Lab Three, Object-based image analysis & machine learning classifiers

Background and Goal:

        The goal of this lab was twofold. One, to develop knowledge in Object-based image analysis and two, gain experience with the eCognition software. The object-based image analysis was done with machine learning, namely SVM and Random Forest classifiers. 

Methods:

        Through the lab, there were classifications done three times. However, as the general process is the same I will be highlighting only the methods of one classification.

Drone imagery in false-color

        The first step is to segment the image into objects. Above can be seen the base image. We achieved segmentation using the multiresolution segmentation process in eCognition.

Drone imagery objects

        From here, we then selected training samples to train the classifier.

Training Samples

        Now all that is left to do is to train and run the classifier. In the above case, I classified the data with a SVM classifier.

Results:

        The classifier will result in an image similar to the one below.

   Result of Classification    

        The result was not perfect. The classifier had a hard time distinguishing vehicles, but it did a very good job classifying the road, as well a differentiating between trees and lawn. 

    When I ran the other classifiers in my lab, I got significant error in my classification. I believe that this error stemmed from my training samples, namely selecting trees un urban areas as a training sample for forests. This led to a vast overrepresentation of the actual downtown Eau Claire area.

Eau Claire area classified with Random Forest

        To avoid this in the future I will be much more carful selecting training samples.

Sources:

Planet Team.(2021).[Planet satellite image].Planet Application Program Interface. In Space of life on                Earth, San Francisco, CA. Provided by Cyril Wilson.

Department of Geography and Anthropology.(2020).[Drone imagery]. University of Wisconsin Eau-                   Claire. Provided by Cyril Wilson.