Gualvez, Aprilla M.

(2020-8993-10687)
Laboratory Exercise No. 6: Image Interpretation and Classification

The creation of maps of Land Use and Land Cover is one of the most significant uses of
remote sensing data, and this function can be managed via a procedure known as image
classification. The process of categorizing every pixel in an image or the raw data from remotely
sensing satellites to produce a specified set of labels or land cover themes is known as image
classification (Lillesand, Keifer 1994). This procedure, which attempts to group every pixel in
the image into unique groups or themes, may be the most significant phase in digital image
analysis because it is essential for label production and target location monitoring. Hence, the
objective of this activity was to learn about and experiment with some of the most popular
classification approaches and to hone image interpretation skills.

I. Delineating Regions of Interest (ROIs)

To start the process, proceed by creating a new project in QGIS then continue by adding
layer. Load your reflectance bands created in Laboratory 4 in the Layers Panel, have it enhanced
through applying enhancement modifications that would improve the display and made land
cover types more distinguishable. In this instance, I applied filters and custom stretched Band 1-3
to my specifications. For the morphological filter, I used the Opening method, and the simple
filter with the sharpen approach.

Fig. 1 Loaded Reflectance Bands

Fig 2. Custom Stretched Fig 3 Applied Filters

 Question 1. What different land cover types you see/think is present in your image?

Vegetation (Grass/Crops) Built-Ups Inland Water (River/Lake)

Forest Bare soil Bay
Harbor Airport Port

Fig 5. Land Cover Types Table

In SCP plugin window, refresh the band list displayed in QGIS, this time, the reflectance
image bands will automatically get listed in the Band Set 1 list. In the Quick wavelength settings,
choose Landsat 8 OLI to automatically establish the Center wavelength of each band and the
Wavelength unit. In order to display the color composite, a Virtual Band Set 1 will pop up on
your Layers Panel when you define the band set. We will use the RGB False-Infrared
combination for this laboratory exercise.

Fig 6 False-Infrared RGB Combination

Continue the process by creating Training input in order to collect Training Areas (ROIs)
and calculate the Spectral Signature thereof which will be useful in image classification. In
the SCP dock click the button create a new training input and define a folder name. The path of
the file is displayed in Training input.

II. Creating ROIs

In creating ROIs, begin by defining the Classes and Macro classes, which are the
identification codes of land cover classes. When attempting to categorize materials with various
spectral signatures inside the same land cover class, it is helpful to employ macro classes, which
are collections of ROIs with distinct Class IDs.
 ROIs are created by clicking on any pixel of the image. Click “Create a ROI Polygon” in
the working toolbar, proceed by clicking on the map to define the vertices of you ROI then close
the polygon vertices by right-clicking. Over the image, a temporary polygon in the color orange
is visible and is semi-transparent. If you are satisfied with the shape of the polygon for your
chosen ROI, configure the Macro class ID, Marco class info, for example "water," then the
Classes ID and Classes info, for instance "Lake," in the ROI creation list. As a result, the spectral
signature will be calculated, and the ROI will be added to the list of ROI Signatures.

Fig 8 ROI Polygon

Fig 7 ROIs

In creating ROI for Built up class, zoom in the map over the brownish-gray region of the
image, these comprises residentials, commercial and industrial subclasses. Click the “Activate
ROI pointer” in the working toolbar. A temporary ROI created with the automatic region
growing algorithm will be generated. It is an image segmentation process which automatically
select clusters for images and characterizes the texture and color present in the image for the
final segmentation.

Fig 8 Built Up Class

 ROI INTERPRETATION KEY
Descriptive Thumbnail
Land Cover Land Cover Tone/Color Appearance on Screenshot of
Subclasses Imagery Samples

Bay 1 Blackish Color in Smooth Texture

RGB 432

Greenish-Blue
Bay 2 Color in RGB 432 Smooth Texture

Smooth Texture,
Water Watershed Dark Blue Color resembles a tree
in RGB 432 branch

Smooth Texture
River Greenish-Blue surrounded by
Color in RGB 432 coarse texture of
vegetation

Blackish Color
Harbor with low Smooth Texture
transparency
White-like smoke

Construction Yellowish-Brown Coarse Texture

Site Color in RGB 432

Bare Soil White Beige in Course Texture,

Mining Site Color in RGB 432 Seed-like shape

Landfill Grayish Color in Irregular surface,

RGB 432 unsmooth texture
 Grayish-Brown Smooth but Edgy
Residences with spots of Red texture

Built Up

Airport Grayish Color Coarse Texture

Saturated Red
High (Red to Dark Coarse Texture
Vegetation Red)
Color in RGB 432

Low Unsaturated Red Unrefined texture,

Vegetation Color irregular surfaces
Vegetation

Saturated Red, Coarse Texture.

