al. 2012, Fernández et al. 2012.
Urban expansion increasingly affects riparian zones, causing hydrologic changes, which then cause changes to soil, vegetation, and the ability to filter pollutants (Groffman et al. 2003). While urbanization and increasing amounts of IC are known stressors on stream ecosystem health (Schueler et al., 2009Walsh et al. 2005, proximal IC, i.e., IC that is adjacent to streams, may disproportionally affect water quality compared to IC distributed throughout the watershed (Wickham et al. 2016 for more accurate estimates of in-stream response to urbanization. The objectives of this study were to 1) determine the effects of spatial resolution on IC estimates in the riparian zone and 2) to compare IC derived from fixed-width and elevation based buffers. Given the smaller spatial area of delineated riparian buffers, it was expected that the NLCD would underestimate proximally distributed IC. Based on the literature, it was expected that fixed-width buffers would underestimate IC when compared to elevation based buffers. Both spatial resolution of IC data and delineation method are critical considerations for accurate modeling of water resources.

Study area
The study area included riparian buffer areas of four Hydrologic Unit Code (HUC) 10 catchments and eighteen HUC 12 catchments in the state of Vermont

Defining the riparian zone
Definitions of the riparian zone are many, varied, and dependent on the resource being evaluated. Some definitions include the aquatic environment while others exclude it, some definitions incorporate soils and hydrology, while others focus on landscape characteristics and geomorphology (Ilhardt et al. 2000). Functional definitions of riparian areas focus on the ecosystem services provided by riparian zones rather than their individual components (Ilhardt et al. 2000, Verry et al. 2004. For the purposes of this study, a functional riparian zone is defined as "a three dimensional space of interaction that includes terrestrial and aquatic ecosystems that extend down into the groundwater, up above the canopy, outward across the floodplain, up the near-slopes that drain to the water, laterally into the terrestrial ecosystem, and along the water course at a variable width." (Verry et al 2004). This study focuses on the riparian zone up to the 50-year flood height.  , Detenbeck et al. 2016). This generated a distance from the NHDPlus flowline, which was then classified into two categories: buffer and non-buffer areas. The raster was then reclassified so that the cells in the buffer were equal to one and the remaining cells were equal to zero. This resulted in a buffer along the flow path of 120m on each side.

Delineating riparian zones
A null conditional was set in ArcGIS Spatial Analyst to make the buffer area =1 and the non-buffered areas=NoData. The bankfull mean depths and associated 50-year floodplain elevations for each HUC 10 and HUC 12 unit are shown in Table 3-1. Bankfull depth was held constant over each HUC 10 and HUC 12.
Ten meter DEMs were chosen for processing as 30m DEMs have been shown to have too coarse a resolution to effectively map riparian zones (Abood et al. 2012, Fernández et al. 2012. DEMs were obtained from the 3D Elevation Program (3DEP) (formerly the National Elevation Dataset) 1/3 arc second (10m) continuous dataset (Gesch et al. 2014). Data were projected into NAD83 Albers using nearest neighbor interpolation because bilinear interpolation altered the scale of the raster and also included more errors than those in the cells calculated by nearest neighbor as determined by visual inspection of the original dataset. Most likely, this is due to bilinear interpolation using a weighted average across the four nearest cells, thus introducing smoothing error in areas with steep elevation gradients. Nearest neighbor, although not recommended for continuous data, calculates each cell value using the nearest cell in the input and does not change the values of cells from the input raster (ESRI 2016). It was necessary to reproject the data prior to processing because the RBDM automatically extracts percent IC by area based on a classified land use land cover dataset. Two HUC 12 units overlapped into New Hampshire. In these cases, the NED data from each state were downloaded and mosaicked together with the mosaic operator set to take the values of the VT raster, where data overlapped. One HUC 12 unit overlapped into Canada so the catchment boundary was clipped at the Canadian border, as the NLCD does not extend into Canada.
The RBDM calculates 360° radial transects from sample points along the stream to calculate elevation data, based on a user specified maximum transect length. Abood et al. (2012) suggest setting a maximum transect length of 250m-500m, while balancing floodplain characteristics and processing time. Preliminary analysis showed setting a 500m transect length had minimal improvement in riparian buffer delineation, resulting in a 0.1% change in IC calculations but greatly increased processing time, so the maximum transect length was set at 250m.

