Geospatial ass essment of riparian zzones: A case study in ......ords: Ripari dation tial ass son Ri r.edu, PhD esf.edu, A C ion and prot ey knowledg s in New Yo and quanti sub-watersh

G

AbstRipaUnforiparbuffeour mclassdeterthat t15–2persiand cbut a Thre

•

•

•

Keywdegra

GeospaHud

1gpu100@sy2ljquack@

tract arian restoratortunately, krian corridorfer vegetationmethod in a sifying vegetrmine spatiathere were m

22% of buffeistently degrcommunity also enable t

ee Summary • We found

over time• Compare

products while ove

• Results ovegetatiorestoratio

words: Ripariadation

tial assdson Ri

yr.edu, PhD

@esf.edu, AC

tion and protkey knowledgrs in New Yon and quantisub-watershtation and no

al trends in bmore areas oer areas haveraded across groups to nohem to asses

Points of Intd decreasinge. However, ed to riparianbased on 30erestimating

of cluster anaon degradatioon efforts on

ian Buffer, R

NWD 2Itht

essmenver Est

Candidate, DEnviron

Associate ProCollege of En

tection provge gaps exisork State. Thfy the spatio

hed of the Huon-vegetatiouffer vegetaf vegetation e clusters of the time spa

ot only rapidss the spatio

terest g trends of vethe 120 m b

n buffers def0 m satellite g agriculture alysis providon. Such info

n the ground.

Remote Sensin

NEW YWATEDepartment of Biol

30 Riley-Robb Hathaca, NY 14853-ttp://wri.cals.corne

nt of riptuary –

Ge Pu1, Lin

Department nmental Scieofessor, Depanvironmenta

ide key benest in terms ofhe primary ootemporal paudson River-on regions wation for 200

than non-vevegetation d

an studied. Udly delineate temporal pat

egetation cobuffer showefined from 1 images undeby 10–14%

ded further information can.

ng, Spatial An

YORK ER RES

logical and Enviro

all, Cornell Univer5701 ell.edu

parian z– Stockp

ndi J. Quacke

of Environmnce and Foreartment of El Science an

efits in termsf understandobjective of tatterns of buf-Stockport C

within aerial p6 and 2015. egetation regdegradation wUltimately, o

buffer vegettterns of veg

over (-1.45 –ed an increas

m aerial imerestimate th, and other nnsights into n be directly

nalysis, Spatia

STATESOURConmental Engineeri

rsity

zones: port Crenbush2

mental Resouestry, Syracu

Environmentnd Forestry, S

s of both cleding the currethis study waffer vegetati

Creek Watersphotos aboveWe tested f

gardless of siwithin each

our approachtation from hgetation vari

– -0.07%, -0.sing trend (0

magery, resulthe proportionnon-vegetatispatial and t

y utilized to a

al Autocorrel

E CES INS

ring

A case reek W

urce Engineeuse, NY tal Resource Syracuse, NY

ean water andent state andas to developion variationshed. With ae 97%, we w

four differenize. Hotspotyear, 30–40

h can enable high spatial iations.

1– -0.01km20.1%, 0.03 kmts show tradn of buffer von land covetemporal pataid buffer m

ation, Land C

STITUT

Em

study iWatershe

ering, SUNY

EngineeringY

d resilient cod effectivenep a method t

n over time. Wan overall acwere able to nt buffer sizet analysis rev0% of buffer

stakeholderresolution a

2) in 30–90 mm2).

ditional mappvegetation byer by 8–9%.tterns of buff

management a

Cover, Vegeta

TE Tel: (607) 254Fax: (607) 255

mail: nyswri@corn

in the ed

Y College of

g, SUNY

ommunities.ess of to delineate We tested

ccuracy of confidently s and found veals that areas are

rs, managers aerial images

m buffers

ping y 19–22%

ffer and

ation

4-7163 5-4449 ell.edu

f

s,

Geospatial assessment of riparian zones: A case study in the Hudson River Estuary – Stockport Creek Watershed

This report was prepared for the New York State Water Resources Institute (WRI) and the Hudson River Estuary program of the New York State Department of Environmental Conservation, with support from the NYS Environmental Protection Fund

Overview The 2015 Hudson River Estuary Action Agenda states that riparian restoration and protection are selected actions that will provide key benefits in terms of both clean water and resilient communities for 2015–2020 (NYS DEC 2015). Unfortunately, key knowledge gaps exist in terms of understanding the current state and effectiveness of riparian corridors in New York State. Yet this is not just a regional problem; since a national panel in 2002 concluded that the United States lacked a detailed map of the location and condition of riparian ecosystems (Brinson et al. 2002), various entities have reiterated the need for a comprehensive, detailed representation of riparian zones. Many of them have also identified this as a critical data gap (Salo and Theobald 2016). Thus, to fill this gap not only supports efforts of the Hudson River Estuary Program (HREP), but also contributes to the broader national attempt in mapping riparian zones. We conducted a pilot project mapping and assessing geospatial changes in riparian corridors within a Hudson Estuary River sub-watershed to aid riparian restoration and protection efforts by HREP, specifically the program of the Hudson Estuary Trees for Tribs (HETT). This program currently offers on-the-ground assistance to qualifying native riparian tree and shrub planting projects within targeted Hudson River watersheds. By applying our method, we generated results that provide valuable information that can help identify key riparian zones to allow targeted and effective prioritization of restoration projects, assisting in long term monitoring of riparian protection efforts, and supporting programs that aim to educate local communities about the critical roles riparian buffers play. The attached in-depth manuscript-style report provides more information on this project and explains the background, processing flow, results, conclusions and restoration recommendations.

Policy Implications There are currently no laws or regulations in the New York State regarding riparian buffer or zone protection. However, HETT and NYS Trees for Tribs are actively working with community groups and local governments to restore and protect riparian areas. We hope our method of delineating buffer vegetation can aid the decision and planning processes of various on-the-ground conservation efforts. The framework established by this study could also be utilized in potential riparian policy making processes for filling the information gap of buffer vegetation current states and historical trend. Outreach Comments We were in contact with Beth Roessler, Stream Buffer Coordinator for the Hudson River Estuary Program, periodically regarding this project. We will continue to reach out to her and share both our findings as well as the methods developed from this study. Furthermore, we will share all our findings and data from this project through a website hosted by SUNY ESF, as well as potential future presentations and publications. Student Training One doctoral candidate at SUNY ESF worked full time as a research project assistant for duration for this project. The training he received from this study includes enhanced understanding of: river channel boundary delineation, sampling design, supervised image classification using Google Earth Engine, and geospatial data analysis using ArcMap. Publications/Presentations Part of this project was presented at as a webinar named Riparian Buffer Assessment on March 19th 2018 sponsored by the New York State Geographic Information Systems Association. The remainder of this report is structured as the draft of a manuscript that will be further developed for future publication.


