Multi-Classifier Comparison using eCognition

The objective of this exercise was to compare the various image classifiers that the eCognition software offers. Five (5) classifiers were used: Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Random Trees (Random Forest), Bayes, and the Decision Tree.

Area of Interest

The area of interest was the region surrounding a section of the Salzach River, located along the Austrian (right) and German (left) border, right above the city of Salzburg.

A subset of Planet's high-resolution surface reflectance imagery with RGB and NIR bands was used.

Image Segmentation

The first step was to segment the image to facilitate the collection of training samples in the next step. Small image segments were chosen to have more detail in features that would be used for training and to improve the classification results.

Training Sample Selection

The area of interest was relatively small; therefore, 10-12 training segments were selected per class. Five classes were specified: arable land, pastures, artificial surfaces, forest and water.

Image segments with selected training samples.

True colour image with overlaid training samples. Top-right panel: classes used in classification. Bottom-right panel: number of samples per class.

Applying the Classifiers

The comparison was done using the default parameters for each classifier. The training samples were also kept the same throughout.

Support Vector Machine (SVM)

Parameters used:

• Features: Mean, and Standard Deviation for Red, Green, Blue and NIR bands

• Kernel type: Linear

• C: 2

 Result of SVM classification. Far-top-right panel: classes used in classification. Near-top-right panel: process tree used. Bottom-right panel: number of samples per class.

K-Nearest Neighbour (KNN)

Parameters used:

• Features: Mean, and Standard Deviation for Red, Green, Blue and NIR bands

• K: 1

Result of KNN classification. TFar-top-right panel: classes used in classification. Near-top-right panel: process tree used. Bottom-right panel: number of samples per class.

Random Trees (Random Forest)

Parameters used:

• Depth, Min sample count, Active variables: 0

• Use surrogates: No

• Max categories: 16

• Max tree number: 50

• Forest Accuracy: 0.01

• Termination criteria type: Both (terminate learning by Max tree number and Forest Accuracy)

Result of Random Trees classification. Far-top-right panel: classes used in classification. Near-top-right panel: process tree used. Bottom-right panel: number of samples per class.

Bayes

Parameters used:

• Features: Mean, and Standard Deviation for Red, Green, Blue and NIR bands

Result of Bayes classification. Far-top-right panel: classes used in classification. Near-top-right panel: process tree used. Bottom-right panel: number of samples per class.

Decision Tree

Parameters used:

• Depth, Min sample count: 0, 0

• Use surrpgates: 0

• Max categories: 0

Result of Decision Tree classification. Far-top-right panel: classes used in classification. Near-top-right panel: process tree used. Bottom-right panel: number of samples per class.

Comparing the results

True colour image

SVM

KNN

Random Trees

Bayes

Decision Tree