An Introduction to Satellite Imagery Analysis

for Land Managers

A step-by-step guide to downloading, interpreting, and

classifying satellite imagery using free software

Image Credit: USGS/NASA Landsat

 PEER!
REVIEWED
ALABAMA EXTENSION

James Cash, former Graduate Student, Nancy Loewenstein, Extension Specialist, Chris Anderson,
Associate Professor, Forestry and Wildlife Sciences, and Arnold “Beau” Brodbeck, Extension
Specialist, all with Auburn University
For more information, contact your county Extension office. Visit www.aces.edu/directory.
The Alabama Cooperative Extension System (Alabama A&M University and Auburn University) is an
equal opportunity educator and employer. Everyone is welcome! Please let us know if you have
accessibility needs. New March 2020, FOR-2072
© 2020 by the Alabama Cooperative Extension System. All rights reserved.
Table of Contents
Part 1. Introduction ............................................................................................................ 1
Land Management Applications .............................................................................. 1
Basic Concepts ...................................................................................................... 1
Part 2. Sentinel and Landsat Imagery ................................................................................ 3
Background and Specifications .............................................................................. 3
Visual Interpretation ................................................................................................ 4
Viewing Online ........................................................................................................ 4
Downloading Imagery ............................................................................................. 5
Part 3. QGIS: Free Open-Source Mapping Software ........................................................ 8
Download and Setup ............................................................................................. 8
Adding Data ........................................................................................................... 8
Creating and Editing Vector Layers ........................................................................ 10
Part 4. Supervised Land Cover Classification .................................................................... 12
Step 1: Create Band Set ........................................................................................ 12
Step 2: Set Training Sites ...................................................................................... 14
Step 3: Classify Land Cover .................................................................................. 16
Step 4: Complete an Accuracy Assessment ........................................................... 17
Step 5: Estimate Surface Area .............................................................................. 19
Part 5. Additional Resources and References ................................................................... 20
Satellite Imagery Access ........................................................................................ 20
Software Downloads ............................................................................................... 20
General Tutorials and User Guides ........................................................................ 20
Specific Topics ........................................................................................................ 20
References .............................................................................................................. 21
 Part 1. Introduction

Satellite imagery analysis and other remote and agriculture. Satellite imagery also can
sensing techniques have become an help managers monitor plant health over
integral component of land and natural large areas, such as detecting insect
resource management, particularly among damage or drought stress.
government agencies, NGOs, and private
Novel uses for satellite imagery and other
companies that can afford to hire remote
remote sensing products are continuously
sensing specialists. Smaller organizations,
being developed. As more land managers,
however, may lack the resources to staff a
foresters, and biologists begin using these
dedicated remote-sensing or GIS
tools, their fresh perspectives are likely to
(geographic information system) position.
help drive innovation.
Fortunately there are free tools available
that allow nonspecialists to access and Basic Concepts
analyze satellite imagery to help achieve
natural resource management objectives. Humans see the world based on light
reflecting off objects in the visible range of
This guide provides a brief introduction to
the electromagnetic spectrum (i.e., red,
the applications and basic concepts of
green, blue). Most consumer-grade
satellite imagery analysis. This is followed
cameras, such as those found in
by step-by-step instructions* on how to
smartphones, capture light within or near
access and analyze up-to-date imagery.
this range to produce images that are
The instructions are based on a case study
similar to what we see with the naked eye.
involving detection and mapping of the
invasive shrub Chinese privet (Ligustrum Satellite sensors used for land cover
sinense) in a bottomland hardwood forest. analysis sample light reflecting off Earth’s
surface in wavelengths outside the visible
Land Management Applications range. These multispectral sensors capture
more information about the surface of the
Broad applications exist for satellite imagery
planet than could be gathered using a
analysis in land and natural resource
management. Most of these fall under the regular red-green-blue (RGB) camera.
umbrella of land cover analysis. Knowing Sampling in the infrared portion of the
spectrum is particularly useful for mapping
the proportion and location of different land
vegetation because leaves tend to strongly
cover types on a property is useful when
reflect near-infrared radiation.
developing management plans and
monitoring the results of actions. Multispectral satellite sensors sample
discrete bands within the electromagnetic
Land cover types that land managers may
spectrum. The number of bands sampled
be particularly interested in mapping include
determines the spectral resolution. The data
forests, wetlands, invasive plant species,
gathered by the multispectral sensor can be

*Step-by-step instructions for software in this guide are based on the Windows operating
system. This workflow is also possible on MacOS X, Android, and Linux systems but may
require slight modifications.

1
used to construct spectral signatures for images in which each pixel in the image is
land cover types. classified into a specific land cover
category. To create a land cover map using
Spectral signatures are representations of supervised classification, the user starts by
how different land cover types reflect light
creating training sites. This is done by
with varying intensity across the manually labeling the land cover in a small
electromagnetic spectrum (figure 1). These
portion of the satellite image. The software
differences in reflectivity arise due to
then calculates the spectral signature of
differences in pigmentation, water content,
each land cover type based on the training
texture, and other factors. Satellites with
sites. Next, it calculates a spectral signature
higher spectral resolution (i.e., more bands)
for each pixel in the image, compares that
can produce more detailed spectral
signature to the signatures derived from the
signatures, which is helpful for
training areas, and sorts the pixel into the
differentiating between land cover types
land cover class that is the closest fit. This
with similar spectral characteristics.
process is covered in greater detail in Part 4
Remote sensing software can be used to of this publication.
create land cover maps from satellite

Figure 1. Example spectral signatures for bare soil, vegetation, and water. This example is based on center wavelengths of Sentinel 2
spectral bands. Modified from seos-project.eu, based on Siegmund and Menz (2005).

