Mapping Russian Olive Using Remote Sensing to map an

Document Sample
Mapping Russian Olive Using Remote Sensing to map an Powered By Docstoc
					                   Mapping Russian Olive
Using Remote Sensing to map an Invasive Tree

September 2006                                           RSAC-0087-RPT1




                                                     1
   United States
   Department of     Forest    Remote Sensing
   Agriculture       Service   Applications Center
Abstract
Hamilton, R.; Megown, K.; Lachowski, H.; Campbell, R. 2006. Mapping Russian olive: using remote sensing to map an
invasive tree. RSAC-0087-RPT1. Salt Lake City, UT: U.S. Department of Agriculture Forest Service, Remote Sensing
Application Center. 7 p.

With funding from the Remote Sensing Steering Committee, a pilot project was initiated to develop a cost-effective method
for mapping Russian olive (Elaeagnus angustifolia L.), an invasive tree species, from scanned large-scale aerial photographs. A
study area was established along a riparian zone within a semiarid region of the Fishlake National Forest, located in central
Utah. Two scales of natural color aerial photographs (1:4,000 and 1:12,000) were evaluated as part of the project. Feature
Analyst, an extension for ArcGIS and several image processing software packages, was used to map the invasive tree. Overall,
Feature Analyst located Russian olive (RO) throughout the imagery with a relatively high degree of accuracy. For the map
derived from 1:4,000-scale photographs, the software correctly located the tree in 85 percent of all 4-by-4 meter transect cells
where Russian olive was actually present. However, smaller trees were sometimes missed and the size of trees and groups of
trees were frequently underestimated. The map derived from 1:4,000-scale photographs was only slightly more accurate than
the map derived from 1:12,000-scale photographs, suggesting that the smaller scale photography may be adequate for mapping
Russian olive.

Key Words
Russian olive, invasive species, remote sensing, Feature Analyst, large scale photos, accuracy assessment


Authors
Randy Hamilton is an Entomologist and Remote Sensing Specialist working at the Remote Sensing Applications Center
and employed by RedCastle Resources.

Kevin Megown is a Senior Project Leader and Biometrician working at the Remote Sensing Applications Center and employed
by RedCastle Resources.

Henry Lachowski is Program Leader for the Integration of Remote Sensing Program at the Remote Sensing Applications
Center in Salt Lake City, Utah.

Robert B. Campbell is an Ecologist working at the Fishlake National Forest in the Intermountain Region in Richfield, Utah.




                                                                ii
                         Table of Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii


Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1


Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2


Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . .4


Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6


Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7


References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7




                 iii
iv
Introduction
Invasive plants (weeds) have infested
hundreds of millions of acres of forest,
rangeland, and grassland throughout the
United States and pose a serious threat
to ecosystem health and function,
biodiversity, and endangered species
(USDA Forest Service 2004). Annually,
invasive plants invade an estimated
700,000 hectares of U.S. wildlife
habitat (Babbitt 1998).