Impervious cover mapping
Percent IC for both the fixed-width and elevation based buffers were extracted from the 30m NLCD 2011 IC data (MRLC 2014) and 1m 2011 high spatial resolution data derived from the National Agriculture Imagery Program (NAIP) for the selected HUCs in Vermont (Morgan et al. in review-a). The extract by mask function of Spatial Analyst was used to extract the NLCD 2011 IC in the buffer layer for both the fixed-width and elevation based buffers. The resulting raster was used to calculate the percent IC within the buffer. Zonal statistics in the Spatial Analyst toolbox were used to calculate NLCD 2011 IC within the buffer area for each HUC 10 and HUC 12.
Because the high resolution data were not contiguous, extract by mask was used to extract only the buffer regions of the high resolution data. Zonal statistics were then run to calculate the mean percent IC for each buffer zone in the HUC 12 and HUC 10 units for the fixed-width buffers. The RBDM tool automatically calculated percent IC from the high spatial resolution data as an optional input. To maintain consistency with the RBDM tool, which removes waterbodies (e.g., lakes) from the final output, waterbodies were removed from the fixed-width buffers as well.

Statistical analysis
A regression analysis was run to identify the relationship between NLCD and NAIP data among HUC 10 and HUC 12 units for both fixed-width and elevation based buffers. The ratio of NLCD to NAIP IC was then compared to the NAIP data within the buffer types to identify where the NLCD data are adequate and where they may be over-or underestimating IC. Last, high spatial resolution IC data were compared among buffer types to identify differences in the total percent IC.

Riparian buffer zones
Elevation buffers generally had greater spatial extents than those generated by the fixed-width method (Figure 3-2). IC characterized from NAIP imagery by both methods ranged from 1.58-8.67% in the buffer zone of HUC 12 units and from 3.79-5.42% in HUC 10 units (Table 3-2). IC characterized from NLCD imagery by both methods ranged from 1.40-8.81% in the HUC 12 units and from 3.07-5.48% in HUC 10 units (Table 3-2). As expected, greater percentages of IC in both types of buffer zones were found in the more urbanized catchments such as those located in Burlington, Rutland, and St. Albans.

Statistical analysis
The NAIP data at the HUC 10 and HUC 12 level predicted NLCD classifications well for both types of buffers, with some noise in the data ( The elevation based buffers showed an outlier (Muddy Brook) that severely overestimated the amount of IC in the buffer area, with a ratio 2. This outlier is not indicated in the fixed-width buffer analysis. With the exception of the Muddy Brook outlier, the NAIP percent IC was comparable between buffer types at lower levels of IC (<5.5%) but showed more scatter at higher levels of IC (Figure 3-7).

Discussion
Choosing the appropriate spatial scale and methodology for delineating riparian buffer zones is critical for defining ecologically relevant riparian buffer regulations. Many municipalities create regulations based on fixed-width buffers because they are easy to define and enforce. Recent advances in delineating riparian zones have moved toward functional definitions and methodologies (Abood et al. 2012, Ilhardt et al. 2000. Although conceptually, it is more difficult to gain public support for functional riparian zones, recent studies suggest that functional riparian zones may be more cost-effective than fixed-width buffers (Tiwari et al. 2016). This is particularly true in areas of wetlands and low production forest.
Surprisingly, spatial resolution had less of an effect on IC estimates in the riparian zone than in full catchments at the same locations (Morgan et al. in review-a).
The NLCD IC estimates still showed a trend of underestimating IC at low levels of development but it was not as strong as at the catchment level. It is possible that with more data points, the relationship would be significant. We expected that there would be more variation at the riparian scale due to the limited spatial extent and more opportunities for small inaccuracies to create large differences between the NLCD and NAIP data. This was not the case as there was no major change to the ratio for either NLCD or NAIP data.
As land use/land cover type can vary across a watershed, it is possible that differential use within the watershed impacted IC classifications more than the spatial resolution of the data, as the type of land use/land cover may influence the accuracy of the NLCD percent IC classifications (Morgan et al. in review-b      Figure 3-4. Relationship of NLCD to NAIP data at the HUC 10 and HUC 12 scale for elevation based riparian buffers. Solid line through origin shows 1:1 ratio. Figure 3-5. Ratio of medium to high spatial resolution IC classifications for fixed width riparian buffers. Solid line represents a perfect ratio of NLCD to NAIP data with slope=0.
Figure 3-6. Ratio of medium to high spatial resolution IC classifications for elevation based riparian buffers. Solid line represents a perfect ratio of NLCD to NAIP data with slope=0. Figure 3-7. Comparison of fixed width and elevation based buffer IC classifications for high spatial resolution NAIP data. Solid line through origin shows 1:1 ratio. Outliers are labeled by catchment name.

Muddy Brook
Moon Brook Appendix 2. R code Used in Analysis for Chapter 2.