2
 Part 2. Sentinel and Landsat Imagery

Background and Specifications failure on Landsat 7, however, creates lines

in the imagery that can complicate
Multispectral satellite imagery is freely interpretation and classification; therefore,
available through several sources. The Landsat 8 is more frequently used.
United States’ Landsat series and the
European Space Agency’s (ESA) Sentinel 2 Sentinel 2 is a pair of satellites that
constellation are perhaps the most launched in 2015 and 2017. Together they
commonly used. have a 5-day temporal resolution. Both
satellites have 13 bands with three spatial
The first Landsat satellite was launched in resolutions: 10, 20, and 60 m (table 2).
1972. Now in its eighth generation, the
Landsat program provides an unparalleled The greater spatial, spectral, and temporal
data archive that is crucial for land cover resolution of Sentinel 2 tends to give it an
change analysis. The current generation advantage over the Landsat system for
(Landsat 8) has 11 bands with spatial many, but not all, land cover applications.
resolutions of 15, 30, and 100 meters (m) Higher spatial resolution allows finer grain
(table 1). Spatial resolution refers to the details to be discerned in the imagery. This
dimensions of the area on the ground that may or may not be important depending on
each pixel in the image represents (e.g., a the scale of the mapping project.
30 m resolution means that each pixel is 30 Greater temporal resolution increases the
´ 30 m). chances of acquiring cloud-free imagery
The temporal resolution of Landsat 8 is 16 within short windows of time, something that
days at the equator, meaning that new can be difficult in regions with long rainy
images are acquired roughly every 16 days seasons. This can be important when trying
for any given location. Areas near the poles to find imagery from a specific time period
are imaged more frequently due to the polar or when relying on near real-time updates
orbit of the satellite. The Landsat 7 is still in for deforestation or flood monitoring. The
operation, giving a combined 8-day greater temporal resolution also provides
temporal resolution for the pair. A sensor Table 2. Sentinel 2 Spectral Bands
(Sources: Congedo, 2016, and ESA, 2015)
Table 1. Landsat 8 Spectral Bands Band
Band Spatial resolution
Spatial(m)
resolution (m)
(Source: USGS) Band 1 - Coastal aerosol
Band 1 - Coastal aerosol
60
60
Band 2 - Blue 10
Band Spatial resolution (m) Band 2 - Blue
Band 3 - Green 10 10
Band 1 - Coastal aerosol 30 Band
Band 3
4 -- Red
Green 10 10
Band 4
Band 5 -- Vegetation
Red Red Edge 20 10
Band 2 - Blue 30
Band 6 - Vegetation Red Edge 20
Band 3 - Green 30 Band 5 - Vegetation Red
Band 7 - Vegetation Red Edge
Edge 20
20
Band 4 - Red 30 Band
Band 6
8 -- Near-infrared
Vegetation Red Edge 10 20
Band 5 - Near Infrared 30 Band
Band 7
8A- -Vegetation
Vegetation RedRed Edge
Edge 20 20
Band 8
Band 9 -- Water Vapor
Near-infrared 60 10
Band 6 - Shortwave Infrared 1 30 Band 10 - Shortwave Infrared (Cirrus) 60
Band 7 - Shortwave Infrared 2 30 Band
Band 8A
11 - -Shortwave
Vegetation Red Edge
Infrared 20 20
Band 8 - Panchromatic 15 Band
Band 9
12-- Water Vapor
Shortwave Infrared 20 60
Band 9 - Cirrus 30 Band 10 - Shortwave Infrared (Cirrus) 60
Band 10 - Thermal Infrared 1 100 Band 11 - Shortwave Infrared 20
Band 11 - Thermal Infrared 2 100 Band 12 - Shortwave Infrared 20

3
more data for time series analyses. As
covered previously, greater spectral
resolution increases the ability to distinguish
among spectrally similar land cover types.
Although Sentinel 2 has many benefits,
Landsat provides the crucial ability to
compare new imagery to a decades-long
archive. In addition, the coarser spatial
resolution can actually be beneficial when
working over very large areas due to the
smaller file size.

Visual Interpretation
Although software can utilize the full set of
bands from a multispectral sensor to
calculate spectral signatures, humans are
still limited to seeing combinations of red,
green, and blue. To visualize the spectral Figure 2. Comparison of natural color (top) vs. false color
infrared (bottom) dormant season Sentinel 2 imagery of a
bands from outside the visible range, we bottomland hardwood forest in Alabama. The false color
can assign those bands to one of the three infrared is particularly useful for distinguishing between
primary colors (red, green, blue), hence deciduous forests invaded by privet—an evergreen shrub—
and uninvaded forests. This is an L1C image with DOS1
creating a false color composite. atmospheric correction and gain increased to 4.0 to
maximize visual contrast. Modified Copernicus Sentinel data
False color composites are typically 2017/Sentinel Hub
designated using three numbers separated
by hyphens; the first number represents the have dropped their leaves, there is much
satellite band assigned to red, the second greater contrast. In our classification case
number to green, and the last number to study in Part 4 we show how phenological
blue. For Sentinel 2, a color combination of differences can be used to detect invasive
4-3-2 is a natural color image, whereas 8-4- evergreen shrubs in deciduous forests.
3 is an infrared false color composite (table Targeting specific time periods can also be
2) that highlights living, nondormant used to distinguish among crops with
vegetation in red/pink (figure 2). Other different planting and harvesting schedules.
combinations are useful for emphasizing
different land cover features, and it is
Viewing Online
typically a good idea to experiment with Sentinel 2 and Landsat 8 imagery can be
different band combinations to identify the viewed for free online using EO Browser at
best option for a given application. https://apps.sentinel-hub.com/eo-browser/.
EO Browser can be used to access up-to-
It is often useful to take vegetation
date and archived Sentinel 2 and Landsat 8
phenology into consideration to help
images as well as several other satellites.
distinguish between similar vegetation
Images are searchable by date and cloud
types. For instance, deciduous and
cover. Using EO Browser allows quick and
evergreen trees may have similar spectral
easy viewing of up-to-date or archived
signatures and be hard to distinguish during
imagery without needing to download large
the growing season. During the dormant
files.
season, however, when the deciduous trees