Ge Pu and Lindi J. Quackenbush

Table of Contents Introduction ..................................................................................................................................... 2 Methods........................................................................................................................................... 4

Study site: selected watershed and water bodies ........................................................................ 4 Data ............................................................................................................................................. 5 Processing Flows ........................................................................................................................ 5

Image Classification ................................................................................................................ 5 Channel boundary and land cover variations at 1 meter (NAIP) ............................................ 8 Land cover comparison between 1 m vs. 30 m datasets ......................................................... 8 Vegetation Degradation Cluster Analysis ............................................................................... 9

Results ........................................................................................................................................... 11 Classification Accuracy Results ............................................................................................... 11 Channel boundary and land cover variations at 1 meter (NAIP) .............................................. 12 Land cover comparison between 1 m (NAIP) vs. 30 m (NLCD) datasets ............................... 15 Vegetation Degradation Hotspot Analysis ................................................................................ 17

Moran’s I results ................................................................................................................... 17 Comparison of Gi* results within each year .......................................................................... 17 Comparison of Gi* results across years ................................................................................. 20

Discussion ..................................................................................................................................... 23 Classification Accuracy Results ............................................................................................... 23 Channel boundary and land cover variations at 1 meter (NAIP) .............................................. 23 Land cover comparison between 1 m (NAIP) vs. 30 m (NLCD) datasets ............................... 25 Vegetation Degradation Hotspot Analysis ................................................................................ 27 Management Recommendations ............................................................................................... 28

Conclusion .................................................................................................................................... 29 References ..................................................................................................................................... 30


2

Introduction Riparian buffers have long been recognized as important landscape features with ecological

importance that often far exceeds their spatial extent (Baker and others 2006). They play key

roles in providing unique habitat for many wildlife species, and serve as corridors for species

migration (Iverson and others 2001). Vegetation within riparian buffer zones, also known as

riparian buffer vegetation or buffer vegetation, helps maintain stream water quality through

filtering contaminants, dissipating energy associated with flood events, and stabilizing

streambanks through root systems (Dillaha and others 1989; Mulamoottil 1996; Lowrance and

others 1997; Correll 2005). Recent studies have also suggested that buffer vegetation extent can

positively influence run-off related effects of agriculture (Webber and others 2015; Chase and

others 2016; Lerch and others 2017).

The 2015 Hudson River Estuary Action Agenda describes riparian restoration and protection as

actions that will provide key benefits in terms of both clean water and resilient communities for

2015–2020 (NYS DEC 2015). Unfortunately, key knowledge gaps exist in terms of

understanding the current state and effectiveness of riparian corridors in New York State. This

gap creates obstacles in restoration and protection efforts. Yet this is not just a recent nor a

regional problem; a national panel in 2002 concluded that the United States lacked detailed maps

of the location and condition of riparian ecosystems (Brinson and others 2002). Since then,

various entities have reiterated the need for a comprehensive, detailed representation of riparian

zones and identified this as a critical data gap (Salo and others 2016). Thus, to fill this gap not

only supports the efforts of the Hudson River Estuary Program (HREP), but also contributes to

the broader national goal of mapping riparian areas.

Past studies have utilized the National Land Cover Dataset (NLCD) to map riparian buffers

(Jones et al., 2010; Weller et al., 2011; Weller and Baker, 2014) because it is a consistent product

for study sites across the United States that is freely available and has convenient access.

However, this dataset was produced from Landsat imagery with pixel sizes at 30 m. This may be

sufficient for mapping buffers at a regional scale, but has limitations in being applied to on-the-

ground management efforts at a local scale due to the impact of land cover mixing within 30 m

pixels (Hollenhorst et al. 2006). Recent studies by Hayes et al., (2014) have begun to utilize high


3

resolution (>1 m) to map riparian buffer vegetations, and their results suggested that the freely

available aerial images from United States Department of Agriculture National Agriculture

Imagery Program (NAIP; USDA-FSA-APFO Aerial Photography Field Office, 2016) could be

used to classify riparian land cover with moderate accuracy. However, more testing is still

needed, e.g. to evaluate similar approaches in non-arid regions of the United States or to

determine utility of applying multi-temporal images. In addition to changes in available imagery,

new methodologies should also take advantage of recent technological advances in remote

sensing processing platforms (e.g. Google Earth Engine). Such approaches reduce the data

management burdens by utilizing cloud storage, reduce processing cost with parallel computing,

and gain easy access to high resolution data (e.g. public orthoimagery) or self-captured very high

resolution data (e.g. from Unmanned Aerial Vehicle imagery).

In addition to mapping the extent of riparian buffers, there are ongoing efforts to analyze the

spatiotemporal patterns of buffer vegetation variations using various approaches (Apan and

others 2002; Shandas and Alberti 2009; Constança Aguiar and others 2011; Fernandes and others

2011). Past studies have utilized various landscape metrics—e.g., mean shape index, patch

statistics, and mean proximity index—to quantify the spatial patterns of vegetation variations.

However, these traditional statistical approaches generally ignore the spatial dependence of data

and do not reveal local landscape metrics (Fernandes and others 2011). Fernandes et al. (2011)

utilized global and local autocorrelation metrics to analyze buffer vegetation landscape patterns

using aerial photos taken for one specific time. Recent studies have just begun to utilize

spatiotemporal autocorrelation as a tool for analyzing and detecting vegetation loss hotspots

(Harris and others 2017). However, effort is still needed to develop approaches for improving

current methods to examine buffer vegetation spatial patterns over time.

The primary objective of this study was to develop a method that not only delineates buffer

vegetation from aerial imagery, but also quantifies the spatiotemporal patterns of vegetation

variation. Testing this new method focused on: (1) quantifying the accuracy of image

classification for mapping buffer vegetation from multi-temporal aerial photos, (2) classifying

NAIP images for mapping riparian buffer land covers and measuring the spatiotemporal changes

of stream channel boundaries and their buffer vegetation, (3) determining the difference between


4

mapping riparian buffer vegetation by classifying NAIP images vs. using products such as the

NLCD, and (4) quantifying the spatiotemporal patterns of buffer vegetation change and

computing the statistical significance of that change.

Methods Study site: selected watershed and water bodies Our project focused on the Stockport Creek watershed (Figure 1), a sub-watershed of the Hudson

River located on the New York–Massachusetts border. Stockport Creek is the second largest

tributary to the tidal Hudson River with 1733 km of streams that comprise 12% of the total

stream length in the Lower Hudson River Basin (Martino 2012). Selection of this watershed

aligns well with the mission of the Hudson Estuary Trees for Tribs program and ensures a direct

link to support current grassroots efforts for riparian restoration and protection from this project.

Figure 1. Stockport watershed in southeastern New York State showing creeks included in study: Claverack Creek, Fitting Creek, Kinderhook Creek, Kline Kill, and Taghkanic Creek.


