Abstract:
Self-organization is the core principle of all learning in Adaptive Resonance Theory (ART), which has been highly successful in accounting for biological visual learning or biologically plausible computational modeling of visual processing. Such processing involves the analysis of visual data that may not be consciously visible, such as changes in fine visual detail in images related to alterations in natural or urban landscapes. Over time, this imaging data may reveal structural changes resulting from measurable human impact or climate change. Capturing these changes in time series of satellite images before they are perceptible to the human eye makes them available at early stages to citizens, professionals, and policymakers. This promotes change awareness and facilitates early decision-making for action.
In this context, unsupervised Artificial Intelligence (AI) is employed, leveraging principles of self-organized biological visual learning for the analysis of time series of satellite images. The Quantization Error (QE) in the output of a Self-Organizing Map prototype serves as a computational metric of variability and change. This neural network metric is sensitive to the intensity and polarity of image pixel contrast and selective to pixel color, proving effective in capturing critical changes in urban landscapes.
Illustratively, satellite images from two regions of geographic interest in Las Vegas County, Nevada, USA, spanning the years 1984-2008, are used. The SOM-QE analysis is integrated with the statistical analysis of demographic data, revealing correlations between human impacts and structural changes in specific regions of interest. By correlating the impact of human activities with the structural evolution of urban environments, this approach expands SOM-QE analysis as a parsimonious and reliable AI approach to the rapid detection of human footprint-related environmental change