4
The first time you use EO Browser, a pop- a classified land cover image is necessary,
up should appear to provide a quick tutorial EO Browser can be used to scout for
on the basic features of the web page. You images that are of appropriate quality to
download and analyze further.
can click the button in the top right
corner to view the tutorial again. EO Browser has the capability to export
imagery; however, the exported images
To view imagery, start by zooming to your
tend to have inconsistent spatial resolutions.
area of interest on the base map. Choose
Consistency in spatial resolution is
the data source, date range, and max cloud
particularly important when using land cover
cover using the Search tab. Note there are
maps to track change over time, as is often
two options for Sentinel 2 imagery, L1C and
the goal. It may be better, therefore, to
L2A. These designations refer to the level of
download imagery from either the US
preprocessing that the images received.
Geological Survey (USGS) or the ESA. The
L2A images have been corrected by the
USGS website
ESA for atmospheric interference, such as
https://earthexplorer.usgs.gov provides data
haze, and tend to have greater color
archives for Landsat 8 and Sentinel 2 along
contrast. L2A images are not always
with a variety of other data sources. A free
available, however, due to data gaps;
account is required, which can be created
therefore, L1C sometimes must be used.
using the Register tab at the top right of the
Less sophisticated atmospheric corrections
web page.
can still be applied to L1C images if
necessary. To find appropriate imagery, begin by
zooming in on your area of interest using
Images are viewed by clicking on Visualize
the Google base map. Click on the map to
under the Results tab. In the visualization
draw a rectangle around your desired site
tab you can choose one of several pre-set
(each click sets a corner pin for your area of
band combinations to view the imagery. The
interest). If you used EO Browser to find a
Custom option allows you to create any
specific image, you can add that image date
band combination you choose. The to the mm/dd/yyyy fields under the Date
button provides additional visualization Range tab. If you still need to identify an
options. The Atmospheric Correction appropriate image, input a range of dates
drop-down menu allows you to apply a that are likely to provide reasonable
DOS1 or statistical correction to Sentinel 2 imagery. Too narrow a range limits the
L1C images, if desired. The Gain and chances that good cloud-free imagery will
Gamma sliders can be used to increase be available, while too wide a range may
contrast in the image, which can be useful bring up old imagery that isn’t
for distinguishing between spectrally similar representative of current ground conditions.
land cover types. Use the tool to return Therefore, keep project objectives and local
to the band combination page. land cover dynamics in mind when choosing
dates. Specific months within the date range
Downloading Imagery also can be selected using the Search
Months drop-down menu. This is useful
For some applications, informal visual when trying to target a specific part of the
interpretation of multispectral imagery using
vegetation growth cycle.
EO Browser or other online option may be
sufficient for natural resource–related Once you’ve chosen your dates and area of
planning or monitoring. For situations where interest, the next step is to choose the

5
image source you want to use. Click on the Now click on the Results button and scroll
Data Sets button. Click the + button next to through the images to decide which to
each data set name to open up further download. A coarse resolution preview
options. Check the box next to the image
image can be displayed by clicking the
source you want. Multiple Landsat options
button for each image. Click the same
are available for use.
button to remove the preview so that new
Level 1 data have basic preprocessing images can be viewed. You may notice that
steps completed but are not atmospherically some images have a significant amount of
corrected. This may or may not be important black space. This is a by-product of the
depending on the needs of the project. images being reprocessed into a different
Atmospheric correction reduces haze and tiling system. Look for images in which your
helps to standardize radiance values, which area of interest isn’t covered by clouds,
is important when using a single trained cloud shadows, or one of the no data black
classifier on multiple images. spaces.

Analysis Ready Data (ARD) are Once you’ve found an image that meets
atmospherically corrected and are usually
your needs, click on the button to view
the best option, but may not be available for
download options. For Sentinel 2 images
all dates. If ARD products are not available
you want to choose the L1C tile in
for a specific date, Level 1 products should
JPEG2000 option. For Landsat you need
be adequate. Sentinel 2 images are only
the Level 1 GeoTIFF data product or
available as L1C products from USGS. If
surface reflectance if using the ARD
L2A atmospherically corrected products are
dataset. Your image will download as a
needed, they can be downloaded from
compressed .gz, .tar, or .zip file (based on
https://scihub.copernicus.eu/dhus/#/home.
the data type) and may take a few minutes
See Song et al. (2001) for more information
depending on your internet connection.
on when atmospheric correction is
Depending on the size and location of your
necessary.
study area you may need to download
Once you’ve selected your data source, multiple images.
click on Additional Criteria. The most
Once the file is downloaded, move it out of
useful feature on this page is the Cloud
your Downloads folder and into a dedicated
Cover filter, which allows you to filter out
project folder of your choice under
images with a cloud cover greater than a
Documents (in your computer's File
specified level. Keep in mind that the cloud
Explorer).The .gz, .tar, or .zip file is a
cover percent is for the entire image (also
compressed file that will need to be
known as a scene), which is over 100
extracted using software such as 7-Zip
kilometers (km) wide for both Sentinel 2 and
(available at https://www.7-
Landsat 8. This means that if your area of
zip.org/download.html). Download the 7-Zip
interest is small, it may be possible to get a
.exe file that matches your system
cloud-free image of the site even if the
requirements (32 bit or 64 bit). Most modern
entire scene has high cloud cover. Don’t be
PCs are 64 bit, but you can check your
too aggressive with the cut-off filter you set
computer by opening the settings and
because you may exclude usable images
clicking on System then About. Whether
from your search results.
you have a 64 bit or 32 bit device will be
listed under System Type. Once you’ve
downloaded the correct .exe file, open it and

6
follow the prompts to install the 7-Zip files should now be fully extracted and
program. ready for processing. For Landsat Level 1
products, the .gz file will extract a .tar file,
Once 7-Zip is installed on your PC, navigate and the previous step will need to be
to the compressed file and right click on it.
repeated (two layers of compression are
Choose 7-Zip then Extract Here. For used due to the larger file size).
Sentinel 2 and Landsat ARD products, the