Bay some spot of dark Irregular shape
Vegetation red pixels

Mangrove Unsaturated Red Coarse Texture

Rice field Unsaturated Red Coarse and Edgy

to Dark Red Texture
 Cloud ---- Bright White in Smooth Texture,
Color irregular shape

----

Shadow Blackish Gray in Edgy Texture

Color

The following ROIs were attained by using Google Maps. Since the image is
quite fragmented and impossible to detect a specific subclass without zooming, it is challenging
to identify the land cover subclasses. I look at the image's Land Cover elements using Google
Maps, compare them, and look up where they are in relation to one another.
III. Assess the Spectral Signatures

To assess the Image Spectral Signatures, show your selected ROI signatures in
the Spectral Signature Plot, this will be accessed by highlighting your spectral signature list then
clicking the tool “Add the highlighted signatures to spectral signature plot”. After the Spectral
Signature Plot window got displayed, proceed by clicking “Calculate Spectral distances” tool,
by doing this, you will be able to calculate the spectral distances of the selected signatures, which
will be very helpful in assessing ROI separability. From the table of values, you could determine
whether the variation of classes could have similarities that would cause classification errors.

Fig 9 Spectral
Signature Plot
 Fig 10
Spectral Distances

Question 2. Inspect the spectral signature plot of the different land cover classes. Is the plot
consistent with the expected land cover properties?

The SCP implements the following algorithms for assessing similarity of spectral

signatures: Jeffries-Matusita Distance, Spectral Angle, Euclidian Distance and Bray-Curtis
Similarity. However, for this Exercise we will just check the values for Jeffries Matusita
separability measure. Values exceeding 1.9 indicate that the ROI pairs are easily distinguishable.
If possible, try to make ROI pairs with lower separability values more separable when selecting
new ROIs or making changes to existing ROIs.

Question 3. Among your identified classes, which pair is most separable? Which is least?
Verify if these most and least statistically separable pairs are also the most and least visually
distinguishable classes. What land cover classes seem to be almost the same? If combining
certain training sites will create better separability between classes, which classes would you
merge and why?
From what I’ve encountered, the most separable pair are the water subclasses and
vegetation subclasses, I guess it’s because of their distinct color and texture. On the other hand,
I consider Built Up subclasses, such as the Residence and the Airport as the least separable pair
and also some of the subclasses in Bare Soil class. From this, it happened that water and
vegetation subclasses appeared to be the most distinguishable classes while Bare Soil and
specifically the Airport in Built Up class, appeared to be the least distinguishable.
If we will look on the specific, my built-up Class failed to create separability between its
subclasses. Since Airport covers a small area in the entire image, its spectral signature is
indistinct with the Residences. Therefore, it somehow collided and turned out to be inseparable
in the classified image. For this reason, I might merge Built Up subclasses.
However, if I will merge between macro classes, I’ll consider merging the Built-Up Class
and the Bare Soil Class since their textural features are almost the same and they cover
significant spatial features in the entire image.

IV. Creating Classification Preview

In order to assess the selected ROIs, we need to perform classification preview. This
allows us to test the algorithm of a small selected area before classifying the whole image or the
classification output itself. To do this, click the “Active classification preview pointer” then click
on the map to start the classification process and from it, a classification preview of your selected
area on the map will be displayed.

Fig 11 Classification Preview

V. Creating the Classification Output

 If the classification preview results were good, based on pixels that were allocated to the
right class identified in the ROI Signature list, we may now proceed to the classification of the
entire image. In this case, I clicked on the band processing tool directing to the classification
button then selecting MC ID, and Maximum Likelihood as its algorithm. Run the process and
define the path of the output and have it saved as a tif file.  You will hear a sound when the
process is finished.

Fig 12 Classification Output (MC ID) Fig 13 Classification Output (C ID)

 Fig 14 Original Image

Satellite images are obtained with the aid of remote sensing. These photos, however, need
to be processed before we can examine them. So, image classification is required in order to use
these images for analysis. Therefore, this makes image classification as one of most significant
stages in image processing and analysis. Earth's surface types correspond to varied spectral
reflectance features; by identifying these distinctive qualities, classification can be done.
Pre-processing is a step in the classification process that involves calibrating, enhancing,
and applying various corrections and enhancements. Its primary goal is to ensure that your data
is accurate and reliable for the following procedures. The procedure continues from there by
choosing representative parts of the image and creating training signatures. After executing
classification algorithms, you can now post-process the classified image once again to examine
and evaluate the precision of your results.
In classifying images, accuracy assessment plays a key role. The accessibility of high-
quality remotely sensed images, the design of an appropriate classification procedure, and the
expertise and experience of the analyst are only a few of the variables that affect how well an
image classification in remote sensing turns out. Thus, this Laboratory exercise made me realize
that making thematic maps was never easy, there are many factors to take into account,
particularly in terms of accuracy.

References:
Congedo Luca (2016). Semi-Automatic Classification Plugin Documentation. DOI:
http://dx.doi.org/10.13140/RG.2.2.29474.02242/1

Thematic Map. Statistics Canada: Canada's national statistical agency / Statistique Canada:
Organisme statistique national du Canada. (2018, September 17). Retrieved November 30,
2022, from https://www150.statcan.gc.ca/n1/pub/92-195-x/2011001/other-autre/theme/def-
eng.htm

Al-doski, J., Mansorl, S. B., & Shafri, H. Z. M. (2013). Image Classification in Remote Sensing,
3(10). https://doi.org/https://core.ac.uk/download/pdf/234663192.pdf
Image Classification. (n.d.)
http://www.sc.chula.ac.th/courseware/2309507/Lecture/remote18.html