NDVI – Normalized Difference Vegetation Index

What is NDVI?

 NDVI stands for Normalized Difference Vegetation Index.

 It is a widely used vegetation index in remote sensing to measure the health, density,
and presence of green vegetation.
 NDVI is calculated using the red (R) and near-infrared (NIR) bands of satellite imagery.

NDVI Formula:

NDVI = NIR – R/ NIR +R

 NIR (Near Infrared): Strongly reflected by healthy vegetation.

 Red: Absorbed by chlorophyll during photosynthesis.

NDVI Value Range:

NDVI Value Meaning

+0.6 to +1.0 Dense, healthy vegetation
+0.2 to +0.5 Sparse vegetation, grasslands
0 to +0.2 Bare soil, very little vegetation
<0 Water bodies, clouds, snow, etc.

Importance of NDVI

NDVI is a simple yet powerful tool that uses satellite data to measure vegetation health and
distribution. It plays a vital role in agriculture, forestry, ecology, and environmental
management by providing timely and accurate information about plant growth and ecosystem
health.

1. Detects Plant Health

 Healthy plants reflect more NIR and absorb more red light.
 NDVI helps detect crop stress, disease, or drought effects.

2. Tracks Vegetation Growth over Time

 NDVI time series allows monitoring of seasonal changes in vegetation.

 Useful in studying crop cycles, forest growth, and climate impacts.
Page 1 of 13
3. Large-Area Monitoring

 NDVI from satellite images (e.g., MODIS, Landsat, Sentinel) helps monitor vegetation
over large regions quickly and cost-effectively.

4. Agricultural Applications

 Farmers and researchers use NDVI to assess crop yield, field variability, and precision
farming decisions.

5. Environmental and Climate Studies

 NDVI helps track deforestation, desertification, and the effects of climate change on
vegetation cover.

6. Disaster Impact Assessment

 After events like wildfires, floods, or droughts, NDVI can assess the damage to
vegetation.

True color/ Natural Color and False Color Composites

1. What is a Color Composite in Remote Sensing?

 A color composite is an image created by combining data from different spectral bands
(usually Red, Green, Blue) to visualize satellite imagery.
 These bands can be from the visible spectrum (what our eyes can see) or from non-
visible parts like near-infrared (NIR).
 There are two main types:
o Natural Color Composite
o False Color Composite

Natural Color Composite

Definition:

A natural color composite uses the red, green, and blue bands of a satellite image in their true
color positions (R → R, G → G, B → B), so the image looks like what the human eye would see.

Band Assignment Example:

 Red band → Red color

Page 2 of 13
  Green band → Green color
 Blue band → Blue color

Appearance:

 Vegetation appears green

 Water appears blue or black
 Urban areas appear gray
 Soil/bare land appears brownish

Example:

 A natural color image from Landsat or Sentinel-2 can look like a photograph.
 Used in urban mapping, general land cover visualization, and visual interpretation.

False Color Composite (FCC)

Definition:

A false color composite uses non-visible spectral bands, like near-infrared (NIR), and assigns
them to visible colors to highlight specific features not easily seen in natural color.

Common Band Assignment Example (Standard FCC):

 NIR band → Red

 Red band → Green
 Green band → Blue

Appearance:

 Healthy vegetation appears bright red (due to high NIR reflectance)

 Water appears dark blue or black
 Urban areas appear cyan or gray
 Bare soil appears light brown or tan

Example:

 A false color composite from Landsat Band 5 (NIR), Band 4 (Red), and Band 3 (Green) is
often used in agriculture, forest monitoring, and vegetation health studies.

Page 3 of 13
Comparison Table:

Feature Natural Color Composite False Color Composite

Bands Used Red, Green, Blue (Visible) NIR + Visible (e.g., NIR, Red, Green)
Looks Like Photo? Yes No
Vegetation
Green Red or magenta
Appearance
Water Appearance Blue or black Dark blue or black
Visual interpretation, urban Vegetation health, agriculture,
Main Use
areas environmental
Detects Invisible
No Yes (e.g., moisture, plant health)
Info?

Conclusion:

 Natural color composites show the Earth's surface as we see it with our eyes, useful for
basic interpretation.
 False color composites enhance hidden features (like vegetation health or moisture
levels) by using non-visible spectral bands.
 Both types are essential in remote sensing for different applications.