7
 Part 3. QGIS: Free Open-Source Mapping Software

QGIS is open source (free) mapping QGIS folder and double-click on QGIS
software that can be used to organize, Desktop [version number].
create, and display geospatial data. This
Once QGIS is open, click the Plugins tab
section covers some basic functions in
on the top banner and choose Manage and
QGIS, including adding, creating, and
Install Plugins. Use the search bar to find
modifying data layers. This will barely
and install the Semi-Automatic
scratch the surface of what QGIS is capable
Classification Plugin (SCP).
of. We suggest users consult the QGIS
documentation available at Adding Data
https://qgis.org/en/docs/index.html for
information on additional functions or Two main types of data can be loaded into
clarification on the topics covered in this QGIS for viewing, editing, and analyzing:
section. There also is a strong online user raster and vector. Satellite imagery is a type
community that can answer questions and of raster data. Raster files are grids of pixels
help you work through errors or other (cells) in which each pixel contains a single
problems. Find them at value. In the case of optical satellite imagery
https://gis.stackexchange.com/questions/tag (e.g., Sentinel 2 and Landsat), the pixel
ged/qgis. values represent the amount of solar
radiation within a spectral band being
Download and Setup reflected off the surface of the earth within
QGIS can be downloaded at the area sampled by each pixel (plus
potential atmospheric interference).
https://qgis.org/en/site/forusers/download.ht
Common raster data formats include .tif and
ml. Two options are available: the latest
.jp2.
release and the long-term release
repository. The latest release will have the Vector files are made up of points, lines, or
most up-to-date features while the long-term polygons. A vector layer may contain many
release will be more stable. In our features (i.e., separate points in a point-
experience either version should work fine. based vector layer). Information about each
feature in a vector layer is stored in an
Once you’ve decided on a version, choose
the QGIS Standalone Installer Version attribute table, which contains one row per
[version number] appropriate for your feature and as many columns of attribute
operating system. Click on the .exe file information as needed. Vector files are
when it is finished downloading and follow useful for delineating property boundaries,
roads, and infrastructure.
the prompts to complete the installation.
The instructions in this guide are based on
Vector layers typically are composed of
version 3.4, but the steps likely will be
three separate files with .shp, .shx, and .dbf
similar for future versions.
extensions, and they also may include .xml,
Once QGIS is installed, launch the software .prj, .sbn, and .sbx. All supporting files for a
by clicking on the Windows button and
searching for it in your app list. Open the

8
layer need to be in the same folder, but represents. You can add multiple bands at
when adding the layer to QGIS you only add once by holding the Ctrl button on your
the .shp file. Raster files often include an keyboard and clicking on each image file.
.xml supporting file, but again you only need Click Open then Add.
to add the actual image file (.tif, .jp2, etc.) to
QGIS. Sentinel 2 data are a little harder to find in
your file folder. Begin by opening the main
Data are added to QGIS using the folder containing the imagery and click on
button, which is located on the far left of the Granule. This should contain a single
third row of the top banner (figure 3). folder. Open it and click on IMG_DATA to
Clicking on the button opens a pop-up (data find the image files (stored as .jp2 files).
source manager) with a list of data types Part 4 of this guide will cover the image
along the left side. Choose Vector or bands that are most important for land cover
Raster depending on the file you plan on classification and demonstrate how to
combine these bands into a band set.
adding. Under Source click on the
button to the right of the page. This will Each of the images that have been added to
open up your file explorer. Navigate to your QGIS will appear under the Layers list on
file, click on it, and select Open. Click Add the left side of the interface. Click the box
at the bottom of your data source manager next to a layer to toggle it on or off. You can
pop-up. change the order of the layers in the list by
clicking and dragging them. The order of the
To add your downloaded Landsat imagery layers in this list determines the order they
to QGIS, open the Data Source Manager are rendered on the screen. Small layers
and click Raster, then open your file should be ordered higher than larger layers
explorer using the button. Navigate to (such as base maps) or they will be
the folder where the imagery is stored. Each obscured. Layers can be removed by right
image band (red, green, blue, near-infrared, clicking on the layer and choosing Remove
etc.) is stored as a separate image file. Layer, which will remove the layer from
QGIS but not delete it from your computer.
There are also additional image files with
information on cloud cover and other image-
quality information. The last two digits in the
file name identify what band the image

Figure 3. Top banner of QGIS interface, with important tabs/buttons identified. Note that the entire bottom row is devoted to the Semi-
Automatic Classification Plugin.

9
Creating and Editing file name, click on the button to the
Vector Layers right of the box to open up your file explorer.
Navigate to your project folder then create
A vector layer of a property boundary is
an appropriate file name in the file name
useful for satellite imagery analysis because
text box and click Save. Note that it is good
you can clip the imagery to your specific
practice to avoid spaces in file names;
area of interest to eliminate unnecessary
instead use ‘_’ or ‘-’ to separate words.
processing and allow for easy calculation of
Under Geometry Type, use the drop-down
the acreage of different land cover types.
menu to select Polygon. The next drop-
Property boundary layers for state and down menu contains an EPSG number,
federal public lands are typically available which likely will be 4326-WGS 84 by default.
either online or can be requested from the This is your coordinate reference system
relevant agency. Private property (see note below for more details). Click OK
boundaries may be more difficult to acquire. to finish creating the layer. The new layer
should now appear in your layers list, but
Sometimes the local tax assessor office has there will not be anything new on the map.
vector files that can be accessed online or
through a direct request, but this is not Next, you can add a base map that will
always the case. If you cannot find a pre- serve as a reference when you are tracing
made property boundary vector file, there the outline of the new vector layer by hand.
are a few options for creating your own. You could use Landsat or Sentinel 2
imagery as a base map, although they may
One option is to trace the boundary by hand not have high enough resolution to see
details that mark property boundaries. In the
in QGIS. Start by clicking on the button
(third row from the top) to pull up a dialogue United States, the National Agriculture
Imagery Program (NAIP) is a good option
box to create your new shape file. For your

