Module 7-8: Image Interpretation, Land Cover Classification, and Accuracy Assessment

Lecture 09-11

Image Interpretation

This is the process of identifying objects or conditions in images and determining their meaning/significance

• Subject

• Geographic Region

• Remote Sensing System

Image interpretation is based on three separate kinds of knowledge:

Classification, Enumeration, Measurement, and Delineation

Tasks of Image Interpretation

Classification

It is the assignment of objects, features, or areas to classes based on their appearance on the imagery.

Detection, recognition, and identification

Classification follows the process of:

Detection

s the assignment of objects, features, or areas to classes based on their appearance on the imagery.

Recognition

s the assignment of objects, features, or areas to classes based on their appearance on the imagery.

Identification

means that the identity of an object or feature can be specified with enough confidence and detail to place it in a specific class.

Enumeration

This task means that the identity of an object or feature can be specified with enough confidence and detail to place it in a specific class

Photogrammetry and photometry

Measurment Can be done either:

Photogrammetry

measurement of distance and height and, by extension, the measurement of volumes and areas

Photometry

measurement is quantitative assessment of image brightness.

Delineation

delineate, or outline, regions as they are observed on remotely sensed images.

• - SIze - Shape - Tone/Hue/Color - Texture - Shadow - Pattern - Association - Site/Location

Elements of Image Interpretation

1. . Look for a scale

2. Look for patterns, shapes, and textures

3. Define the colors (including shadows)

4. Find North

5. Consider your prior knowledge

5 tips and strategies for Image Interpretation

Image Classification

The task of assigning pixels in an image to specific categories or classes of interest (e.g., land cover types)

Unsupervised classification and Supervised classification

There are 2 classification methods:

Unsupervised Classification

• involves the algorithm dividing/classifying the image into classes without using any training samples

• The algorithm identifies clusters of pixels with similar spectral signatures without prior knowledge of what those clusters represent.

Supervised Classification

• involves the use of training samples/data with known classes to train a classification algorithm.

• The algorithm "learns" the spectral signatures of different classes based on the p

Accuracy Assessment

• Quantitatively assess how well a classification worked.

• Evaluate errors for each class, ultimately, determining the reliability of the produced map.

• Aerial photo

• Ground truth data (with GPS information)

• Other GIS files

Some possible reference data include:

true

T/F. The dates on the remote sensing data (e.g. satellite data used for classification) and your validation data must be the same or does not have a significant gap.

Create our own validation data.

What if ground truth data is unavailable?

Human errors such as, but not limited to:

a. Misinterpretation of some features (due to low resolution image) b. Personal biases in selecting the regions of interest or ROIs.

But there are some implications of using this method (creation of own validation data)

• Difference in resolution (Landsat 8 has 30m spatial resolution)

• Map products like Google Earth “mosaics” several images from different time frames to produce a “cloudless” data.

Why not use high resolution imagery like Google Earth or Bing?

error matrix

• Examines the relationship between the reference data and the predicted outputs per class.

• One of most common methods to evaluate accuracy.

confusion matrix or contingency table

the error matrix is also called as 

Reference Class; Predicted Class

In error matrix, A table of nxn cells corresponding to the number of classes, where the columns represent the —and the rows represent the —.

Total Accuracy, User’s Accuracy, Producer’s Accuracy, and Kappa Statistics

What are the types of Metrics

. Total (Overall) Accuracy

Calculated by dividing the total number of correctly classified samples by the total number of reference pixels

Formula for Total Accuracy

Only shows the overall performance of the classifier (average) and does not give insight on the distribution of error along all classes.

What is the main problem in total accuracy

User’s Accuracy

Corresponds to the error of commission (inclusion) - considers the number of pixels which does not actually belong to that specific classification.

Formula for User’s Accuracy

the class is over-represented

• what does a low user’s accuracy mean?

• —> many pixels labeled as that class actually belong to others.

Producer’s Accuracy

• Corresponds to the error of omission (exclusion) - considers the number of pixels not included in the correct classification.

• Represents the probability that a reference (ground truth) pixel of a given class was correctly mapped in the classification.

• It evaluates how well the classifier captured actual ground conditions.

Formula for Producer’s accuracy

many reference samples of that class were placed in other categories on the map.

What does a low producer’s accuracy

Kappa Statistic

• Is a discrete multivariate method for accuracy assessment.

• considers the expected agreement by chance, taking into account the class distribution and preventing biased results. (ex more vegetation than bare soil)

Formula for Kappa Statistic

k

In Kappa Statistic, this is a measure of agreement of accuracy between two measures.

+1.0

Kappa Statistic Interpretation:

• A value of — indicates perfect agreement,

0

Kappa Statistic Interpretation:

• indicates agreement equivalent to random chance,

negative values

Kappa Statistic Interpretation:

• negative values indicate less agreement than expected by chance

-1.0

Kappa Statistic Interpretation:

• is a perfect disagreement.

• Incorrect training data

• Misregistration between the map and reference imagery

• Poor classifier performance (e.g., wrong spectral signatures)

• Heavily unbalanced class distributions causing spurious results

A negative k often arises from:

Positive kappa