5

As shown in Figure 1, five streams from the Stockport Creek watershed were considered for this

study: Claverack Creek, Fitting Creek, Kinderhook Creek, Kline Kill, and Taghkanic Creek

These particular streams were selected due to the limitation of the available aerial images to

identify stream channels that are smaller than 1 m. Given that streams higher than 3rd order in the

Stockport Creek watershed are generally larger than 1 m, we utilized this as a guideline in

selected our study streams.

Data This study utilized public aerial photography provided through the NAIP. The NAIP dataset is

comprised of airborne orthorectified images acquired at 1 m ground sampling distance (GSD).

The NAIP program collects imagery during the growing season on a semi-annual basis, with

images of the study are available from 2006–2015. The program generally acquires visible image

bands (blue, green, red), with near-infrared (NIR) information also available for the study site

from 2011. This combination of high spatial, moderate spectral, and lower temporal resolution

enables interpolation of detailed information on the boundaries of river channels and

characterization of the presence of buffer vegetation. With a focus on quantifying the changes in

buffer vegetation extent, NAIP images from 2006 and 2015 were utilized in this study.

NLCD was also used to compare the differences in buffer land cover delineation results from 1

m (NAIP-derived) and 30 m (NLCD) spatial resolutions. NLCD is a land cover dataset derived

from Landsat imagery using decision tree classification (Homer and Fry 2012). This dataset was

produced at 30 m pixel size and available for the study site on 2006, 2011 and 2016. Since NAIP

was not available in other two years, we utilized the 2006 NLCD (30 m) to compare with our

2006 NAIP-derived dataset (1 m).

Processing Flows Image Classification Prior to performing image processing, the channel boundaries of the five selected streams were

manually delineated in Google Earth Engine (GEE) using the USDA NAIP imagery. Since the

location of the channel varied between observations, channel boundaries were digitized


6

separately for both 2006 and 2015 NAIP images. Historical hydrological data was not available

for the stream, hence we were not able to confirm that the river depth was similar in 2006 and

2015. However, visual assessment comparing the two NAIP images, suggest that there were no

significant channel expansion or contractions between the two years.

Having identified the channel boundaries, the second step of the process buffered each channel

boundary to create the limit of the riparian buffers. It is important to note that there is no

universal agreement on buffer distance for riparian management (Salo and others 2016; Dempsey

and others 2017). Thus, we selected four incremental distances—30, 60, 90 and 120 m—to

accommodate such uncertainty. A similar approach was used by Dempsey et al. (2017). By

selecting multiple buffer distances we could assess the effect of buffer size on the buffer

vegetation analysis.

The third step of the process was classifying the riparian buffer pixels as vegetation or non-

vegetation. For each available year, the NAIP images were clipped using the 120 m buffer

boundaries, and then each pixel in the buffer zone was classified into vegetation or non-

vegetation classes. Reference points were computer-generated randomly within the riparian

buffer boundaries and visually assigned to vegetation or non-vegetation classes and divided into

training and validation samples. The total number of reference samples varied between 2006 and

2015 (Table 1), with more points sampled in 2006 than 2015. Such adjustment was done to

accommodate the smaller number of non-vegetation points in 2006 and ensure there were an

adequate number of points for validation. The general guideline for determining the number of

reference points was referenced from Warner et al. (2009).

Table 1. Distribution of sampling points.

Year Total Vegetation Non-Vegetation2006 3310 3130 179 2015 3185 2664 521

Classification of the pixels within the buffer zone as vegetation or non-vegetation utilized the

random forest method (Immitzer and others 2012; Lee and others 2016) based on the training

samples generated. The input layers for the random forest classification included the reflectance


7

for pixels within each NAIP image band. With the NIR band available in 2015 image, a

normalized difference vegetation index (NDVI) derived from the red and NIR bands was also

used as a classification input to help differentiate vegetation and non-vegetation classes.

Classifier training was based on 70% of the reference points while the rest were reserved for

accuracy verification through generating confusion matrices. These matrices are used to describe

classification accuracy and characterize errors (Foody 2002). Optimization of the number of

trees for the random forest classifier was performed in R while buffering, clipping and

classification of the images were done in GEE.

Upon producing the classified riparian buffer land cover maps, post processing procedures were

utilized to: (1) reduce the speckle effect of image misclassification, and (2) remove agriculture

from the classified buffer vegetation land cover. The speckle, or “salt and pepper,” effect is

common to pixel-based classification of fine spatial resolution images (Ke and others 2010). This

effect created misclassification where small groups of vegetation pixels (especially in riparian

forests or agricultural fields) were mistakenly classified as non-vegetation. We utilized the

connected pixel function in GEE to eliminate isolated pixels that were likely misclassifications.

Prior to using the function, visual inspection confirmed the majority of the speckles were under a

patch size of 50 pixels. Thus, filters were used to eliminate the speckles in both vegetation and

non-vegetation classes. Having smoothed the classification, we manually converted vegetation in

agriculture fields within the riparian buffers into the non-vegetation land cover class. This

decision was reached based on the recommendation of the New York State Department of

Environmental Conservation, which states farm fields are not considered as riparian buffer

vegetation. (NYS DEC 2018).

Ultimately, the final classified land cover maps contained two distinct classes: vegetation and

non-vegetation (Table 2). We clipped the final classified maps to 30, 60, and 90 m for comparing

the effect of buffer size variations on the outcome of the land cover variations and further

analysis, but focused the classification accuracy result on the 120 m buffers, since that included

the smaller zones.


8

Table 2. Land cover classes in the classified NAIP maps.

Land cover class Description Example features

Non-vegetation Land not covered by any type of vegetation or vegetated land within agricultural fields.

Impervious surfaces; agriculture fields; dairy farms

Vegetation Land covered by vegetation outside agriculture fields.

Natural forests and shrubs; tree nurseries; lawns*

*Lawns were observed in the images, but cover a significantly lower portion of land than agriculture fields.

Channel boundary and land cover variations at 1 meter (NAIP) We compared the variations of channel boundaries and land cover compositions using the

classified 2006 and 2015 NAIP images. Cumulative area and percentage of both vegetation and

non-vegetation land cover classes were computed for each year. Temporal difference of each

land cover classes over their respective buffer areas were also computed. These differences were

further validated by manual interpolations by comparing 2006 and 2015 NAIP images. Visually

clear and examples of significant changes in buffer vegetation were recorded and presented using

time series imagery.

Land cover comparison between 1 m vs. 30 m datasets In order to assess the impact of spatial resolution on riparian vegetation characterization, the

project goals originally intended to compare NAIP and Landsat classification results. However,

with low class accuracy (e.g. 50%) for the Landsat classification, instead of deriving a 30 m

product, we compared the 2006 classified NAIP land cover with the 2006 NLCD, which was

derived from Landsat data. We compared the total area and percent area for land cover types

generated at the two different image source scales within the 30 m, 60 m, 90 m and 120 m

riparian buffers. Since the NAIP-derived land cover types were initially different from those in

the NLCD, we consolidated and modified land cover types from both datasets in order to

compare them in a meaningful way. Table 3 shows the corresponding land cover classes between

NAIP-derived and NLCD datasets.