The best defenses against invasive weeds
are prevention, early detection, and
eradication. However, once populations
become well established, management
objectives shift from eradication to
containment—establishing a perimeter
around the infestation, eliminating
small outlying patches, and gradually
reducing the perimeter of the primary
infestation. At this stage, mapping
established populations becomes critical.
An accurate map provides baseline             Figure 1—Russian olive, an invasive tree, invades riparian environments in semiarid
information and becomes a strategic           regions of the western United States. Photo courtesy of J. Scott Peterson, USDA NRCS,
planning tool for future management           www.forestryimages.org
efforts. Combating invasive weeds
without accurate baseline maps has been      Within the Fishlake National Forest          Casady and others 2005; Lass and
likened to fighting wildfires without        (FLNF), located in central Utah,             others 2005; Maheu-Girouxa and de
knowledge of their locations (Dewey          Russian olive invaded and became well-       Blois 2005). Biological and
1995). Unfortunately, the cost of            established in riparian zones of the more    phenological characteristics that
traditional field-based mapping can, in      arid regions of the Forest. Because this     distinguish a weed from its
some cases, be prohibitive.                  tree was largely ignored until recently,     surroundings, as viewed from above, are
                                             when Sevier County designated it as a        critical for successful mapping using
Russian olive (Elaeagnus angustifolia L.),   noxious weed, FLNF had not mapped            remote sensing. Russian olive on the
a thorny shrub or tree with origins in       the location and extent of the invasion.     Fishlake National Forest was considered
Southeastern Europe and Western Asia,        The Fishlake National Forest is              a good candidate for mapping because
was intentionally introduced and planted     proactive and has an aggressive weed         of the distinctive silver-gray color of its
in the United States for windbreaks,         management program. Nevertheless,            leaves, petioles, and current-year
erosion control, wildlife habitat, and       budgetary constraints severely limit the     branchlets (figure 2) (Katz and Shafroth
other horticultural purposes (figure 1)      scope of their weed management               2003). In addition, Russian olive trees
(Katz and Shafroth 2003). The tree is        activities. The need to map Russian          are large compared to other invasive
particularly well adapted to semiarid        olive placed additional strain on the        weeds and occur in dense stands. On
and saline environments. Early in the        Forest’s limited budget.                     the Fishlake National Forest, these trees
20th century, Russian olive escaped                                                       are generally confined to sparsely
cultivation and spread, particularly into    In 2005, the USDA Forest Service             vegetated areas along riparian corridors,
moist riparian environments in arid or       Remote Sensing Steering Committee            which further increases the ease of
semiarid regions of the western United       awarded funding for a proposal submitted     mapping them using remote sensing.
States (Stannard and others 2002).           by the Fishlake National Forest to, in
Although the invasive nature of Russian      part, develop a cost-effective remote        Specific objectives of this study were to
olive was known years ago, public and        sensing alternative to field surveys for     evaluate and compare the utility of two
private agencies continued to promote        mapping Russian olive.                       scales of scanned natural color aerial
planting it for various purposes as                                                       photography for mapping Russian olive
recently as the 1990s, and it is still        Remote sensing has been used with           infestations within a pilot study area on
commercially available (Katz and             varying degrees of success to map weed       the Fishlake National Forest.
Shafroth 2003).                              infestations (Hunt and others 2003;
                                                          1 | RSAC-0087-RPT1
                                                                                           process of noting a few features that
                                                                                           were correctly identified and others that
                                                                                           were incorrectly identified; digitizing
                                                                                           additional samples when features were
                                                                                           missed in the classification; and finally,
                                                                                           reprocessing the image. This iterative
                                                                                           process continues until a satisfactory
                                                                                           result is achieved. Although the output
                                                                                           from Feature Analyst can be remarkably
                                                                                           good, some manual editing may be
                                                                                           required to produce the final map.

                                                                                           Feature Analyst provides the user with
                                                                                           several parameters to fine-tune the
                                                                                           classifier to achieve optimal results.
                                                                                           First, the user specifies a kernel or input
                                                                                           representation, which approximates the
                                                                                           size and shape of the feature of interest.
                                                                                           To accommodate the circular shape of
                                                                                           the Russian olive tree crowns, we used a
  Figure 2—The distinct silver/gray coloration of Russian olive foliage can distinguish    circle input representation with a seven-
  it from its surroundings, making it a good target for mapping using remote sensing.      pixel diameter for the 1:4,000-scale
  Photo courtesy of Paul Wray, Iowa State University, www.forestryimages.org.              photography and a five-pixel diameter
                                                                                           for the 1:12,000-scale photography
                                                                                           (figure 4).

