Landsat Satellite Image Segmentation Using the Fuzzy ARTMAP Neural Network

Abstract

This application illustrates how the fuzzy ARTMAP neural network can be used to monitor environmental changes. A benchmark problem seeks to classify regions of a Landsat image into six soil and crop classes based on images from four spectral sensors. Simulations show that fuzzy ARTMAP outperforms fourteen other neural network and machine learning algorithms. Only the k-Nearest-Neighbor algorithm shows better performance (91% vs. 89%) but without any code compression, while fuzzy ARTMAP achieves a code compression ratio of 6:1. Even with a code compression ratio of 50:1 fuzzy ARTMAP still maintains good performance (83%). This example shows how fuzzy ARTMAP can combine accuracy and code compression in real-world applications.