9

Table 3. Corresponding land cover classes between NAIP-derived and NLCD datasets.

NAIP-derived land cover classes NLCD land cover classes Vegetation Forest, Shrubland, Herbaceous, Wetlands

Agriculture* Planted/Cultivated Other* Developed, Barren, Water

* Indicates classes that were combined in the final Non-vegetation class in the classified NAIP images.

Vegetation Degradation Cluster Analysis In order to identify vegetation degradation in the riparian buffers, we utilized Getis-Ord (Gi*;

Getis and Ord, 2010) and Moran’s I (Moran 1950) statistics to determine spatial clusters of non-

vegetation land cover. Gi* statistics are local indicators of spatial associations. Goetz et al. (2017)

used Gi* to identify the location and degree of spatial clusters of forest loss. We used a similar

method to quantify the loss of riparian buffer vegetation over time based on the NAIP-based land

cover maps. Areas with locally high levels of non-vegetation and vegetation land cover were

quantified, along with their Z scores (standard deviation) and p values (statistical probabilities).

We defined hot-, cold- and nonsignificant-spots as shown in Table 4.

Table 4. Definitions of hot, cold and nonsignificant spots.

Detailed Label Definition Hotspot Location surrounded with high presence of non-vegetated land cover, with

|Z score| > 1.96 and p < 0.05. Coldspot Location surrounded with high presence of vegetated land cover, with |Z

score| > 1.96 and p < 0.05. Nonsignificant Spots Any location that is not a statistically significant hot or cold spots.

In addition to the local spatial association analysis, a global indicator—Moran’s I —was also

utilized to quantify the significance of clustering from a global perspective. This global statistics

can indicate the existence and degree of spatial autocorrelation. Moran’s I values show whether

global spatial patterns are dispersed (0). Since

Moran’s I results are distributed normally, computed statistics can be assessed for statistical

significance with Z scores and p values (Fischer and Getis 2009). It is also worth noting that Gi*

is mathematically related with Moran’s I (Fischer and Getis 2009; Getis and Ord 2010), thus one

can relate the two based on the existence of autocorrelation and statistical significance.


10

We developed a model in ArcMap to obtain the Gi* and Moran’s I statistics, which involves three

main processing steps: data aggregation, spatial weight generation, and hotspot analysis. At the

data aggregation stage, classified NAIP images at 1 m pixel size were aggregate into new images

with pixel sizes (30, 60, 90 or 120 m) based on their selected buffers. NAIP-derived land cover

pixels were recorded using 1 m pixels and then transformed into points located at the center of

each pixel. Points that had vegetation were removed for later steps. Each transformed point

represents one incident of absence of buffer vegetation for one year in one selected buffer size.

Vegetation change was considered using the 30, 60, 90 and 120 m buffers, with the extent of the

constructed buffers dictating the study area of interest for one particular buffer. An array of

aggregation grids were then constructed using a selected cell size. We set the aggregation cell

size equal to the corresponding buffer size for avoiding inclusion of null data outside the buffer

area (> buffer size), while not exceeding the computational capacity of the program (< 35 m).

The spatial weight matrix generation stage constructed matrices for the selected buffer sizes.

Spatial weight (W) matrices encompass the preconceived understanding of spatial relationships

(Fischer and Getis 2009). There are various approaches to construct such matrices; we sought to

use the W matrix as a geometric indicator of spatial nearness and so utilized the popular k-

nearest neighbors spatial weight matrix (Bailey and Gatrell 1995), which matches the underlying

conceptual spatial relation in the input data. The assumption here is that at any particular

location, nearby land cover has much larger impact than points farther away. Utilizing the kth

nearest neighbor also ensures that all features have the same number of neighbors prior to the

hotspot analysis. We set the number of neighbors to 10 to ensure an adequate number of

neighbors were analyzed. Empirical testing varying the number of neighbors from 10 to 100 in

increments of 10 confirmed that 10 neighbors was adequate to identify individual spatial clusters,

but did not overstate hotspot size. Row standardization was applied in constructing the spatial

weights so that a given observation is not considered a neighbor of itself. Researchers have found

the approach useful in equally weighting observations (Fischer and Getis 2009).

At the hotspot analysis stage, both global (Moran’s I) and local (Gi*) spatial autocorrelation

statistics were computed based on the constructed spatial weight matrixes. As mentioned before,


11

Moran’s I indicates the existence and degree of global spatial correlation, while Gi* indicates the

location of local clusters and spatial nonstationarity. Both statistics are normally distributed, thus

Z score and p value were also computed to represent statistical significance. The Gi* statistics

were used to identify hot-, cold-, and nonsignificant spots according to the definitions shown in

Table 4.

In order to compare across years, we generated a new set of labels to classify the changes among

the hot-, cold- and nonsignificant spots between 2006 and 2015 (Table 5). These labels were

utilized to identify the spatiotemporal variation of hot- and coldspots. Four main categories were

established: new, diminishing, persistent, and intensifying. Each of these was then separated into

corresponding hot- and coldspot, e.g. “New Hot” is a 2015 hotspot that was nonsignificant in

2006. Classification using generated labels was done with a series of automatic processes

through an ArcMap Model.

Table 5. Labels of temporal variations between hot- and coldspots based on land cover maps classified from NAIP images.

Main Categories Detailed Label 2006 Condition 2015 Condition

New New Hot No significant pattern High local non-vegetation

New Cold No significant pattern High local vegetation

Persistent Persistent Hot High local non-vegetation High local non-vegetation

Persistent Cold High local vegetation High local vegetation

Intensifying Intensifying Hot High local non-vegetation Higher local non-vegetation

Intensifying Cold High local vegetation Higher local vegetation

Diminishing Diminishing Hot High local non-vegetation No significant pattern

Diminishing Cold High local vegetation No significant pattern

Transition Transition Zone Transition between hot and cold between 2006 and 2015

Results Classification Accuracy Results The confusion matrix generated after performing the two-class image classification (Table 6)

shows that overall accuracies of classification were high (at or above 97%). Individual class

accuracies were generally at or above 90%, except for the producer’s (PA) and user’s (UA)


12

accuracies of the 2006 non-vegetation class (74% and 87%, respectively). Regardless of year, the

vegetation class had higher PA and UA (from 6–25%) than the non-vegetation class. Vegetation

class accuracies were similar across the two years, while non-vegetation accuracies were higher

in 2015 than in 2006.

Table 6. Confusion matrix of NAIP image classification results. Producer’s accuracy reflects how well the reference data was classified, and user’s accuracy shows the reliability of classes in the classified images.

