Chapter 1: Types of Imagery
Imagine the satellite sends different kinds of photos.
1. Panchromatic imagery
Black and white image
Very sharp details
Like an old newspaper photo
2. Multispectral imagery
Few color bands
Shows vegetation, water, soil separately
Like a normal color photo with extra powers
3. Hyperspectral imagery
Hundreds of colors
Can identify materials
Like seeing fingerprints of objects
4. Thermal imagery
Shows heat
Hot and cold areas
Like a temperature map
5. Radar imagery
Uses microwaves
Works day and night
Sees through clouds
Chapter 2: Image Rectification and Restoration
Your image arrives. It is crooked, dull, and noisy.
So first you FIX it.
A. Geometric Correction
Problem
The image is not in the right place. Roads do not match maps.
Why this happens
Earth is curved
Satellite is moving
Sensor angle distortion
Solution
Geometric correction means
👉 putting the image in the correct map position
Uses
Ground Control Points
Map projection
After this, the image fits perfectly on GIS layers.
B. Radiometric Correction
Problem
The image looks too bright or too dark.
Why
Sun angle
Atmospheric scattering
Sensor errors
Solution
Radiometric correction adjusts pixel values
👉 so brightness truly represents ground conditions
C. Noise Removal
Problem
Random dots appear in the image.
Cause
Sensor malfunction
Transmission errors
Solution
Noise removal smooths the image
👉 removes unwanted random values
Chapter 3: Image Enhancement
Now the image is correct, but still boring 😴
Enhancement makes it visually useful.
A. Contrast Manipulation
Story version
Your image is greyish. Everything looks same.
Contrast manipulation
Expands pixel value range
Makes dark darker
Bright brighter
Types
Linear stretching
Histogram equalization
👉 Features become visible.
B. Spatial Feature Manipulation
This is where shapes and patterns are enhanced.
1. Spatial Filtering
Think of a filter as a decision-maker.
Low-pass filter smooths image
High-pass filter sharpens edges
2. Convolution
Convolution is how filtering is done.
Story
A small matrix called kernel moves pixel by pixel
Calculates new values
Changes image appearance
Kernel + image = new image
3. Edge Enhancement
Purpose
Highlight boundaries
Roads, rivers, buildings become clear
Edges are areas where pixel values change suddenly.
4. Fourier Analysis
This sunds scary but listen.
Idea
Image = combination of patterns
Some patterns are smooth
Some are sharp
Fourier analysis separates image into
Low frequency components
High frequency components
Then you enhance or remove what you want.
Chapter 4: Multi-Image Manipulation
Now you combine multiple bands to extract information.
A. Spectral Ratioing
You divide one band by another.
Why
Removes shadow effects
Highlights specific features
Example
Vegetation index uses band ratios.
B. Principal Component Transformation
You have many bands saing similar things.
PCA
Combines correlated bands
Creates new components
First component has maximum information
Reduces data but keeps meaning.
C. Canonical Component Transformation
Used whe
Comparing two datasets
Or two time periods
Finds maximum correlation between them.
D. IHS Enhancement
IHS means
Intensity
Hue
Saturation
Story
Take color image
Separate brightness from color
Replace brightness with sharper image
Combine back
Used for pan-sharpening.
E. Decorrelation Stretching
Story
Bands are very similar
Colors look dull
Decorrelation removes similarity
Stretch colors separately
Result
Very colorful image
Small differences become visible
Used in geology and landform studies.
THE STORY OF IMAGE CLASSIFICATION 🌍🖍️
Imagine you receive a satellite image of Earth.
Every pixel is just a number.
Classification means teaching the computer what each pixel represents.
Forest. Water. Road. Building. Crop.
Chapter 1: Image Classification
Image classification is the process of
👉 assigning each pixel to a land cover category based on its spectral values.
Two main ways
Supervised classification
Unsupervised classification
Then comes hybrid methods and fixes.
CHAPTER 2: SUPERVISED CLASSIFICATION
You are the teacher 👩🏫
Computer is the student 💻
You already know what forest, water, and buildings look like.
Assembling the training data
Story
You point at the image and say
“This is forest”
“This is water”
“This is road”
These selected pixels are called
👉 training samples or ROIs
They represent real ground conditions.
Graphical representation of spectral response patterns
Each land cover has a spectral signature.
When you draw graphs
X axis = wavelength
Y axis = reflectance
Forest, water, soil form different curves.
This helps check whether classes are separable.
Quantitative expressions of category separation
Now numbers talk.
Computer calculates
Distance between class means
Variance within classes
If classes overlap too much
👉 classification will be poor.
Good separation means good accuracy.
Self-classification of training data set
Before classifying whole image
the computer tests itself.
It classifies the training pixels again.
If forest pixels become water
your training data is wrong.
This step checks reliability.
Interactive preliminary classification
You run a test classification.
You look at the result
Too much confusion
Wrong pixels
You go back
Modify training samples
Add better ROIs
This is trial and correction.
Representative sub-scene classification
Instead of classifying whole image
you classify a small representative area first.
If result is good
you apply same rules to entire image.
SUPERVISED CLASSIFICATION ALGORITHMS
Now comes the math part. But story first.
1. Minimum Distance to Means
Story
Each class has a centre point called mean.
Computer checks
Which class mean is closest to the pixel.
Closest mean wins.
Features
Simple
Fast
Ignores class spread
2. Parallelepiped Classification
Story
Each class is given a box in spectral space.
If pixel falls inside box
it belongs to that class.
Problems
Overlaps cause confusion
Some pixels remain unclassified
3. Gaussian Maximum
Likelihood Classification
Story
This method is very intelligent.
It checks
Distance from mean
Variance of class
Probability of belonging
Pixel is assigned to class with highest probability.
Most accurate supervised method
but computationally heavy.
CHAPTER 3: UNSUPERVISED CLASSIFICATION
Story
You tell computer
“Group similar pixels”
Computer forms clusters
based on spectral similarity.
Later you label clusters
forest, water, urban.
Common algorithms
K-means
ISODATA
CHAPTER 4: HYBRID CLASSIFICATION
Best of both worlds.
Story
Start with unsupervised classification
Identify clusters
Use them as training data
Apply supervised classification
Improves accuracy.
CHAPTER 5: CLASSIFICATION OF MIXED PIXELS
Real world pixels are messy.
One pixel may contain
tree + soil + road.
Spectral Mixture Analysis
Story
Pixel is treated as a mixture.
Computer estimates
percentage of each material.
Uses pure pixels called endmembers.
Fuzzy Classification
Story
Pixel is not forced into one class.
It can be
60% forest
40% agriculture
This matches real landscapes.
CHAPTER 6: POST-CLASSIFICATION SMOOTHING
After classification
image looks noisy.
Tiny isolated pixels everywhere.
Post-classification smoothing
Applies majority filter
Removes salt and pepper noise
Produces cleaner map
