ARTMAP neural networks for information fusion and data mining: map production and target recognition methodologies

Area of Science:

• Computer Science
• Artificial Intelligence
• Geospatial Analysis

Background:

• MIT Lincoln Laboratory developed a hierarchical system for geospatial image analysis using an ARTMAP neural network.
• The existing system is limited to target/non-target identification and does not generate comprehensive maps.

Purpose of the Study:

• To extend the capabilities of existing geospatial image analysis systems.
• To develop a spatial data mining system capable of identifying and mapping arbitrarily many target classes.
• To address challenges posed by highly skewed class distributions in mapping problems.

Main Methods:

• Development of a new mapping methodology building upon the ARTMAP neural network.
• Implementation of canonical algorithms and a benchmark testbed for evaluating recognition networks and processing options.
• Utilizing training pixels from a spatially distinct region, potentially with different class distributions than the target map.

Main Results:

• The new system successfully learns to identify and distribute numerous target classes.
• The methodology is designed to handle highly skewed class distributions effectively.
• Candidate recognition networks, pre-processing, post-processing, and feature selection options were evaluated.

Conclusions:

• The developed spatial data mining system offers a significant advancement over previous methods for geospatial image analysis.
• The methodology establishes a new standard for various spatial data mining tasks, particularly those with imbalanced datasets.
• The default ARTMAP network, with specified canonical parameters, is a suitable choice for this family of neural networks in diverse applications.