Year Producer (%) User (%) Overall Accuracy (%) Vegetation Non-Vegetation Vegetation Non-Vegetation2006 99 74 98 87 98 2015 98 92 98 90 97

Channel boundary and land cover variations at 1 meter (NAIP) Changes in the meander over time led the total length of the river centerline to increase by 2.48

km (2%) from 2006 to 2015. The total area within the riparian buffer zones increased, depending

on the buffer width considered, by up to 0.1 km2. Composition of vegetation and non-vegetation

within the riparian buffers varies both spatially and temporally. Within each year, the percentage

of vegetation decreases as the buffer length increases with a corresponding increase in the non-

vegetation class (Figure 2 and Figure 3). There is generally a larger percentage of vegetation

(77–91%) than non-vegetation (8–22%) in the buffer regardless of buffer size.

Figure 2. Area of non-vegetation land cover and vegetation land cover for 30, 60, 90, and 120m riparian buffers in 2006 and 2015 NAIP images.

0

5

10

15

20

25

30

30 60 90 120

Are

a (k

m2 )

30 60 90 120

Buffer Distance (m)

2006 2015


13

Figure 3. Percentage of land cover classes within buffer areas.

Comparing the 2006 and 2015 land cover results, the percentage of riparian vegetation within the

riparian buffer zone decreased by 1.45% within the 30 m buffer, 0.57% within 60 m and 0.07%

within 90 m (Figure 4). However, there is a marginal increase (0.1%) within the 120 m buffer.

Figure 4. Percent change of riparian vegetation and non-vegetation land cover from 2006 to 2015.

Large land cover changes between the classified 2006 and 2015 images were cross validated

with visual-interpretation of the intermediate NAIP images—i.e. including the images acquired

-0.04-0.02

00.020.040.060.08

0.10.120.140.160.18

0 30 60 90 120 150

Perc

ent c

hang

e fr

om 2

006

to 2

015

Buffer Distance (m)

Geospatial a

in 2008,

detectabl

vegetatio

regenerat

these loc

Figure 5. N42.498° N

Figure 6. Acenters at

assessment of

2009, 201

le variation o

on to non-v

tion or expa

ations is sho

NAIP images sand 73.494° W

An example of42.374° N and

riparian zone

1, and 2013

of riparian v

egetation la

ansion of ve

own in Figur

showing ripariW. Area highlig

f NAIP imagesd 73.709° W. A

es: A case stud

3. This ana

egetation. A

and cover as

egetation fro

re 7.

ian vegetationghted in red is

s showing ripaArea highlighte

dy in the Huds

14

lysis reveal

Among them,

s shown in

om non-vege

decrease alons a section of th

arian vegetatioed in red is a s

son River Estu

led 23 locat

, only one lo

Figure 5.

etation land

ng the Kinderhhe 120 meter c

n regrowth alosection of the 1

uary – Stockp

tions that ex

ocation show

The other 1

cover (Figu

hook Creek. Imchannel buffer

ong the Kinder120 meter chan

port Creek Wa

xhibited vis

wed conversi

19 locations

ure 6). A m

mage centers ar.

rhook Creek. Innel buffer.

atershed

sually

ion of

s had

ap of

t

Image


15

Figure 7. Map of visually confirmed changes (2006 to 2015) in riparian vegetation within the 120 meter buffer zone. Land cover comparison between 1 m (NAIP) vs. 30 m (NLCD) datasets Results of the comparison of the amount of each land cover class within the different buffers are

shown in Table 7 and Figure 8. There are small differences in the total buffer area between the

NLCD and NAIP-derived data (less than 1%), hence we also compared the proportion of each

land cover type in the buffer. Figure 8 shows that the NLCD (30 m) underestimated the

proportion of riparian buffer vegetation by 19–22% compared to the NAIP-derived mapping (1

m) while overestimating agriculture by 10–14%, and other non-vegetation land covers by 8–9%.


16

Our classified 2006 buffer land cover map at 1 m pixel size has a higher overall accuracy than

the reported 2006 NLCD region 10 (Northeast US) average. Wickham et al. (2013) reports the

overall accuracy of the land cover classes in the 2006 in region 10 of NLCD mapping zone has

an overall accuracy of 71% for 21 classes. Meanwhile, our classified buffer land cover maps

from 2006 NAIP image has an overall accuracy of over 97% for 2 classes.

Table 7. Area of Vegetation, Agriculture and Other land cover classes for both NAIP-derived and NLCD datasets within 30, 60, 90 and 120 meter buffers.

Buffer Size (m) Vegetation (km2) Agriculture

(km2) Other (km2) Total (km

2)

NA

IP

-der

ived

30 6.61 0.48 0.15 7.24 60 12.08 1.94 0.30 14.32 90 17.01 3.74 0.47 21.22

120 21.68 5.57 0.65 27.90

NLC

D 30 5.04 1.50 0.75 7.29

60 9.28 3.53 1.55 14.36 90 13.10 5.95 2.26 21.31

120 16.46 8.52 3.03 28.01

Figure 8. Proportion of total buffer area (for 30, 60, 90 and 120 m buffers) within Vegetation, Agriculture and Other land cover classes for both NAIP-derived and NLCD datasets.


17

Vegetation Degradation Hotspot Analysis Moran’s I results When the buffers are considered as a whole, the Moran’s I statistics show that regardless of

buffer size, the 2006 and 2015 buffers show statistically significant (p < 0.01) clustering of non-

vegetated land cover. As shown in Table 8, all Moran’s I values are above 0.3. These values all

have Z scores above 4 and p values below 0.01, showing that statistically significant global

clustering exists.

Table 8. Moran’s I results. All results in this table have a Z score above 4 and p value below 0.01.

Year Buffer Size (m)

30 60 90 120 2006 0.34 0.33 0.32 0.32 2015 0.33 0.31 0.3 0.3

Comparison of Gi* results within each year In this analysis, the term hotspot is used to refer to statistically significant clusters of non-

vegetation land cover while coldspots are clusters of vegetated cover. As expected, the total area

of hot- and coldspots generally increases as the size of buffer increases (Table 9), although there

is no clear pattern on the magnitude of such increases. When we normalized total area with

respect to buffer size, 15–22% of buffer areas have clusters of vegetation degradation within

each year. Variation in buffer size does not show significant impact on the composition (0–6%)

of non-zero hot- or coldspots. Areas of individual hotspot increase in intensity and expand

outwards from 30–120 m buffers (Figure 9 and Figure 10). There are no coldspots in 2006 for

the 30 or 60 m buffers or in the 30 m buffer in 2015. In the buffer sizes where both hotspots and

coldspots were detectable, composition of these spots over their respective buffer areas are

similar.

Table 9. Cumulative area of coldspots and hotspots (in km2) and composition of areas over their respective buffer areas are in parentheses.