Role of Training Sites in Supervised Classification

What is Supervised Classification?

 Supervised classification is a method in remote sensing where the computer assigns

land cover classes (e.g., water, forest, urban) to pixels in an image based on known
examples.
 These known examples are called training sites or training samples.

What Are Training Sites?

 Training sites are specific areas in the satellite image where the land cover type is
already known (from field surveys, maps, or expert knowledge).
 These areas are selected by the user to "teach" the computer what each land cover class
looks like spectrally (based on reflectance values in different bands).

Role of Training Sites in Supervised Classification:

Page 4 of 13
 1. Provide Reference Data:
o Training sites act as the ground truth to help the software learn the spectral
patterns of different classes.
2. Build Classification Model:
o The algorithm (like Maximum Likelihood, SVM, or Decision Tree) uses the data
from training sites to create statistical or machine learning models that can
classify the rest of the image.
3. Improve Accuracy:
o Well-selected training sites improve the accuracy and reliability of the
classification results.
4. Guide the Class Labels:
o Training sites help define what class each pixel in the image should be assigned
to, based on similarity to known samples.
5. Validation and Testing:
o Some training sites may be used for validation, checking how well the
classification performs compared to actual ground conditions.

Key Characteristics of Good Training Sites:

 Should be representative of the entire class (e.g., don’t choose just one type of forest if
many types exist).
 Should be uniform in land cover.
 Should cover enough pixels to allow the algorithm to learn.
 Should be well-distributed across the image.
 Example:

In a land use map:

 Select areas of pure water bodies as one training site.

 Choose known urban areas, cropland, and forest as others.
 The classifier learns from these and applies the learned patterns to the rest of the image.

Conclusion:

Training sites are the foundation of supervised classification. They provide the essential data
that guides the algorithm in identifying and labeling different land cover types across the
image. Without good training sites, the classification will likely be inaccurate or misleading.

Page 5 of 13
 Limitations in Performing Change Detection Using Classified Images from Different Time
Periods

Change Detection definition

Change detection in remote sensing refers to the process of identifying differences in land
cover or land use over time by analyzing satellite images from two or more time periods.

When using classified images (i.e., images already processed into land cover classes), change
detection compares the class labels between the two time periods to find areas of change.

Limitations in Change Detection Using Classified Images:

1. Geometric Misalignment (Registration Errors)

 Even a slight shift between the two images can result in false changes.
 Misaligned pixels may appear to have changed class when they actually haven't.

2. Differences in Classification Accuracy

 If one image is classified more accurately than the other, differences in quality can lead
to incorrect detection of change.
 Example: A misclassified forest in one year may appear as deforestation in the analysis.

3. Different Classification Schemes

 Images may have been classified using different land cover classes or naming systems.
 Incompatible class labels can make it hard to compare results.
 Example: One image uses "cropland" and "grassland", the other just "agriculture".

4. Atmospheric and Seasonal Differences

 Changes in season, sun angle, or atmospheric conditions (clouds, haze) affect how land
cover appears in images.
 This can lead to misclassification, which affects change detection.

5. Classification Method Variations

 Using different classifiers (e.g., SVM in one year and Maximum Likelihood in another)
can produce inconsistent results.
 Differences in training data also impact classification outcomes.

Page 6 of 13
6. Mixed Pixels and Boundary Issues

 Pixels near land cover boundaries may change classes due to small variations, even if
there is no actual land cover change.
 This leads to edge effects in change maps.

7. Human Error in Training Site Selection

 If training sites for classification were selected inconsistently across time periods, the
resulting land cover maps may be biased, leading to incorrect change detection.

8. Post-Classification Errors

 Post-classification comparison methods depend entirely on the accuracy of the

classified images.
 Errors are compounded — if each classification has 85% accuracy, the combined
accuracy could be much lower.

9. Difficulty in Interpreting “Noisy” Changes

 Some detected changes may not reflect actual land use change (e.g., temporary water,
seasonal crops).
 This can make results hard to interpret without additional information.

Why Hyperspectral Images Have High Dimensionality

Definition of Hyperspectral Images:

1. Hyperspectral images capture data in many narrow, contiguous spectral bands across a
wide range of wavelengths (often hundreds of bands).
2. Unlike multispectral images, which have a few broad bands (e.g., 3–10 bands),
hyperspectral sensors record detailed spectral information.