Methods                                       recognize features of interest through an    Because Feature Analyst currently
                                              iterative “training” process. First, the     cannot process data sets larger than 2
A pilot study area was established            user digitizes a few sample features.        GB, the imagery was cropped,
within the Fishlake National Forest           Using the spectral and spatial               eliminating some areas unsuitable for
along a six-mile segment of Salina            information contained in the samples,        Russian olive establishment, then subset
Creek, just east of Richfield, Utah           Feature Analyst attempts to identify         into sections slightly less than 2 GB in
(figure 3). Natural color aerial              similar features in the imagery. The user    size. To increase processing speed, a
photographs were acquired at 1:4,000          then proceeds through an iterative           resampling factor was set, which
and 1:12,000 scales over the area on 11
September 2005 using a Zeiss RMK-A
camera equipped with an 80mm lens.
Negatives were subsequently scanned at
14 microns yielding 5.6 and 16.8
centimeter (0.18 and 0.55 foot) pixels
for the respective scales. The digital
images were then orthorectified using
the Leica Photogrammetry Suite.

For this study, Feature Analyst 4.0, an
extension for ArcGIS, ERDAS Imagine,
and other image processing software
packages, was used to map Russian
olive. The image classification method
implemented in Feature Analyst is
known as feature extraction. Unlike
traditional image classification, feature
extraction uses both spectral and spatial       Figure 3—The study area, a six-mile section of Salina Creek, is located east of
properties of an image to identify the          Richfield, Utah within the Fishlake National Forest. Natural color 1:4,000-scale
                                                scanned and orthorectified aerial photographs of the study area are shown overlaid
feature of interest (tree crowns in our
                                                on a shaded relief of the area.
case). Feature Analyst “learns” to
                                                            2 | RSAC-0087-RPT1
                                             imagery, no field visits were made prior        each transect. When Russian olive was
                                             to classifying the imagery. However, we         encountered, the tree was outlined
                                             visited the site as part of an accuracy         directly on the transect datasheet. Back
                                             assessment. To assess the accuracy of the       at the office, outlines drawn on the
                                             map, the study area was sampled using a         transect datasheets were manually
                                             transect design. First, the study area was      digitized over the imagery in ArcGIS.
                                             divided into 1-kilometer segments. Line         The digital transects were intersected
                                             transect (100-meters in length by 4-            with the digitized Russian olive tree
                                             meters wide, subdivided in 4-meter              outlines and with the Feature Analyst
                                             segments) were randomly located on the          classification of Russian olive (figure 5).
                                             imagery within the riparian zone of             The percent composition of digitized
                                             every other 1-kilometer segment (figure         and classified Russian olive was
                                             5). Transects were not allowed to cross         computed for each 4-by-4 meter
                                             the stream due to high water. Because           transect cell. These data were then
                                             of this constraint and because the              categorized using three different scales
 Figure 4—A circle input representation,     riparian zone was relatively narrow             (for comparison). (A) the Daubenmire
 7 pixels in diameter, was used to
                                             (usually less than 100-meters on either         cover class scale, frequently used for
 accommodate the circular shape of
                                             side of the stream), transects generally        collecting ecological data (0, <1, 1–5,
 Russian olive tree crowns in the 1:4,000-
 scale imagery.                              paralleled the stream. The orientations         5–25, 25–50, 50–75, 75–95, and 95–
                                             of transects were fixed at some whole           100 percent); (B) 10-percent intervals
                                             number multiple of 22.5 degrees from            or classes (0, 0–10, 10–20, 20–30, 40–
                                             north. For each transect, a field               50, 50–60, 60–70, 70–80, 80–90, and
effectively allows the software to           datasheet with transect overlaid on the         90–100 percent); and (C) present/
resample the imagery to a larger pixel       imagery was printed on a color laser            absent intervals (classes). Contingency
size, thereby reducing the amount of         printer.                                        tables comparing the Feature Analyst
data that is processed. For this study,                                                      classification to the field observations
resampling factors of 4 and 2 were set       In July 2006, transects were located on         were created for the three different data
respectively for the 1:4,000 and             the ground using a combination of GPS           categorizations for each scale of imagery
1:12,000-scale data. Default settings        (global positioning system) units, transect     as well as for the unedited and manually
were accepted for all other parameters.      locator maps, and the individual field          edited maps.
The processing time was further reduced      datasheets. We walked the length of
by creating a polygon mask to eliminate
other regions of the imagery not suited
for Russian olive establishment.