Year Category Buffer Size (m)

30 60 90 120

2006 Coldspot 0 (0) 0 (0) 8.43 (23%) 10.70 (23%) Hotspot 2.22 (15%) 5.17 (21%) 7.82 (21%) 10.3 (22%)

2015 Coldspot 0 (0) 3.45 (13%) 7.59 (21%) 9.98 (21%) Hotspot 2.31 (15%) 5.02 (19%) 7.68 (21%) 10.34 (22%)


18

Figure 9. Locations of statistically significant hot- and coldspots in 2006. Hotspots represent significant clusters of non-vegetation land cover within the respective riparian buffer size, while coldspots represent significant clusters of vegetation.


19

Figure 10. Locations of statistically significant hot- and coldspots in 2015. Hotspots represent significant clusters of non-vegetation land cover within the respective riparian buffer size, while coldspots represent significant clusters of vegetation.


20

Comparison of Gi* results across years When comparing across years, the total area of hot- and coldspots also revealed that the

cumulative area of hotspots increased in the 30 and 120 m buffers, while decreased in the 60 and

90 m buffers. The magnitude of these changes ranges from 0.04–0.15 km2 (Table 9). Coldspots

in the 90 and 120 m buffers decreased by 0.72–0.84 km2. As the buffer size increase, areas that

exhibit increasing (New and Intensifying) and constant (Persistent) local non-vegetation

clustering also increases (Figure 11a). The magnitude of increase is most significant in the

Persistent Hotspot category. For coldspots, Persistent, New and Diminishing spots showed

varying magnitude of increases for 90 and 120 m buffers. No areas were observed that show a

transition from hotspot to coldspot or vice-versa.

The composition of total area for each of the newly developed labels reveals different patterns

(Figure 11b). The composition for all areas that show increasing (New and Intensifying) and

constant (Persistent) magnitude in non-vegetation clustering observed a general decreasing trend

from 30–120 m buffers. For coldspots, Persistent and New spots also show a decreasing trend

from 30-40%, while Diminishing spots show an increasing trend from 90–120 m buffer all below

10%.

Regardless of the buffer size, there are groups of persistent and intensifying hotspots near

location A and B (Figure 12). Location A is located in the town of Claverack, NY along the

Claverack Creek, while location B is located to the SW of the village of Valatie, NY along the

Kinderhook Creek. The apparent magnitude of these spots increases as the buffer size increases.

Persistent and intensifying coldspots appear near location C (in Nassau, NY along Kinderhook

Creek) and D (in Hudson, NY along Stockport Creek) for the 90 and 120 m buffers. Near the

town of Stephentown, towards the NE corner of the town, there is a mixture of both persistent

hot and cold spots.


21

(a)

(b)

Figure 11. (a) Total area of changes and (b) composition of area respective buffer area in each category of temporal variations between 2006 and 2015 hot- and coldspots. Definition of each category is shown in Table 5.

020406080

100120140

Transition IntensifyingCold

Persistent Cold New Cold DiminishingCold

Dinimishing Hot New Hot Persistent Hot Intensifying Hot

Tota

l Are

a(S

quar

e K

ilom

eter

s)

30 m 60 m 90 m 120 m

0%10%20%30%40%50%

Com

posi

tion

Ove

rR

espe

ctiv

e B

uffe

r Are

a(%

)


22

Figure 12. Locations of variations between hot- and coldspots from 2006 to 2015. Hotspots represent significant clusters of non-vegetation land cover within the respective riparian buffer size, while coldspots represent significant clusters of vegetation. Labels A–D highlight distinct spots. Location A is located in the town of Claverack, NY along the Claverack Creek, while location B is located to the SW of the village of Valatie, NY along the Kinderhook Creek. Persistent and intensifying coldspots appear near location C (in Nassau, NY along Kinderhook Creek) and D (in Hudson, NY along Stockport Creek) for the 90 and 120 m buffers.


23

Discussion Classification Accuracy Results Overall accuracy of both 2006 and 2015 NAIP image classifications was satisfactory at above

90%. In terms of UA and PA for individual map classes, the vegetation classes had higher

accuracy than non-vegetation classes in both 2006 and 2015. The lower accuracy of the non-

vegetation delineation results might be caused by the number of sampling points not being

sufficient to represent the range of cover types that were included in the non-vegetation class.

The improvement of UA and PA values in 2015 compared to 2006 was likely related to either

the improved GSD or the value of the near-infrared band information. Nevertheless, with all UA

and PA values generally above 80% across all map classes, map users should be able to

confidently use land cover information within our classified NAIP maps regardless of which year

is applied.

Channel boundary and land cover variations at 1 meter (NAIP) The percentage of riparian vegetation extent within the buffer areas showed an increase from

2006 to 2015 for the 90 and 120 m buffers (Figure 4). As the buffer zone gets closer to the

channel, higher percentages of riparian vegetation were converted into agriculture or impervious

land covers. Even though the magnitude of the changes is small, the area represented by such

percentages can be significant. For example, a 0.07% decrease at 90 m buffer zone means a total

area of 10,000 m2 of riparian vegetation were converted from 2006 to 2015.

Our results confirmed that different buffer size can lead to different interpretation on how buffer

vegetation composition varied from 2006 to 2015 (Table 9). Such differences can impact

downstream riparian management decisions, which could lead to potential over or under

planning of vegetation restoration efforts. Future efforts of this study will explore the use of

variable width riparian buffer delineation as proposed by Salo and Theobald (2016) to reduce the

impact of buffer size variations. However, whether it is fixed width or variable width delineation

approach, our developed method can be easily modified to accommodate both methodologies in

providing insights on riparian buffer vegetation.


24

Through comparing our visual validation results and NAIP-derived buffer land cover maps, one

particular benefit of using image classification became clear. With visual inspection, we could

only detect a total of 23 locations of land cover changes within the buffer. Our image

classification on the other hand was able to detect small pixel-based variations that were missed

by manual assessment. Such differences confirm the benefit of using image classification not

only to detect small variations, but also in terms of time and resource savings compared to

manual interpolation.

Some caution is advised in utilizing the reported area differences between the image

classifications for the two years. Some of the variations are likely contributed by the combined

effects of the change in pixel size and misregistration of the datasets. The 2015 NAIP imagery

has a pixel size of 50 cm compared to the 1 m pixels in 2006, which increased the capability of

image classification in 2015 (Figure 13). There is also a slight misregistration between the

datasets with areas where horizontal positions are clearly mismatched between the two image

dates. Further steps will be taken to quantify the impact of the registration uncertainties, for

example through determining the horizontal accuracies of the 2006 and 2015 NAIP images in the

field. For now, percentages of land cover types within the buffer area are utilized for analysis

since it is less impacted by the misalignment between the two images.


25

Figure 13. Example of misalignment between 2006 and 2015 NAIP images and their classified results. Location of each image boundary is the same. Within the classified images, white color represents classified vegetation land cover, black color represents non-vegetation. Image centers are at 42.507° N 73.506° W.