Coordinate reference systems: Displaying the surface of the curved earth on a flat screen
or piece of paper is difficult, as there always will be trade-offs between minimizing distortions
in surface area, shapes, and/or angles. Even displaying the surface of the earth on a digital
sphere or ellipsoid is difficult because the earth isn’t perfectly round or smooth. A geographic
coordinate system is a digital model of the ellipsoidal surface of the earth, while a projected
coordinate system is a transformation of an ellipsoid onto a flat surface. Your choice of
coordinate reference system depends on your location (some are location-specific, others are
global) and will affect the physical appearance of your map as well as measures of area,
length, and angles. Using layers with different coordinate reference systems may lead to location
error in your maps, although the on-the-fly projection feature of QGIS typically limits such
problems. We recommend consulting the following web pages for more information on
choosing an appropriate coordinate reference system and managing your coordinate
reference system in QGIS:

https://docs.qgis.org/3.4/en/docs/gentle_gis_introduction/coordinate_reference_systems.html

https://docs.qgis.org/3.4/en/docs/user_manual/working_with_projections/working_with_projec
tions.html

10
for high resolution aerial imagery. It can be have a paper copy of a map with a property
downloaded from the USGS Earth Explorer border, you can digitize the map into a
website and added to QGIS using the same raster through a process called
steps detailed above for Landsat/Sentinel 2 georeferencing. You can then trace the
imagery. When searching for NAIP imagery boundary into a vector file using the steps
under the Data Sets tab in Earth Explorer, above. Step-by-step instructions for
choose Aerial Imagery and then NAIP. The georeferencing can be found at:
imagery will be downloaded as a .zip file https://docs.qgis.org/3.4/en/docs/training_m
that will need to be unzipped in order to add anual/forestry/map_georeferencing.html.
the TIF image file to QGIS.
Finally, you can create a property boundary
To draw the property boundary, select your layer by traveling the boundary in person
new layer in the layers list (it will now be with a GPS and uploading the file to QGIS.
The following link contains instructions for
underlined) and click the icon (third row
transferring files from a GPS to QGIS:
from the top) to toggle editing. Next, click
https://docs.qgis.org/3.4/en/docs/user_man
to the right of the toggle editing button ual/working_with_gps/index.html.
to start drawing your new feature. Your You can change the appearance of a vector
mouse cursor will change to crosshairs and layer by double clicking on its name in the
you can draw your property boundary (using layers list to open up the Symbology menu.
the base map as guidance) by left clicking At the top of the pop-up will be an option
to add new vertices. Right click to finish called Fill and below it Simple Fill. Click on
drawing and add a feature I.D. number in simple fill to open up more options for
the pop-up box. You can add more features changing the fill and outline of the layer.
(i.e., new points, polygons, or lines) to the Changing the appearance of a layer can
same layer by repeating this process, or you help to increase contrast among layers or
can edit previous features using the vertex create a more visually appealing map. You
editor . With the vertex editor you can can make a vector polygon transparent by
add, delete, or move the vertices (corners) choosing No Brush under the Fill Style drop-
down menu.
of the polygon. Click the button again to
save your edits and exit editing mode. If you

11
 Part 4. Supervised Land Cover Classification

At this point you should be familiar with the Step 1: Create Band Set
basics of accessing, visualizing, and
downloading satellite imagery as well as The first step in the supervised classification
some of the basic functions of QGIS. Now process is to combine the separate bands
we will describe how to create a land cover into a band set (also known as a band
map using the Semi-Automatic stack). Start by adding your imagery to
Classification Plugin (SCP) version 6.3.0. QGIS using the steps detailed in Part 3. For
Landsat 8 imagery, you need bands 2 to 7,
The instructions are based on a case study and for Sentinel 2 imagery you need bands
from Alabama that focused on detecting and
mapping Chinese privet, an invasive 2 to 8, 8A, 11, and 12. Next, click the
evergreen shrub, in a bottomland hardwood button (bottom row of the top banner) to
forest. The same steps will apply to other open the SCP (figure 4). Click on the Band
land cover mapping applications. For our Set option on the left side of the dialogue
example we will use Sentinel 2 imagery
from early March. Our research and other box. Click on the button to load your
studies have shown this is the most image bands. Highlight the bands by
effective time frame for detecting privet clicking on the first one, holding the shift key
because it is the time of year with the on your keyboard, then scrolling down and
greatest spectral difference between privet clicking on the last one.
and the deciduous overstory.

Figure 4. Band set tool in Semi-Automatic Classification Plugin

12
Once they are all highlighted in blue, click may take a few minutes for the process to
run. A new layer that ends in
the button. All the bands should now B0stack_raster.tif should now be present in
be listed under the Band Set Definition your QGIS layers list. This raster layer is a
section. multiband image stack that contains
At this point you could continue working with information from all of the image bands.
the full Landsat or Sentinel scene, but doing Go through the band set process one last
so is computationally expensive and will time to add your new multiband stack image
take too long to run on most computers. To to the SCP. Clear the previous band set
clip the image bands to your study site, click
on the Preprocessing tab on the left side of using the button. Click the button
the SCP menu. Scroll through the tabs at under Multiband Image List. Open the
the top until you find Clip Multiple Rasters. Multiband Image List drop-down menu
Check the box next to Use Vector for and click on the band stack image.
Clipping and choose your property
The individual bands should now be loaded
boundary vector file from the drop-down
under the Band Set Definition. Make sure
menu. The vector file needs to be loaded
the Quick Wavelength Settings are set on
into the QGIS project (listed in your layers
the correct satellite. Check to make sure the
list) in order for it to show up (see Part 3).
center wavelengths for each band are
You may need to click for it to appear. correct. If they are listed as 1.0, 2.0, 3.0,
Click run then select a folder to save the etc., you can temporarily choose a different
clipped images in. The clipping process may satellite under quick wavelength settings
take a few minutes. and then switch back to the correct satellite.
This should cause the center wavelengths
Now go back to the band set page. Click to change to the correct values. You do not
need to press run again. Simply adding the
to clear the previous (unclipped) band
new bands to the band set menu tells the
set and click to load the new clipped SCP that those are the bands you will be
images into the single band list. These working with.
images should contain the prefix clip. The band numbers used to set band
Highlight the clipped images and add them combinations in the SCP are based on the
to the band set using the button. Next, order of the bands in the band set definition
open the drop-down menu next to Quick menu, not the actual band numbers from
Wavelength Setting and choose the the satellite. Because we are using only a
satellite type you are using. This tells the subset of the satellite bands, the band
software which wavelength to assign to numbers in the SCP don’t line up with the
each of the bands. Make sure the bands are real satellite band numbers. Table 3 shows
listed under band set definition in the same the conversion between the SCP band
order as the bands in the quick wavelength numbers and the real Sentinel 2/Landsat 8
setting listing for your satellite. Next, check band numbers, assuming you are using the
the box next to Create Raster of Band Set correct band order for each satellite from
the quick wavelength settings.
(Stack Bands) and click . Your
file explorer will pop up. Choose your
project folder and click Select Folder. It