The time required to train Feature
Analyst can be reduced by training first
on a small representative subset of the
imagery, then applying the resulting
model to the rest of the imagery. We
evaluated this approach in several areas,
but the results were generally
unsatisfactory. Therefore, Feature
Analyst was trained using an entire 2
GB section of imagery and retrained for
each additional section.

After Russian olive was mapped by
Feature Analyst, the maps for the two
scales of imagery were manually edited
to eliminate obvious errors in the
classification.
                                               Figure 5—A 100-meter accuracy assessment transect overlaid on 1:4,000-scale
Because of the high spatial resolution of      imagery. Transects were intersected with the Feature Analyst classification of Russian
the imagery and the relative ease of           olive (blue) and the digitized polygons of Russian olive identified in the field (yellow).
recognizing Russian olive on the
                                                           3 | RSAC-0087-RPT1
Producer’s, user’s, and overall accuracy               things as Russian olive. These errors of                 Results and
were computed from each contingency
table as well as the kappa statistic
                                                       commission are addressed by the user’s
                                                       accuracy. A high user’s accuracy, for
                                                                                                                Discussion
(Lillesand and Kiefer 1994). To                        example, tells us that almost everything                 The final output from Feature Analyst
accommodate measurement errors                         mapped as Russian olive is Russian olive                 consisted of sets of polygons outlining
resulting from sketching and digitizing                and that the analyst mislabeled very few                 Russian olive tree crowns or clumps of
the field data, the accuracy statistics                other things as Russian olive.                           trees (figure 6). Depending on the
were also computed using a 10-percent                                                                           specific data categorization and whether
fuzzy tolerance applied to the 10-                     The kappa statistic measures the extent                  or not the map was edited, overall
percent interval categorization. In other              to which the accuracies are due to true                  accuracies ranged from 70–91 percent
words, we assumed that the true                        agreement between the map and field                      for the 1:4,000-scale imagery and 67–
percentage of Russian olive in a cell                  verification data versus chance agreement.               88 percent for the 1:12,000-scale
could deviate from its computed                        Kappa typically ranges between 0 and 1.                  imagery, with respective kappa statistics
percent-RO interval by as much as 10                   A high kappa value indicates that the                    ranging from 0.49–0.81 and 0.39–0.73
percent, or one 10 percent interval                    classification is much better than could                 (table 1). In general, the accuracy of the
above or below the computed interval.                  be achieved by chance alone.                             map derived from 1:4,000-scale imagery
For example, if the percent Russian
olive computed from digitized field data
was in the 80–90 percent interval, then
we assumed that the true value could
fall within the 70–80, 80–90, or 90–
100 percent intervals.

Producer’s and user’s accuracies are
measures of accuracy computed for each
map class (e.g. each percent-RO class)
in a classification, while overall accuracy
is a summary statistic of the accuracies
of all classes combined. Fundamentally,
producer’s accuracy tells us how well the
analyst mapped Russian olive actually
found in the transects. For example, a
high producer’s accuracy indicates that
the analyst’s map correctly identified
most Russian olive trees or, in other
words, very few Russian olive trees were
omitted or missed by the analyst’s map.
However, producer’s accuracy does not                     Figure 6—Map of Russian olive (blue outlines) created from 1:4,000-scale scanned
tell us whether the analyst committed                     aerial photographs using Feature Analyst.
errors by incorrectly labeling other

   Table 1—Overall accuracies and kappa statistics for the edited and unedited Russian olive classifications derived from 1:4,000
   and 1:12,000-scale imagery for each percent-RO categorization, as well as the fuzzy assessment of the 10-percent interval
   categorization.

