Aerial Image Segmentation Workbench

Upload a training aerial image, label pixels by class, train a CART-like decision tree, segment a second image, and inspect the model with EDA, feature importance, tree construction, confusion matrix, and ROC curves.

Status: Load images to begin.
Selected Class: Roads

Controls

Classes

How to use:

  1. Upload a training aerial image.
  2. Select or create a class and click representative pixels or small areas in the training image.
  3. Train the decision tree.
  4. Upload a second image and click Segment test image.
  5. For confusion matrix and ROC, click validation points on the test image using the true class.
  6. Use mouse wheel over canvases to zoom. Use Shift + wheel for horizontal pan and Alt + wheel for vertical pan.

Image Workspace

Training Image

100%

Test Image (click here for validation labels)

100%

Segmentation Output

Prediction distribution shown below.

Sample / Model Summary

Dataset Summary

Model Summary

Validation Metrics

EDA

Class Distribution

Mean Features by Class

Brightness (V channel) Histogram

Feature Statistics

Decision Tree

Tree Construction Plot

100%

Feature Importance

Split Construction Log

Evaluation

Confusion Matrix

ROC Curves (One-vs-Rest)

Confusion Matrix Table