13
 Table 3. Band Number Conversions train the software to recognize the different
land cover types on the property through the
SCP Band Landsat 8 Sentinel 2 use of training sites (also called regions of
Number
interest or ROIs). You can either train the
1 2 2
software to recognize all the main land
2 3 3
cover types on the property (in which case
3 4 4
every pixel in the image will be sorted into a
4 5 5
5 6 6 land cover type during the classification
6 7 7 step) or you can choose only one or a few
7 N/A 8 land cover types to target (in which case
8 N/A 8A pixels that don’t match those land cover
9 N/A 11 types will be assigned a “no data” value
10 N/A 12 during the classification step). The latter
technique requires setting spectral
thresholds for determining when a pixel is
You can now close the SCP menu to view too dissimilar from any of the target land
your clipped band stack imagery. Find the cover types and is thus assigned a no data
value.
tool in the fourth row
of the QGIS banner at the top of your Setting an appropriate threshold can be a
screen. Make sure the dark circle next to trial and error process. For this tutorial we
RGB is clicked (it will have a smaller grey will map all primary land cover types on our
circle inside). Type in a band combination property—uninvaded deciduous hardwoods,
(see table 3) and press Enter on your privet-invaded deciduous hardwoods,
keyboard. It may take a few minutes for loblolly pine (Pinus taeda), swamp, open
your computer to render the result. Note water, and fields—to avoid that step.
that colors may not appear how you would
Find the SCP dock in the bottom left corner
expect for a natural color combination. You
of your screen (below the layers list). If it is
can use the and buttons to the
right of the RGB tool to increase color not there, click the button to open it.
contrast in the image. Choose the Training Input tab on the left

side of the dock (figure 5). Click the

Step 2: Set Training Sites*
button to create a new training file. Create
At this point you should have a clipped band at least three training sites per land cover
stacked image of your property visible on type to capture the natural spectral variation
the QGIS main screen. Your next step is to within each. Each land cover type is

*In some cases you may need to classify more than one image, such as if your study area spans
more than one Landsat or Sentinel scene or if you are working on a change-detection project. In
such cases it is generally best to use a separate set of training sites (and thus separate training
signatures) for each image due to variations in spectral properties across different acquisition
dates (affected by atmospheric and orbital conditions). You can run each image segment
through the classification process separately and combine the final maps for display or analysis.
However, if you must use the same training signatures for multiple images, you need to use
atmospherically corrected surface reflectance products.

14
considered a macroclass (MC), and each the MC ID by 1 and change the land cover
training site will have its own class (C). name for MC info. Repeat this process for
all your land cover types.
To create the first training signature, add the
appropriate land cover name in the MC info You can use a few methods to aid in
field of the SCP dock. You can add a class creating your training sites. The first and
name as well, but it isn’t necessary. Go to most obvious is on-the-ground knowledge.
the bottom row of the QGIS top banner and You can either rely on you prior experience/
familiarity with the property or conduct
click the button. Draw a polygon around informal ground surveys to identify suitable
a relatively small area on the image that is training sites for each land cover type.
representative of the land cover type you
entered in the MC Info field. Left click to Another option is to visually interpret the
add new vertices; right click to close the Landsat/Sentinel imagery to select training
polygon. Once you have finished drawing sites (see Part 2). If you can’t visit the
the polygon, click the save button in the property on foot and you have difficulty
interpreting the multispectral imagery, you
SCP dock ( ). can also try using a high-resolution base
To add another training site for the same map, such as Bing or Google imagery, for
land cover type, make sure that MC ID and reference.
MC info are the same as the last site but There is no set limit on how large or small
that the C ID is different (it should increase your training polygons should be. If the
by 1 automatically). The C info field can be polygon is too large you risk including pixels
changed or left the same. Repeat this in the polygon that belong to a different land
process until you have at least three training cover type. This can cause the spectral
sites per land cover type. To start training
sites for the next land cover type, increase

Figure 6. Training polygons for our example property are

shown here in yellow on top of a false color infrared Sentinel
Figure 5. SCP dock training input page image. When you create your own polygons they will appear
dark grey.