                                        1:4,000 Edited                1:4,000 Unedited                  1:12,000 Edited             1:12,000 Unedited
        Categorization              Overall                         Overall                         Overall                         Overall
                                                     Kappa                           Kappa                           Kappa                          Kappa
                                   (Percent)                       (Percent)                       (Percent)                       (Percent)
    (A) Daubenmire classes         75              0.55            73              0.53            71              0.44           68              0.42
    (B) 10 percent intervals*      74 (41)         0.52 (0.33)     70 (40)         0.49 (0.32)     69 (29)         0.42 (0.19)    67 (28)         0.39 (0.18)
    (B) 10 percent intervals*
                                   86 (71)         0.85 (0.60)     82 (71)         0.81 (0.61)     80 (54)         0.79 (0.37)    77 (54)         0.76 (0.37)
    (fuzzy tolerance)
    (C) Present/ absent            91              0.81            88              0.75            88              0.73           86              0.69
  * Overall accuracies and kappa statistics are presented both with and without (in parentheses) the 0-percent or absent class. The absent class had a
  disproportionately high

                                                                        4 | RSAC-0087-RPT1
was better than that of the map derived       how the field data were digitized and         disproportionate number of samples on
from the 1:12,000-scale imagery.              in how Feature Analyst drew the               overall accuracy and the kappa statistic
Nevertheless, the increase in accuracy        polygons. Because of this, we imposed         is illustrated in table 1, where these
was modest, suggesting that 1:12,000-         a 10-percent fuzzy tolerance to the 10-       statistics are presented for the 10-
scale photography may be adequate for         percent interval categorization               percent categorizations both with and
mapping Russian olive in many cases.          (categorization B) accuracy assessment        without the absent category. Applying
                                              data.                                         the fuzzy tolerance and removing the
Except for intervals (classes) containing                                                   “absent” category increased the
either no Russian olive or a very high        In addition to imposing the fuzzy             producer’s and user’s accuracies of the
percentage of Russian olive, producer’s       tolerance, we also eliminated the 0-          percent-RO intervals substantially,
and user’s accuracies were quite low,         percent or “absent” category from the         (compare tables 3 and 4).
ranging from only 3–46 percent and            fuzzy tolerance contingency table. This
4–48 percent respectively for the             category had an exceptionally high            The overall ability of Feature Analyst to
Daubenmire and 10-percent (non-               number of samples (n=889) compared            locate Russian olive was assessed by
fuzzy) categorizations (tables 2 and 3).      to the other categories (n<105), which        collapsing the contingency tables to two
In many cases, it appeared that the low       confounded the accuracies (particularly       classes—present and absent. For both
accuracies for intermediate values of         the overall accuracies) and kappa             scales of imagery, overall accuracies were
percent-RO were due to variability in         statistics. The influence of the              in excess of 85 percent (table 1). User’s


  Table 2—User’s and producer’s accuracies for Daubenmire percent-RO intervals (categorization A) for edited and non-edited RO
  classifications derived from 1:4,000 and 1:12,000-scale imagery.

    Percent-RO            1:4,000 Edited             1:4,000 Unedited               1:12,000 Edited         1:12,000 Unedited
     Intervals        User’s      Producer’s       User’s     Producer’s         User’s    Producer’s      User’s     Producer’s
        Absent         92%            95%           92%            90%            87%          96%           88%           91%
          <1            5%            5%            11%            14%             6%          10%           6%            10%
        1–4.9          13%            19%           11%            16%            12%          8%            10%           8%
        5–24.9         35%            35%           34%            41%            27%          27%           25%           29%
       25–49.9         40%            36%           36%            36%            28%          21%           24%           22%
       50–74.9         45%            43%           42%            46%            16%          11%           19%           14%
       75–94.9         43%            41%           48%            43%            29%          29%           27%           27%
        95–100         85%            57%           85%            66%            76%          45%           76%           45%


  Table 3—User’s and producer’s accuracies for 10 percent-RO intervals (categorization B) for edited and non-edited RO classifications
  derived from 1:4,000 and 1:12,000-scale imagery.