Reasons for High Dimensionality:

1. Large Number of Spectral Bands:

1. Hyperspectral sensors record reflectance or radiance in hundreds of narrow

bands, typically ranging from visible to infrared wavelengths.
2. This means each pixel contains a high-dimensional spectral vector instead of
just a few values.

Page 7 of 13
 2. Fine Spectral Resolution:

1. The bands are narrow and closely spaced (e.g., 5–10 nm bandwidth), capturing
subtle differences in material properties.
2. This increases the number of variables or features per pixel.

3. Spectral Richness:

1. The detailed spectral information allows identification of chemical and physical

properties of materials.
2. This leads to a large feature space representing many spectral signatures.

4. Spatial Coverage with Rich Spectral Data:

1. Each pixel is described by a high-dimensional vector of spectral values,

combining spatial and spectral information.

Implications:

1. High dimensionality leads to challenges such as the “curse of dimensionality” — where

data processing and analysis become computationally intensive.
2. Dimensionality reduction techniques (e.g., PCA, MNF) are often required.

Hyperspectral images have high dimensionality because they capture detailed spectral
information in hundreds of narrow and contiguous bands, providing rich data about surface
materials but also increasing data complexity.

Role of Microwave Remote Sensing

What is Microwave Remote Sensing?

 Microwave remote sensing uses microwave wavelengths (typically 1 mm to 1 meter) to

observe Earth’s surface.
 It includes active sensors like Synthetic Aperture Radar (SAR) and passive sensors like
microwave radiometers.
 Unlike visible and infrared sensors, microwaves can penetrate clouds, rain, and
vegetation and work day or night.

Key Roles and Applications:

1. All-Weather, Day-Night Imaging

Page 8 of 13
  Microwaves penetrate clouds, fog, and rain, making it possible to monitor Earth’s
surface under all weather conditions.
 Can operate day and night because it does not rely on sunlight.
 Vital for continuous monitoring of areas with frequent cloud cover (e.g., tropical
regions).

2. Ocean and Sea Surface Monitoring

 Measures sea surface roughness, helping monitor waves, currents, and wind speed
over oceans.
 Used in sea ice detection and tracking oil spills.
 Important for weather forecasting and marine navigation.

3. Vegetation and Soil Moisture Monitoring

 Microwave signals interact with vegetation and soil differently than optical signals.
 Enables estimation of soil moisture content, which is crucial for agriculture and drought
monitoring.
 Helps in assessing biomass, forest structure, and crop conditions.

4. Surface Topography and Land Use Mapping

 Synthetic Aperture Radar (SAR) can generate high-resolution digital elevation models
(DEMs).
 Useful for mapping terrain, detecting landslides, and monitoring urban growth.

5. Snow and Ice Studies

 Microwave sensors can penetrate snow cover and help measure snow water equivalent.
 Monitor glaciers, polar ice sheets, and changes in snow cover, important for climate
studies.

6. Disaster Management

 Provides critical data during floods, hurricanes, earthquakes, and volcanic eruptions.
 Penetrates clouds and smoke to help assess damage and support emergency response.

Advantages of Microwave Remote Sensing:

 All-weather capability: Can see through clouds and rain.

 Day and night operation: Independent of sunlight.

Page 9 of 13
  Penetration ability: Can probe beneath vegetation canopy and dry soils.
 High spatial resolution: Especially with SAR technology.

Microwave remote sensing plays a vital role in Earth observation by enabling continuous,
reliable monitoring of the planet’s surface regardless of weather or lighting conditions. Its
unique ability to penetrate clouds and provide information on soil moisture, vegetation, ocean
surfaces, and topography makes it indispensable for environmental monitoring, agriculture,
disaster management, and climate studies.

Supervised vs. Unsupervised Classification: Advantages and Disadvantages

1. Supervised Classification

Advantages

1. Uses prior knowledge: Relies on labeled training data (training sites), making it more
accurate when good training samples are available.
2. Specific classes: Allows classification into predefined land cover classes, matching user
requirements.
3. Better accuracy: Usually provides higher classification accuracy if quality training data
are available.
4. Flexible: Can use various algorithms like Maximum Likelihood, SVM, Random Forest,
etc.
5. User control: Users can guide the classification process with expert knowledge.