15
signatures of the two land cover types to be the training signatures set in the last step,
too similar to each other and increase the and sort the pixel into the land cover class
classification error rate. If the polygons are that is the best fit.
too small you risk not capturing the full
Click on the Classification tab on the left
variation in spectral signatures. Additionally,
you may get the following warning when side of the SCP dock (figure 7). At the top of
the menu, make sure that the box next to
using small polygons: “The following
MC ID is checked. Below that is the
signature will be excluded if using Maximum
Algorithm section. Use the drop-down
Likelihood.” This means that the training
menu to select between three different
polygon was too small for the software to
options: minimum distance, maximum
gather enough information on the variability
likelihood, and spectral angle mapping.
of the spectral signatures of the pixels within
These three options are essentially different
the polygon; thus, the Maximum Likelihood
ways of defining which land cover type is
classifier can’t use the information from that
the best fit for each pixel (i.e., which one
polygon. See figure 6 for a map of the
has the most similar spectral signature).
training polygons used in our case study.
Alternatively, a minimum threshold for
You also can set training signatures using spectral similarity can be used, although we
won’t cover that in this tutorial. With no
the button and a region-growing
threshold, every pixel in the image will get
algorithm, although that technique won’t be
classified into the most similar land cover
covered in this guide.
type, even if the best fit is still fairly different.
Step 3: Classify Land Cover Below the algorithm section is a section
In this step, the software will calculate the called Land Cover Signature
spectral signature of each pixel in the Classification. This is essentially another
image, compare that spectral signature to type of classification algorithm. This
algorithm sorts pixels into a land cover type
only if its signature falls completely within
the range of spectral signatures from the
training polygons of a single land cover
type. If it falls outside the spectral range of
one of the land cover types or if it falls within
the range of multiple types, the pixel is left
unclassified. You can click the box next to
LCS to activate this classification mode.
Checking the box below it (next to
Algorithm) tells the software to revert to
using whichever algorithm is selected in the
drop-down box for pixels that are
unclassified by the LCS algorithm. Selecting
the Only Overlap box tells the software to
only use the backup algorithm for pixels with
spectral signatures that overlap multiple
cover types. Pixels that don’t match any of
the land cover types are left unclassified.

Figure 7. SCP dock classification page

16
 There is no universal best algorithm. The and choose the one that best fits your
one that performs best in a given situation needs.
will depend on the spectral characteristics of
the land cover types being mapped (figure Once you are satisfied with the classification
8). For this reason it is typically best to
experiment with several algorithms to previews, you can click to
identify the best choice. The SCP makes produce a permanent file. If none of the
this easy via the preview function. algorithms produce useful results, you may
need to go back to step 2 and modify your
Once you’ve set up the classification spectral signature training polygons. It is
algorithm options, go to the bottom row of possible that you accidently mislabeled one
the QGIS top banner and click the of your training polygons, created a training
button and then click on your map. This will polygon that included multiple land cover
create a temporary file with a preview of the types, created a land cover type that was
land cover classification produced by the too broad (i.e., there was too much spectral
settings you’ve chosen. You can increase or variation), or missed an important land
decrease the size of the preview box using cover type altogether.
the arrows next to the ‘S’ field to the right of
Step 4: Complete an Accuracy
the preview button (the default is 200
pixels). You can change your classification Assessment
settings and then click to rerun the At this point you should have a classified
land map that appears fairly accurate based
preview. Click to remove the previews. on your informal visual assessment. The
You can use this technique to visually last step is to complete a formal accuracy
assess the accuracy of the land cover map assessment. How much time and resources
produced by each classification algorithm you spend on this step will depend on your
planned uses for the map. If you plan on

Figure 8. Land cover maps produced by three different classification algorithms applied to the same Sentinel 2 image. All three
maps highlight the same general area as invaded by privet (black); however, there is a notable difference on the east side of the
property. The circled area is a low-density pine stand with little privet cover, but the minimum distance and spectral angle
algorithms mislabeled significant portions as privet-invaded hardwood forest. The best algorithm for a different property and set of
land cover types will likely vary.

17
using acreage estimates and/or location If your reference data and map data don’t
information from the map as an integral part line up spatially, your accuracy statistics will
of planning or monitoring a management be unreliable. For this reason it may be
program, you need to put considerable better to assess accuracy using larger grids
effort into verifying the accuracy of your or polygons. There is no set rule for the
map. number of samples you should use,
although a minimum of 50 per land cover
There are many techniques available for
class has been suggested as a general
conducting accuracy assessments. Remote
guideline.
sensing professionals may disagree over
which technique is best suited for a given Once you’ve established your random
project. In this section we will provide an sample of points or polygons, there are
introduction to the basics of accuracy several ways to assess their true land
assessments along with links to several cover. One option is to use a GPS to
tools and associated tutorials that can be navigate to the samples on the ground and
used to assess accuracy in slightly different assess the land cover in person. This can
ways. be time-consuming but may be the most
reliable method, assuming you have access
Most accuracy assessments are conducted
to a GPS with adequate location accuracy.
through (1) randomly sampling a portion of
You could also assess your samples using
your study area, (2) assessing the true land
high-resolution imagery such as from the
cover within that random sample, and (3)
NAIP.
comparing the true land cover (also known
as the reference data) to the land cover When using moderate resolution imagery,
predicted by your classified land cover map. such as Landsat and Sentinel, there may be
The accuracy statistics derived by occasional mixed pixels in which two or
comparing the randomly sampled reference more land cover types are present within
data to the land cover map are assumed to the pixel. Whichever land cover makes up
be representative of the accuracy of the the majority (greater than 50 percent cover
entire map. threshold) of the pixel (or polygon if using
larger sampling units) is usually the best
The random samples of your study area
choice for labeling the sample. However, if
may be single points (representing single
you are specifically interested in detecting a
pixels) or a larger grid/polygon. Random
certain land cover type (for example, an
points may be easier to generate and
invasive plant species), you could try
sample in the field, but there is an important experimenting with several cover thresholds
limitation. All GPS devices have a certain to identify the smallest threshold at which
amount of location error, and in the land cover of interest can be reliably
recreational-grade GPS devices this error
identified. The cover threshold at which a
can be 10 or more meters. Additionally, the
specific land cover type can be reliably
satellite imagery itself will have some
identified will be affected by your training
amount of error in its location data.
areas, so if your goal, for example, is to detect
Combined, these two sources of location an invasive plant at low densities, you should
error can make it difficult to navigate to a have training areas that reflect that.
point on the ground that aligns with a
specific pixel. Once you’ve finished collecting your
reference data, you can compare it to the
land cover map to produce accuracy