Land cover comparison between 1 m (NAIP) vs. 30 m (NLCD) datasets The difference of total buffer areas between NAIP derived land cover and NLCD in Table 7 is

caused by the impact of the different pixel size (NAIP-derived 1 m and NLCD 30 m) in defining

the boundary. This is due to the nature of raster images to generate pixelated boundaries instead

of smooth edges like vector datasets (Figure 14). We utilized proportions within the buffer zones

in order to minimize the impact of this known issue.

2006 NAIP image 2015 NAIP image

Classified 2006 NAIP image Classified 2015 NAIP image

0 20 m


26

Figure 14. Buffer area difference caused by the impact of pixel size (NAIP-derived 1 m and NLCD 30 m) on delineation of feature borders.

The NLCD has been utilized in prior riparian buffer land cover studies (Jones and others 2010;

Weller and others 2011; Weller and Baker 2014) because it is a consistent product for study sites

across the United States that is freely availability with convenient access. However, this study

suggests that the NLCD underestimates the buffer vegetation land cover and overestimates the

non-vegetation cover classes compared to the higher spatial resolution products. Management

decisions based on this interpretation could lead to a potential waste of resources. The variation


27

between NAIP-derived (1 m) and NLCD (30 m) land cover is mainly due to the impact of land

cover mixing within a pixel at lower spatial resolution, which is well documented (Hollenhorst et

al. 2006). In order to mitigate the potential for misinterpretation, we recommend future studies

use higher spatial resolution images (≤1 m) for riparian buffer land cover analysis. Using higher

resolution images not only reduces mixed pixel issues, but also creates more detailed land cover

maps to better aid riparian restoration efforts on the ground.

Vegetation Degradation Hotspot Analysis The magnitude of Moran’s I values are all above 0.3 (Table 8), which means that the 2006 and

2015 buffers did not have a dispersion (


28

meaningful. As we normalize these increase over the increased cumulative buffer sizes, we

found the area composition of new labels were decreasing. Our normalization process revealed

new information that is more meaningful than comparing using area alone.

Cluster analysis revealed landscape patterns in the land cover data, which is essential to our

understanding of how natural buffer vegetation was altered over time. Such information is not

easily detected by manual interpretation, especially over large areas. The cluster results can be

beneficial for managers to target and prioritize their restoration or protection efforts. They can be

particularly helpful in situations where resources for buffer conservation efforts are limited.

Management Recommendations Based on our findings, we can offer four management recommendations for the Stockport

Watershed:

1) Selection of buffer size can dictate the outcome of buffer vegetation analysis. Depending

on the selected size, the vegetation composition will show different patterns over time.

However, as the buffer size increased, the percentage of vegetation land cover increased

while non-vegetation decreased between the two years. Thus, we advise to use similar

multiple-buffer-approach like in this study to gain comprehensive results instead of using

one single buffer size. If the outcome of the majority of the buffer sizes show a

decreasing trend in vegetation cover, then it is more appropriate to reach the conclusion

of buffer vegetation is declining. We also recommend more prioritization toward smaller

buffer sizes since the land immediately adjacent to the streams are more significant that

those further away.

2) Our results show that NLCD at 30 m cell size may not be appropriate for assisting in on-

the-ground buffer vegetation assessment and restoration efforts. We recommend using

NAIP images as much as possible for assisting those efforts not only because they are

free to use, but also higher spatial resolution data can improve the efficiency of map

interpretation during planning or executing restoration efforts. In areas where high

solution images are not available, it is vital to recognize the shortcomings of using NLCD

in studying riparian buffers.


29

3) According to our results, we recommend prioritizing vegetation management efforts in

the colored areas shown in Figure 12. These are areas where statistically significant

clustering of vegetation and non-vegetation land cover were observed. We advise to

focus restoration efforts in the areas where non-vegetation presence is increasing over

time (Intensifying), remains high (Persistent), or in areas where vegetation has been

replaced by non-vegetation (New Hotspot). Protection efforts should be directed to

locations where vegetation is increasing (Intensifying), remaining high (Persistent) or has

replaced non-vegetation (New Coldspots).

Conclusion We developed a method to both delineate riparian buffer vegetation and analyzing

spatiotemporal patterns of vegetation degradation. It is not only among the first attempts to map

buffer vegetation using publicly available high resolution datasets with a cloud computing

platform, but also to implement the concept of hotspot analysis into buffer vegetation

evaluations. It streamlined the process from obtaining images to generating degradation hotspots.

Testing of our new method show high image classification accuracy and improved detection

capabilities over using NLCD. Decreasing trends of vegetation land cover at was observed at our

test site. Results of cluster analysis provided further insights of both spatial and temporal patterns

of buffer vegetation degradation. Such information can be directly utilized to aid buffer

management and restoration efforts on the ground.

The framework establish in this study can be implemented to other regions of the US with minor

modifications due to the nature of the data and platforms applied. With our new method,

stakeholders who do not have extensive training in remote sensing can monitor spatiotemporal

variations in buffer vegetation, and prioritize their riparian buffer restoration and protection

efforts.


30

References Apan AA, Raine SR, Paterson MS. 2002. Mapping and analysis of changes in the riparian

landscape structure of the Lockyer Valley catchment, Queensland, Australia. Landsc Urban Plan 59:43–57.

Bailey TC, Gatrell AC. 1995. Interactive spatial data analysis. Longman Scientific & Technical Essex

Baker ME, Weller DE, Jordan TE. 2006. Improved methods for quantifying potential nutrient interception by riparian buffers. Landsc Ecol 21:1327–45. http://link.springer.com/10.1007/s10980-006-0020-0

Brinson M, MacDonnell J, Austen J, Beschta L, Dillaha A, Donahue L, Gregory V, Harvey W, Molles-Jr. C, Rogers I, Stanford A. 2002. Riparian areas: functions and strategies for management.

Chase JW, Benoy GA, Hann SWR, Culp JM. 2016. Small differences in riparian vegetation significantly reduce land use impacts on stream flow and water quality in small agricultural watersheds. J Soil Water Conserv 71:194–205.

Constança Aguiar F, Rosário Fernandes M, Teresa Ferreira M. 2011. Riparian vegetation metrics as tools for guiding ecological restoration in riverscapes. Knowl Manag Aquat Ecosyst:21. http://www.kmae-journal.org/10.1051/kmae/2011074

Correll DL. 2005. Principles of planning and establishment of buffer zones. Ecol Eng 24:433–9.

Dempsey JA, Plantinga AJ, Kline JD, Lawler JJ, Martinuzzi S, Radeloff VC, Bigelow DP. 2017. Effects of local land-use planning on development and disturbance in riparian areas. Land use policy 60:16–25.

