Spatial Dynamics

I realize this message is only useful to a handful of people, but it can seem quite daunting and bewildering when first using Classification and Regression Trees ( CART) algorithms for classifying imagery,

I’ve been utilizing CART as a method of supervised classification for a few years. Not only is it the primary tool for the USGS in the creation of NLCD datasets and NOAA for its CCAP classification, it’s one of the most efficient methods for classifying imagery available.

Currently, CART seems to be the de facto algorithm for classifying Landsat imagery among U.S. Government organizations. CART could be considered unique compared to most supervised classification techniques since results are (generally) empirical and reproducible.

One of the most time-consuming aspects (as would be with any method) is resampling datasets to have compatible properties (projection, extent, cell size, etc). While the construction of the decision trees only takes a few minutes, the classification usually takes about an hour or so, depending on the size of the dataset and processor capability.

For me, there are three necessary tools to run CART:

1. ERDAS Imagine 8.x or 9.x

2. The USGS’s NLCD Mapping Tool

- Its free and designed for ERDAS Imagine. A decent but old PowerPoint is available. but skip to slide 15

3. See5/Cubist software

-Trial versions are available (with extreme limitations on sampling size).

While See5 and Cubist cost $900 individually, it’s likely the smallest, yet most valuable and expensive <1Mb algorithm you’ll ever purchase.

Tags:   CART · Cubist · Erdas IMAGINE · regression analysis · See52 Comments