    Percent-RO            1:4,000 Edited            1:4,000 Unedited               1:12,000 Edited          1:12,000 Unedited
     Intervals        User’s     Producer’s       User’s      Producer’s         User’s   Producer’s       User’s     Producer’s
        Absent         92%           95%            92%           90%             87%         96%           88%            91%
       0.01–9.9        40%           42%            33%           40%             32%         29%           29%            29%
       10–19.9         21%           23%            20%           27%             17%         23%           14%            23%
       20–29.9         19%           19%            19%           19%             14%         10%           12%            13%
       30–39.9         16%           16%            15%           16%             7%           5%            6%            5%
       40–49.9         18%           17%            16%           19%             11%          8%           10%            8%
       50–59.9         29%           27%            21%           23%             5%           4%            4%            4%
       60–69.9         25%           24%            23%           24%             7%           3%            6%            3%
       70–79.9         19%           15%            18%           15%             10%         11%           10%            11%
       80–89.9         30%           29%            30%           29%             9%           9%            6%            6%
       90–100          87%           65%            82%           67%             73%         47%           73%            47%


                                                            5 | RSAC-0087-RPT1
  Table 4—User’s and producer’s accuracies computed with a 10 percent fuzzy tolerance for 10 percent-RO intervals (categorization
  B) with the absent category removed for edited and non-edited RO classifications derived from 1:4,000 and 1:12,000-scale imagery.

     Percent-RO          1:4,000 Edited            1:4,000 Unedited                1:12,000 Edited           1:12,000 Unedited
      Intervals      User’s     Producer’s       User’s      Producer’s          User’s     Producer’s     User’s       Producer’s
       0.01–9.9        71%          89%           78%             85%             53%          79%           54%           77%
       10–19.9         68%          82%           73%             80%             51%          71%           51%           72%
       20–29.9         60%          72%           65%             72%             46%          55%           44%           57%
       30–39.9         53%          57%           52%             57%             39%          41%           36%           42%
       40–49.9         54%          56%           51%             55%             33%          33%           33%           31%
       50–59.9         52%          55%           48%             55%             30%          29%           31%           29%
       60–69.9         59%          54%           55%             59%             32%          25%           34%           25%
       70–79.9         69%          57%           68%             62%             55%          35%           55%           35%
       80–89.9         90%          73%           88%             76%             75%          59%           75%           58%
       90–100          96%          78%           96%             80%             85%          65%           85%           64%

   Table 5—User’s and producer’s accuracies for a present/absent categorization (categorization C) for edited and non-edited RO
   classifications derived from 1:4,000 and 1:12,000-scale imagery.

    Percent-RO           1:4,000 Edited            1:4,000 Unedited                1:12,000 Edited           1:12,000 Unedited
     Intervals       User’s     Producer’s       User’s      Producer’s          User’s     Producer’s      User’s      Producer’s
        Present        91%          85%            82%            86%             90%          74%           83%           77%
        Absent         92%          95%            92%            90%             87%          96%           88%           91%




