Segmentation of Satellite Imagery Using Hierarchical Thresholding and Neural Networks

View More View Less
  • a Computer Sciences Corporation, Monterey California
  • | b Naval Research laboratory, Monterey, California
© Get Permissions
Full access

Abstract

A significant task in the automated interpretation of cloud features on satellite imagery is the segmentation of the image into separate cloud features to be identified. A new technique, hierarchical threshold segmentation (HTS), is presented. In HTS, region boundaries are defined over a range of gray-shade thresholds. The hierarchy of the spatial relationships between collocated regions from different thresholds is represented in tree form. This tree is pruned, using a neural network, such that the regions of appropriate sizes and shapes are isolated. These various regions from the pruned tree are then collected to form the final segmentation of the entire image.

In segmentation testing using Geostationary Operational Environmental Satellite data, HTS selected 94% of 101 dependent sample pruning points correctly, and 93% of 105 independent sample pruning points. Using Advanced Very High Resolution Radiometer data, HTS correctly selected 90% of both the 235-case dependent sample and the 253-case independent sample pruning points.

The strength of this approach is that artificial intelligence, that is, reasoning about the sizes and shapes of the emergent regions, is applied during the segmentation process. The neural network component can be trained to respond more favorably to shapes of interest to a particular analysis problem.

Abstract

A significant task in the automated interpretation of cloud features on satellite imagery is the segmentation of the image into separate cloud features to be identified. A new technique, hierarchical threshold segmentation (HTS), is presented. In HTS, region boundaries are defined over a range of gray-shade thresholds. The hierarchy of the spatial relationships between collocated regions from different thresholds is represented in tree form. This tree is pruned, using a neural network, such that the regions of appropriate sizes and shapes are isolated. These various regions from the pruned tree are then collected to form the final segmentation of the entire image.

In segmentation testing using Geostationary Operational Environmental Satellite data, HTS selected 94% of 101 dependent sample pruning points correctly, and 93% of 105 independent sample pruning points. Using Advanced Very High Resolution Radiometer data, HTS correctly selected 90% of both the 235-case dependent sample and the 253-case independent sample pruning points.

The strength of this approach is that artificial intelligence, that is, reasoning about the sizes and shapes of the emergent regions, is applied during the segmentation process. The neural network component can be trained to respond more favorably to shapes of interest to a particular analysis problem.

Save