Disadvantages

1. Requires training data: Needs sufficient and representative training samples, which
may be time-consuming or expensive to collect.
2. Bias risk: Poor or unrepresentative training data can lead to biased or inaccurate results.
3. Not good for unknown classes: Cannot identify new or unexpected land cover types
outside training data.
4. Labor-intensive: Requires expertise to select training sites and tune algorithms.
5. Sensitive to noise: If training data contain errors, classification accuracy suffers.

2. Unsupervised Classification

Advantages

 No prior knowledge needed: Automatically groups pixels based on spectral similarity

without requiring training data.

Page 10 of 13
  Exploratory: Useful for discovering unknown or unexpected classes.
 Quick and easy: Less user input is required, making it faster for initial analysis.
 Good for data exploration: Helps in understanding data structure before detailed
classification.

Disadvantages

 Class labeling needed: After clustering, classes need to be interpreted and labeled by
the user, which can be subjective.
 Lower accuracy: Generally less accurate than supervised methods because it doesn’t use
ground truth information.
 May generate meaningless clusters: Some clusters may not correspond to real land
cover types.
 User interpretation required: The user must decide what each cluster represents.
 Difficulty with mixed pixels: May not separate spectrally similar classes well.

Feature Supervised Classification Unsupervised Classification

Training Data Required (labeled samples) Not required
High (training sites, class
User Input Low (initial clustering only)
names)
Generally higher if training is
Accuracy Generally lower
good
Class Definition Predefined by user Determined by algorithm (clusters)
Discovery of
Limited to known classes Can discover unknown classes
Unknowns
Interpretation Clear, based on training Subjective, requires labeling
When good ground truth When ground truth is limited or
Application
available unknown

 Supervised classification is best when you have reliable training data and want specific
land cover classes.
 Unsupervised classification is useful for preliminary analysis and exploring unknown
data patterns but usually needs further interpretation and validation.

Page 11 of 13
 Evaluation of Hybrid Classification Methods

What is Hybrid Classification?

 Hybrid classification combines the strengths of both unsupervised and supervised

classification methods.
 Typically, unsupervised classification is performed first to generate clusters, followed by
supervised classification where the user labels clusters and refines the classification.

Benefits of Hybrid Classification:

1. Improved Accuracy:
o By combining both methods, hybrid classification can achieve better accuracy
than either method alone.
o Unsupervised clustering helps identify natural spectral groups, reducing bias
from poor training data.
2. Efficient Use of Training Data:
o Requires fewer and more focused training samples compared to purely
supervised classification, since clusters represent spectral groups.
3. Reduces User Bias:
o Initial unsupervised clustering reduces subjectivity in selecting training sites, as
clusters are data-driven.
4. Handles Complex Data:
o Works well with heterogeneous landscapes by grouping similar pixels first and
then labeling them, improving classification of mixed or transitional areas.
5. Flexibility:
o Allows iterative refinement: user can adjust class labels after unsupervised
clustering, improving class definitions.
6. Better at Discovering Unknown Classes:
o Unsupervised step can reveal classes or land cover types that were not
predefined, which can then be included in supervised labeling.

Drawbacks of Hybrid Classification:

1. More Complex and Time-Consuming:

o The two-step process requires additional time and expertise compared to pure
methods.
2. Cluster Interpretation Challenges:
o Assigning meaningful class labels to clusters can be subjective and requires
expert knowledge.
3. Error Propagation:
Page 12 of 13
 Errors in the unsupervised clustering step may affect the supervised
o
classification, impacting final accuracy.
4. Computationally Intensive:
o Performing both unsupervised and supervised steps may require more
processing power, especially for large datasets.
5. Requires Expert Judgment:
o Successful hybrid classification depends on the user’s ability to interpret clusters
correctly and select appropriate training data.

Aspect Benefits Drawbacks

Generally higher than pure Dependent on cluster labeling
Accuracy
methods accuracy
Training Data Less training data needed Requires expert interpretation
Process Complexity Flexible and iterative More complex and time-consuming
Bias Reduces initial bias Subject to bias during cluster labeling
Computational More computational resources
Can handle complex data
Load needed
Discovery of
Can reveal unknown classes Cluster interpretation challenges
Classes

Hybrid classification methods provide a balanced approach that leverages the advantages of
both supervised and unsupervised techniques, often resulting in improved accuracy and
efficiency. However, they require expert judgment, more processing time, and careful
interpretation of clusters, which can be challenging.

Page 13 of 13