TIPS

Creating a Training Set

Feature Analyst® learns from a small and simple set of training examples (i.e. sample features hand-digitized by the user) and classifies the remainder of the image. When extracting features of interest, the GIS analyst draws representative samples that identify the important aspects of the object: color, orientation, surroundings, etc. Drawing training polygons that clearly identify the feature of interest will return clean results with a minimum of clean up required.

Follow these rules of thumb for drawing a good training set:

1. Neatly trace the border. In the airport example, a good training set includes polygons that extend to the edge of the airplane, but not past. This trains Feature Analyst to extract pixels right up to the outline of your features of interest. Zoom in on the image to better analyze the feature you are extracting.

2. Capture the variety. Select representative samples throughout the image to give Feature Analyst a variety of examples to learn from. In the example, the selected airplanes are scattered about the image and they represent the range in color, size shape, etc.

In the above example, the three selected training polygons illustrate a good sample set that represents the variety in the airplane color, brightness, contrast, size, shape, background, etc. Feature Analyst takes all of these aspects into account when searching for similar features.

After drawing your training set, use the Feature Analyst Set Up Learning tool to define the type of feature you are looking for. In the Feature Selector category, choose Manmade Feature (>5 m). In Advanced Learning Settings, under the Learning Options tab, set the aggregate areas to a minimum area of 200 pixels.

Drawing a Bad Training Set

The quality of the training set makes the difference between great initial results and terrible results. If the training set is poorly digitized and not representative of the target features, there is no combination of learning parameters (input representation, pattern size, etc.) that will give good results. The following examples illustrate: (1) Careless Training; and (2) Inadequate Training.

Bad training examples consist of the following:

1. Careless Training - poorly drawn examples result in misidentified features. The “Careless Training” example is loosely digitized. Not surprisingly, the results from the bad training set contain large amounts of clutter and include false positives in the buildings and paved surfaces.

2. Inadequate Training - too few examples results in insufficient feature extraction. The “Inadequate Training” example consists of one poorly drawn polygon that does not closely follow the contour of the airplane.

Click here to view the print version of this Tip of the Week.