Estimating generalized measures of local neighbourhood context from multispectral satellite images using a convolutional neural network
Alex Singleton; Dani Arribas-Bel; John Murray; Martin Fleischmann (2022). Computers, Environment and Urban Systems, 95, 101802. DOI: 10.1016/j.compenvurbsys.2022.101802
Abstract
The increased availability of high-resolution multispectral imagery captured by remote sensing platforms provides new opportunities for the characterisation and differentiation of urban context. The discovery of generalized latent representations from such data are however under researched within the social sciences. As such, this paper exploits advances in machine learning to implement a new method of capturing measures of urban context from multispectral satellite imagery at a very small area level through the application of a convolutional autoencoder (CAE). The utility of outputs from the CAE is enhanced through the application of spatial weighting, and the smoothed outputs are then summarised using cluster analysis to generate a typology comprising seven groups describing salient patterns of differentiated urban context. The limits of the technique are discussed with reference to the resolution of the satellite data utilised within the study and the interaction between the geography of the input data and the learned structure. The method is implemented within the context of Great Britain, however, is applicable to any location where similar high resolution multispectral imagery are available.
Extended Summary
This research develops a novel method for measuring local neighbourhood context from high-resolution satellite imagery using machine learning techniques. The paper addresses the challenge of characterising urban spatial structure at a very fine scale by applying a convolutional autoencoder (CAE) to multispectral satellite data from Copernicus Sentinel 2 across Great Britain. The methodology processes 10-metre resolution satellite imagery covering 1,710,715 unit postcodes, creating 160m × 160m square buffers around each postcode centroid. Each buffer contains four spectral bands: near-infrared, red, green, and blue. The research implements a convolutional autoencoder architecture that learns latent representations from these satellite images without requiring pre-defined labels, making it an unsupervised learning approach. The CAE reduces the dimensionality of each 16×16×4 pixel image to 64 abstract features representing different aspects of urban context. To address the initial sensitivity of the model that produced fragmented classifications, spatial smoothing was applied by calculating weighted averages for each postcode using values from neighbouring postcodes within the 160m buffer. This process created more spatially contiguous representations of context. The smoothed outputs were then subjected to k-means cluster analysis, with seven clusters identified as optimal through clustergram analysis. The resulting classification distinguishes between different types of urban context, including inner city and town centres, spacious residential properties, high-density urban cores, rural residences, rural properties and farms, inner suburban townhouses, and outer suburbs. External validation using ancillary data demonstrates that the clusters effectively differentiate between various built environment characteristics such as housing types, building ages, road network features, and vegetation indices. The research shows particular strength in distinguishing morphometric aspects of urban structure, including street connectivity measures and building density patterns. However, the study acknowledges limitations in differentiating socio-economic characteristics, which appears constrained by the 10-metre resolution of the input data. This work contributes to the broader field of geodemographics and urban morphology analysis by providing a scalable, data-driven approach to measuring local context. The methodology offers significant potential for urban planning applications, policy targeting, and neighbourhood classification systems, particularly in contexts where traditional census or survey data may be limited or outdated.
Key Findings
- Convolutional autoencoders successfully extract meaningful urban context measures from 10-metre resolution satellite imagery across Great Britain’s 1.7 million postcodes.
- Spatial smoothing of raw neural network outputs produces more spatially contiguous neighbourhood classifications compared to unprocessed results.
- Seven distinct urban context types were identified, ranging from inner city centres to rural properties, validated through independent datasets.
- The method effectively captures morphometric urban features like street connectivity but shows limitations in distinguishing socio-economic characteristics.
- Higher resolution satellite data would likely improve the technique’s ability to differentiate between areas with similar built environments but different social characteristics.
Citation
@article{singleton2022estimating,
author = {Alex Singleton; Dani Arribas-Bel; John Murray; Martin Fleischmann},
title = {Estimating generalized measures of local neighbourhood context from multispectral satellite images using a convolutional neural network},
journal = {Computers, Environment and Urban Systems},
year = {2022},
volume = {95},
pages = {101802},
doi = {10.1016/j.compenvurbsys.2022.101802}
}