and producer’s accuracies ranged              was 15 or 25 percent higher than the            other cases it may not.
between 74 and 96 percent, depending          edited area mapped by Feature analyst
on the scale of the imagery and whether       from the 1:4,000 and 1:12,000-scale             For reference, the two scales of scanned
the map was edited (table 5). For the         photography respectively.                       photography for this project were
edited map derived from the 1:4,000-                                                          acquired at a cost of approximately
scale imagery, Feature Analyst correctly                                                      $5,100. Acquiring a single scale of
identified Russian olive in 85 percent of     Costs                                           photography over a larger area would
transect cells where it occurred, while       The expenses involved with mapping an           decrease the relative cost (e.g., cost per
91 percent of cells where Feature             invasive plant using remote sensing can         acre) of acquisition. It is estimated that
Analyst mapped Russian olive contained        vary widely depending on the specific           three full-time person weeks would be
Russian olive.                                objectives, imagery requirements, vendor        required to processes the scanned
                                              availability, location, availability of an      imagery, with an additional 1–2 person
Overall, Feature Analyst located Russian      in-house analyst, the analyst’s level of        weeks to complete an accuracy
olive throughout the imagery with a           experience, and a variety of other factors.     assessment. At a cost of $300 per day,
relatively high degree of accuracy.           In addition, the per-acre cost of aerial        processing would cost $4,500, with an
However, some trees (especially smaller       photography is scale-dependent and              additional $3,000 for an accuracy
trees) were occasionally not identified       decreases with increasing acreage. Our          assessment. Using these estimates, the
by Feature Analyst. Also, the polygons        pilot study area was very small, yielding       total cost, including imagery,
created by Feature Analyst often              a high cost per acre. Because of the            processing, and accuracy assessment is
underestimated the actual area occupied       many variables that can affect the cost,        estimated at $12,600. The Fishlake
by the particular tree or group of trees.     the economics and feasibility of any            National Forest estimated that doing a
In other words, errors of omission were       proposed weed mapping project should            ground survey to map this same region
usually greater than errors of                be carefully evaluated before selecting a       of Salina creek would cost between
commission. This is illustrated by the        specific mapping method. In some cases,         $12,000 and $14,000.
fact that within transects, the total area    remote sensing may prove to be the best
of Russian olive observed in the field        and most economical approach, while in          Compared to field surveys, one benefit

