Source-linked AI summary
Using convolutional networks and satellite imagery to identify patterns in urban environments at a large scale
Adrian Albert, Jasleen Kaur, Marta Gonzalez
TL;DR
Urban land-use data are costly to collect and scarce in many regions. This paper combines satellite imagery, Urban Atlas labels, and deep convolutional networks to classify urban environments and compare neighborhoods through learned representations. It constructs an open dataset and reports encouraging classification and neighborhood-comparison results, while image and city variation constrain the data.
Problem
Urban land-use information needed for planning and infrastructure applications is labor-intensive and costly to collect, and public aerial-imagery benchmarks remain limited.
Method
The paper pairs Urban Atlas land-use labels with satellite imagery, trains VGG-16 and ResNet classifiers, and uses extracted features to compare urban environments.
Results
The classifiers achieve encouraging land-use classification results, and learned features identify visually similar neighborhoods across cities.
Takeaways & Limitations
An open dataset of over 140,000 samples across 10 classes and 6 European cities supports further remote-sensing and urban-environment research.
Takeaways & Limitations
Satellite appearances vary with season, physical form, and landscape change, and selected cities required at least 50 samples per class.
Abstract
from arXiv · showhide
Urban planning applications (energy audits, investment, etc.) require an understanding of built infrastructure and its environment, i.e., both low-level, physical features (amount of vegetation, building area and geometry etc.), as well as higher-level concepts such as land use classes (which encode expert understanding of socio-economic end uses). This kind of data is expensive and labor-intensive to obtain, which limits its availability (particularly in developing countries). We analyze patterns in land use in urban neighborhoods using large-scale satellite imagery data (which is available worldwide from third-party providers) and state-of-the-art computer vision techniques based on deep convolutional neural networks. For supervision, given the limited availability of standard benchmarks for remote-sensing data, we obtain ground truth land use class labels carefully sampled from open-source surveys, in particular the Urban Atlas land classification dataset of $20$ land use classes across $~300$ European cities. We use this data to train and compare deep architectures which have recently shown good performance on standard computer vision tasks (image classification and segmentation), including on geospatial data. Furthermore, we show that the deep representations extracted from satellite imagery of urban environments can be used to compare neighborhoods across several cities. We make our dataset available for other machine learning researchers to use for remote-sensing applications.
1 INTRODUCTION
The paper addresses the costly scarcity of urban land-use data by combining widely available satellite imagery with deep convolutional networks. It introduces a dataset and uses learned representations to compare urban environments.
- Urban land-use classification supports planning, zoning, permitting, real-estate, and infrastructure applications.
- Professional surveys make urban land-use data labor-intensive, infrequent, slow, and costly, especially limiting availability in poorer and developing regions.
- Satellite imagery and advances in deep learning enable broader analysis of built infrastructure and urban environments.
- The study compares VGG-16 and ResNet classifiers, constructs an Urban Atlas-based dataset, and uses extracted features to compare urban environments.The dataset covers 10 European cities and 10 consolidated land-use classes, with the data made publicly available.
- Land-use classes combine physical attributes with expert judgments about cultural and socio-economic function.The paper treats these classes as a useful discretization of a more continuous spectrum of urban organization patterns.
2 LITERATURE
Prior work applies remote sensing, machine learning, and deep learning to land-use classification and related urban-analysis tasks. However, public aerial-imagery benchmarks remain limited and existing datasets do not fully support detailed cross-city urban comparison.
- Remote-sensing research spans land-use classification, object detection, segmentation, geolocation, poverty estimation, and urban quality-of-life analysis.
- Deep learning for satellite-image land-use classification is a relatively recent and still developing research area.
- Public benchmark data for aerial-image land-use classification remain in relatively short supply.
- Existing benchmark datasets: DeepSat supports pre-training but does not enable detailed land-use analysis and cross-city comparison of urban environments.
- Other open-source efforts: TerraPattern and DeepOSM pair Google Maps satellite imagery with OpenStreetMap labels for convolutional classification and image similarity or search.
3 THE URBAN ENVIRONMENTS DATASET
The Urban Environments dataset derives standardized Urban Atlas labels and freely available satellite imagery through a sampling strategy designed to balance classes and preserve representative examples. Validation uses city-centered grids, while dataset selection reflects substantial geographic and visual variation.
- Urban Atlas: Urban Atlas covers approximately 300 European cities and 20 standardized land-use classes; this work consolidates selected classes into 10 final classes.
