Frequency distribution signatures and classification of within-object pixels.

Area of Science:

• Remote Sensing
• Geographic Information Systems (GIS)
• Image Analysis

Background:

• Geographic object-based image analysis (GEOBIA) commonly classifies image objects using simple statistical measures of pixel values.
• Existing methods often overlook the full frequency distribution of pixel data within objects.
• This limits the detailed characterization of earth surface features.

Purpose of the Study:

• To explore variability in pixel frequency distributions within image objects.
• To evaluate a classification method using full histogram signatures versus standard statistical characteristics.
• To assess the effectiveness of histogram signatures for mapping land cover, land use, and socioeconomic status.

Main Methods:

• Utilized high spatial resolution Quickbird satellite multispectral data for Accra, Ghana.
• Examined frequency distributions (histograms) of multispectral pixel values within image objects.
• Compared curve matching of histogram signatures against nearest neighbor classifiers using parametric statistics.

Main Results:

• Image objects representing land cover and land use exhibit distinct, often non-normal, frequency distributions.
• Histogram signatures demonstrated a close match to training data for most image objects.
• Curve matching of histogram signatures showed slightly superior classification performance compared to standard statistical classifiers.

Conclusions:

• Full pixel frequency distributions (histogram signatures) provide valuable information for GEOBIA.
• Histogram-based classification offers a more nuanced approach to mapping land cover and land use.
• This method holds potential for improved socioeconomic status mapping using remote sensing data.