Dillaha TA, Reneau RB, Mostaghimi S, Lee D. 1989. Vegetative Filter Strips for Agricultural Nonpoint Source Pollution Control. 32. http://elibrary.asabe.org/abstract.asp?aid=31033&t=3

Fernandes MR, Aguiar FC, Ferreira MT. 2011. Assessing riparian vegetation structure and the influence of land use using landscape metrics and geostatistical tools. Landsc Urban Plan 99:166–77.

Fischer MM, Getis A. 2009. Handbook of applied spatial analysis: software tools, methods and applications. Springer Science & Business Media

Foody GM. 2002. Status of land cover classification accuracy assessment. Remote Sens Environ 80:185–201. http://www.utsa.edu/lrsg/Teaching/ES6973/accuracy2.pdf. Last accessed 20/03/2017


31

Getis A, Ord JK. 2010. The Analysis of Spatial Association by Use of Distance Statistics. Perspect Spat Data Anal:127–45. http://link.springer.com/10.1007/978-3-642-01976-0_10. Last accessed 06/12/2017

Goetz SJ, Hansen M, Harris NL, Goldman E, Gabris C, Nordling J, Minnemeyer S, Ansari S, Lippmann M, Bennett L, Raad M, Potapov P. 2017. Using spatial statistics to identify emerging hot spots of forest loss. Environ Res Lett. http://iopscience.iop.org/article/10.1088/1748-9326/aa5a2f/pdf. Last accessed 06/12/2017

Harris NL, Goldman E, Gabris C, Nordling J, Minnemeyer S, Ansari S, Lippmann M, Bennett L, Raad M, Hansen M. 2017. Using spatial statistics to identify emerging hot spots of forest loss. Environ Res Lett 12:24012.

Hayes MM, Miller SN, Murphy MA. 2014. High-resolution landcover classification using random forest. Remote Sens Lett.

Hollenhorst P, Host E, Johnson B. 2006. Scaling issues in mapping riparian zones with remote sensing data: quantifying errors and sources of uncertainty. In: Scaling and Uncertainty Analysis in Ecology. Springer. pp 275–95.

Homer C, Fry J. 2012. The National Land Cover Database. US Geol Surv Fact Sheet:1–4. http://pubs.usgs.gov/fs/2012/3020/

Immitzer M, Atzberger C, Koukal T. 2012. Tree species classification with Random forest using very high spatial resolution 8-band worldView-2 satellite data. Remote Sens.

Iverson LR, Szafoni DL, Baum SE, Cook E a. 2001. A riparian wildlife habitat evaluation scheme developed using GIS. Environ Manage 28:639–54. http://www.ncbi.nlm.nih.gov/pubmed/11568844

Jones K, Slonecker E, Nash S, Neale C, Wade G, Hamann S. 2010. Riparian habitat changes across the continental United States (1972-2003) and potential implications for sustaining ecosystem services. Landsc Ecol 25:1261–75.

Ke Y, Quackenbush LJ, Im J. 2010. Synergistic use of QuickBird multispectral imagery and LIDAR data for object-based forest species classification. Remote Sens Environ.

Lee S, Im J, Kim J, Kim M, Shin M, cheol Kim H, Quackenbush LJ. 2016. Arctic Sea Ice Thickness Estimation from CryoSat-2 Satellite Data Using Machine Learning-Based Lead Detection. Remote Sens.

Lerch RN, Lin CH, Goyne KW, Kremer RJ, Anderson SH. 2017. Vegetative Buffer Strips for Reducing Herbicide Transport in Runoff: Effects of Buffer Width, Vegetation, and Season. JAWRA J Am Water Resour Assoc 53:667–83.

Lowrance R, Altier LS, Newbold JD, Schnabel RR, Groffman PM, Denver JM, Correll DL, Gilliam JW, Robinson JL, Brinsfield RB. 1997. Water quality functions of riparian forest buffers in Chesapeake Bay watersheds. Environ Manage 21:687–712.


32

Martino F. 2012. Description and inventory of the stockport creek watershed. http://www.stockportwatershed.org/docs/Stockport_Watershed_Description_and_Inventory.pdf

Moran PAP. 1950. Notes on Continuous Stochastic Phenomena. Biometrika 37:17–23. http://www.jstor.org/stable/2332142

Mulamoottil G. 1996. Wetlands: environmental gradients, boundaries, and buffers. CRC Press

NYS DEC. 2015. Hudson River Estuary Action Agenda 2015-2020. https://www.dec.ny.gov/docs/remediation_hudson_pdf/dhreaa15.pdf

NYS DEC. 2018. Riparian Buffers. NYS DEC. https://www.dec.ny.gov/chemical/106345.html. Last accessed 31/05/2018

Salo JA, Theobald DM. 2016. A Multi-Scale, Hierarchical Model to Map Riparian Zones. River Res Appl 22:1085–95. http://doi.wiley.com/10.1002/rra.1112%5Cnpapers2://publication/doi/10.1002/rra.1112

Salo JA, Theobald DM, Brown TC. 2016. Evaluation of Methods for Delineating Riparian Zones in a Semi-Arid Montane Watershed. J Am Water Resour Assoc 52:632–47.

Shandas V, Alberti M. 2009. Exploring the role of vegetation fragmentation on aquatic conditions: Linking upland with riparian areas in Puget Sound lowland streams. Landsc Urban Plan 90:66–75.

Warner TA, Foody GM, Nellis MD. 2009. The SAGE Handbook of Remote Sensing. Sage Publications https://books.google.com/books?id=zk5656l9ulkC&pgis=1

Webber D, Mickelson SK, Richard TL, Ahn HK. 2015. Vegetative buffer and fly ash pad surface material system application for reducing runoff, sediment and nutrient losses from livestock manure windrow composting facilities. Int J Energy, Environ Econ 23:641.

Weller DE, Baker ME. 2014. Cropland riparian buffers throughout Chesapeake Bay Watershed: Spatial patterns and effects on nitrate loads delivered to streams. J Am Water Resour Assoc 50:696–712. http://doi.wiley.com/10.1111/jawr.12207

Weller DE, Baker ME, Jordan TE. 2011. Effects of riparian buffers on nitrate concentrations in watershed discharges: New models and management implications. Ecol Appl 21:1679–95.

Wickham JD, Stehman S V, Gass L, Dewitz J, Fry JA, Wade TG. 2013. Accuracy assessment of NLCD 2006 land cover and impervious surface. https://ac.els-cdn.com/S0034425712004579/1-s2.0-S0034425712004579-main.pdf?_tid=7febc0ec-dafa-11e7-88e0-00000aab0f26&acdnat=1512615688_4e901c4e034f160ca338754523b8f2e2. Last accessed 06/12/2017

Documents

Geospatial ass essment of riparian zzones: A case study in ......ords: Ripari dation tial ass son Ri r.edu, PhD esf.edu, A C ion and prot ey knowledg s in New Yo and quanti sub-watersh