                                                            6 | RSAC-0087-RPT1
of mapping Russian olive from aerial          References                                                For additional information, contact:
                                                                                                        Henry Lachowski
photography is that the image                                                                           Remote Sensing Applications Center
processing and analysis can be done           Babbitt, B. 1998. Statement by Secretary of the           2222 West 2300 South
                                              Interior on invasive alien species. Proceedings,          Salt Lake City, UT 84119
during winter months. This frees field
                                              National Weed Symposium, BLM Weed Page. April
crews for other projects during summer        8-10.
                                                                                                        phone: 801-975-3750
months and allows the bulk of the work                                                                  e-mail: hlachowski@fs.fed.us.
to be done during the winter when             Casady, G.M.; Hanley, R.S.; Seelan, S.K. 2005.
personnel are less busy.                      Detection of leafy spurge (Euphorbia esula) using
                                                                                                        This publication can be downloaded from the
                                              multidate high-resolution satellite imagery. Weed
                                                                                                        RSAC Web sites: http://fsweb.rsac.fs.fed.us
                                              Technology 19:462-467.
                                                                                                        and http://www.fs.fed.us/eng/rsac
Conclusions                                   Dewey, S.A.; Jenkins, M.J.; Tonioli, R.C. 1995.
Russian olive within riparian areas of        Wildfire suppression - a paradigm for noxious weed
semiarid regions of the Fishlake              management. Weed Technology 9:621-627.
                                                                                                        The Forest Service, United States
National Forest proved to be a good                                                                     Department of Agriculture (USDA), has
                                              Hunt, E.R. Jr.; Everitt, J.H.; Ritchie, J.C.; Moran, M.
target for mapping using large-scale,         S.; Booth, D.T.; Anderson, G.L.; Clark, P.E.;
                                                                                                        developed this information for the guidance
                                                                                                        of its employees, its contractors, and its
scanned, natural color aerial                 Seyfried, M.S. 2003. Applications and research
                                                                                                        cooperating Federal and State agencies
photography. In particular, the silver/       using remote sensing for rangeland management.
                                                                                                        and is not responsible for the interpretation
                                              Photogrammetric Engineering and Remote Sensing
gray foliage of Russian olive coupled         69:675-693.
                                                                                                        or use of this information by anyone except
                                                                                                        its own employees. The use of trade, firm,
with its large size and clumped
                                                                                                        or corporation names in this document is
distribution distinguished the tree from      Katz, G.L.; Shafroth, P.B. 2003. Biology, ecology         for the information and convenience of
its surroundings. Also, its preference for    and management of Elaeagnus angustifolia L.               the reader. Such use does not constitute
                                              (Russian olive) in western North America. Wetlands        an official evaluation, conclusion,
moist riparian areas allowed us to
                                              23:763-777.                                               recommendation, endorsement, or approval
simplify the image processing by                                                                        by the Department of any product or
excluding non-riparian areas from             Lass, L.W.; Prather, T.S.; Glenn, N.F.; Weber, K.T.;
                                                                                                        service to the exclusion of others that
                                                                                                        may be suitable.
further analysis.                             Mundt, J.T.; Pettingill, J. 2005. A review of remote
                                              sensing of invasive weeds and example of the
                                              early detection of spotted knapweed (Centaurea
Feature Analyst proved to be an               maculosa) and babysbreath (Gypsophila paniculata)
effective tool for mapping Russian olive.     with a hyperspectral sensor. Weed Science
                                                                                                        The U.S. Department of Agriculture (USDA)
Overall, the software was able to locate      53:242-251.
                                                                                                        prohibits discrimination in all its programs
Russian olive with a high degree of                                                                     and activities on the basis of race, color,
                                              Lillesand, T.M.; Kiefer, R.W. 1994. Remote                national origin, age, disability, and where
accuracy; however, smaller trees were
                                              Sensing and Image Interpretation, 3rd ed. John            applicable, sex, marital status, familial
sometimes missed and the size of trees        Wiley and Sons, Inc. New York, NY. 750 pp.                status, parental status, religion, sexual
and groups of trees was frequently                                                                      orientation, genetic information, political
                                                                                                        beliefs, reprisal, or because all or part of
underestimated. Training the classifier       Maheu-Girouxa, M.; de Blois, S. 2005. Mapping the         an individual’s income is derived from
on a subset of the imagery and applying       invasive species Phragmites australis in linear           any public assistance program. (Not all
                                              wetland corridors. Aquatic Biology 83:310-320.
the model to the rest of the imagery in a                                                               prohibited bases apply to all programs).
                                                                                                        Persons with disabilities who require
batch processing mode did not produce
                                              Stannard, M.; Ogle, D.; Holzworth, L.; Scianna J.;        alternative means for communication of
reliable results. Therefore, the classifier   Sunleaf, E. 2002. History, biology, ecology,              program information (Braille, large print,
had to be retrained for each section of       suppression and revegetation of Russian-olive sites       audiotape, etc.) should contact USDA’s
                                                                                                        TARGET Center at (202) 720–2600
imagery, which is a time-consuming            (Elaeagnus angustifolia L.). Plant Materials
                                              Technical Note No. 47. Boise, ID. U.S. Department         (voice and TDD). To file a complaint
process.                                      of Agriculture, Natural Resources Conservation            of discrimination, write to USDA, Director,
                                              Service, 14 p.                                            Office of Civil Rights, 1400 Independence
                                                                                                        Avenue, S.W., Washington, D.C. 20250–
                                                                                                        9410, or call (800) 795–3272 (voice) or
                                              U.S. Department of Agriculture, Forest Service.           (202) 720–6382 (TDD). USDA is an equal
                                              2004. National Strategy and Implementation Plan           opportunity provider and employer.
                                              for Invasive Species Management. Rep. FS-805.
                                              Washington, DC: U.S. Department of Agriculture,
                                              Forest Service. 17 p.




                                                               7 | RSAC-0087-RPT1
8 | RSAC-0087-RPT1