18
measurements. Remote sensing specialists should appear at the bottom right of the
have developed many ways of defining QGIS interface. Click on this to open your
accuracy, many of which are based on an report (figure 9). The numbers listed under
error matrix (also known as a confusion Value correspond to your land cover types,
matrix). See Congalton 1991 for more although the value labeled as ‘0’ may be a
information on error matrices. no data value from the outline of your
property (likely shown in black on your
The SCP has a built-in accuracy
map). Pay attention to the area units, which
assessment function. A step-by-step tutorial
will be dependent on the projection of the
is available at
raster layer.
https://fromgistors.blogspot.com/2014/09/ac
curacy-assessment-using-random- Alternatively, the AcATaMa plugin provides
points.html?spref=yml. area estimates for land cover types with
confidence intervals (a measure of
There is also a stand-alone plugin for QGIS
uncertainty) based on your accuracy
called AcATaMa (Accuracy Assessment of assessment. This can be extremely useful
Thematic Maps) that is built to simplify when interpreting map outputs in the
accuracy assessments, including the context of management decisions.
generation of random points, reference data
collection, and analysis. Documentation and
a tutorial for this plugin can be found at
https://smbyc.github.io/AcATaMa/.

Step 5: Estimate Surface Area

Once you’ve developed a land cover map
with adequate accuracy for your project, the
last step is to estimate the surface area of
the land cover types that you mapped. Click
on the Processing tab at the top of the
QGIS interface and choose Toolbox. This
should open a list of options on the right
side of your screen. Choose Raster
Analysis Then Raster Layer Unique
Figure 9. Raster layer unique values report
Values Report.
Use the drop-down menu under Input
Layer to select your classified map. Click
Run at the bottom of the menu. A blue link

19
 Part 5. Additional Resources and References

This guide provided a very basic Landsat 8 User Guide:

introduction to satellite imagery analysis and https://www.usgs.gov/media/files/landsat-8-
related functions in QGIS. There are many data-users-handbook
additional tools available within the SCP Sentinel 2 User Guide:
and QGIS. The following web links and https://sentinel.esa.int/web/sentinel/user-
publication citations are good resources for guides/sentinel-2-msi
expanding your GIS and remote sensing–
related knowledge and skill sets. Some of Specific Topics
these resources have already been Creating and Exporting Professional Maps:
referenced in this guide, while others are https://docs.qgis.org/3.4/en/docs/training_m
new. anual/map_composer/index.html

Additional Resources Coordinate Reference Systems:

https://docs.qgis.org/3.4/en/docs/gentle_gis
Satellite Imagery Access _introduction/coordinate_reference_system
s.html
Sentinel Hub EO Browser:
https://apps.sentinel-hub.com/eo-browser/ Using Projections in QGIS:
https://docs.qgis.org/3.4/en/docs/user_man
USGS EarthExplorer: ual/working_with_projections/working_with_
https://earthexplorer.usgs.gov/ projections.html
ESA Copernicus Open Access Hub: QGIS and GPS Data:
https://scihub.copernicus.eu/dhus/#/home https://docs.qgis.org/3.4/en/docs/user_man
ual/working_with_gps/index.html
Software Downloads
Georeferencing:
QGIS:
https://docs.qgis.org/3.4/en/docs/training_m
https://qgis.org/en/site/forusers/download.ht
anual/forestry/map_georeferencing.html
ml
AcATaMa plugin documentation (accuracy
7-Zip: https://www.7-zip.org/download.html
assessments):
General Tutorials and User Guides https://smbyc.github.io/AcATaMa/

QGIS Documentation: SCP classification accuracy assessment

https://qgis.org/en/docs/index.html tutorial:
https://fromgistors.blogspot.com/2014/09/ac
QGIS User Forum: curacy-assessment-using-random-
https://gis.stackexchange.com/questions/tag points.html?spref=yml
ged/qgis
Semi-Automatic Classification Plugin
Tutorials:
https://fromgistors.blogspot.com/p/semi-
automatic-classification-plugin.html

20
References
Cash, J. S. 2019. “Evaluating remote Song, C., C. E. Woodstock, K. C. Seto, M.
sensing and prescribed fire methods to aid P. Lenney, and S. A. Macomber. 2001.
in the restoration of bottomland hardwood “Classification and change detection using
forests invaded by Chinese privet Landsat TM data: when and how to correct
(Ligustrum sinense).” Auburn University atmospheric effects.” Remote Sensing of
master’s thesis: 111. Environment 75: 230–244.

Congalton, R. G. 1991. “A review of [USGS] United States Geological Survey

assessing the accuracy of classifications of (n.d.) What are the best Landsat spectral
remotely sensed data.” Remote Sensing of bands for use in my research?. Available at
Environment 37: 35–46. https://www.usgs.gov/faqs/what-are-best-
landsat-spectral-bands-use-my-
Congedo L (2016). Semi-Automatic research?qt-news_science_products=0#qt-
Classification Plugin Documentation. DOI: news_science_products
http://dx.doi.org/10.13140/RG.2.2.29474.02
242/1
[ESA] European Space Agency (2015)
Sentinel-2 User Handbook. Available at
https://sentinel.esa.int/documents/247904/6
85211/Sentinel-2_User_Handbook

Knipling, E. B. 1970. “Physical and

physiological basis for the reflectance of
visible and near-infrared radiation from
vegetation.” Remote Sensing of
Environment 1: 155–159.
[NASA] National Aeronautics and Space
Administration (1999) Land cover
classification: how scientists differentiate
between land cover types.
https://earthobservatory.nasa.gov
/features/LandCover/land_cover_2.php
[SEOS] Science Education through Earth
Observation for High Schools (n.d.) 1.
Physical Basics. Available at
https://www.seos-
project.eu/remotesensing/remotesensing-
c01-p06.html

Siegmund A, Menz G (2005) Fernes nah

gebracht - Satelliten- und Luftbildeinsatz zur
Analyse von Umweltveränderungen im
Geographieunterricht. Geographie und
Schule 154: 7.

21