- Urban Atlas: Land-use distributions vary widely across cities, motivating careful sampling for training data.
- Data sampling and acquisition: Google Maps Static API supplies freely available satellite imagery, with up to 25,000 free API requests per day.
- Data sampling and acquisition: The sampling strategy balances highly imbalanced classes and assigns images to polygon labels while seeking representative satellite views.
- Dataset variation: Images within the same class vary because of acquisition season, differing physical forms, and landscape changes since Urban Atlas compilation.
- Validation data: Validation uses 100 × 100 cells over a 25km × 25km city-centered grid, assigning each cell the class with maximum polygon intersection area.
- Validation data: The validation dataset contains approximately 140,000 images across 10 classes from 6 European cities, selected where each class had at least 50 samples.
4 EXPERIMENTAL SETUP
The experiments compare VGG-16 and ResNet architectures, including transfer-learning settings across cities, and use learned representations to study urban-environment similarity.
- Neural network architectures and training: The study compares VGG-16 and ResNet architectures for urban-environment classification.Both architectures are evaluated throughout the experiments.
- Transfer learning: Transfer learning initializes urban-classification models from architectures pretrained on datasets from other domains.The motivation is that pretrained networks already encode basic shapes and patterns, potentially improving performance and training speed.
- Transfer learning: Models are trained on one or more cities and evaluated on different geographical locations to measure cross-city transfer.The experiments test whether urban-environment similarities support transfer across locations.
- Feature extraction: The convolutional networks provide image representations of D = 4096 dimensions for VGG16 and D = 2048 dimensions for ResNet-50.The codes are taken from the last network layer and represent high-level land-use concepts.
- Feature extraction: The extracted representations are used with a KD-tree to quantify similarity among local urban environments across city grids.The similarity analysis focuses on 25km × 25km, 100 × 100-cell grids around city centers.
5 RESULTS AND DISCUSSION
The models qualitatively reproduce urban land-use maps and support cross-city comparison, although performance varies by class, city, and spatial scale. Diverse-city training transfers better across locations, while 250m tiles outperform smaller analyzed scales for high-level land-use concepts.
- Classification results: The model qualitatively recreates urban land-use maps from satellite imagery across six cities.Figure 1 compares ground-truth grids with model predictions on 100 × 100 cells spanning 25km × 25km.
- Classification results: Agricultural, water, industrial, public, commercial, forest, green-urban, low-density, airport, and sports-leisure classes show particularly good qualitative agreement.Road distributions are harder to reconstruct because roads occur across scenes with different planning classifications.
- Transfer learning and classification performance: Training on all cities transfers better to other locations than training on individual cities alone.Pairwise cross-city performance is generally poor, while more diverse training data yields encouraging transferability.
- Classification results: ResNet-50 consistently achieved slightly better classification results than VGG-16 in the reported experiments.The remaining results are therefore reported using ResNet-50.
- Per-class performance: Airports and forests are relatively easier to distinguish, while high-density urban fabric is more easily confused with medium- or low-density fabric.The latter difficulty is especially apparent when comparing similar cities such as Rome and Barcelona.
- Choosing the spatial scale: sensitivity analysis: 50m, 100m, and 150m spatial scales perform worse than the selected 250m scale.Smaller tiles may capture insufficient variation in buildings, vegetation, roads, and other urban-form elements needed for high-level classification.
- Urban-environment similarity: The learned embeddings reveal city- and class-level structure that can support coarse neighborhood comparisons.Most low-density urban, forest, and water classes are well resolved, while sports and leisure facilities blend into other environments.
6 CONCLUSIONS
The paper uses convolutional networks and satellite imagery to classify urban land use, compare neighborhoods, and support low-cost analysis where ground-truth planning data is unavailable.
- Over 140,000 samples from six European cities form an open dataset covering 10 consistent land use classes.
- The method predicts land use classes, embeds convolutional features into a lower-dimensional manifold, and identifies visually similar neighborhood images.
- Classification is more difficult for high-, medium-, and low-density environments, while agricultural land, forests, and airports are more visually consistent across cities.
- Industrial areas show greater inter-city than intra-city similarity, suggesting that learning can transfer across geographies for several classes.
- Satellite imagery may provide a low-cost way to analyze urban environments where ground-truth urban-planning information is unavailable.
A PRACTICAL TRAINING DETAILS.
Training uses an 80/20 training-validation split, separate test-time ground-truth samples, augmented RGB inputs, and stochastic-gradient optimization in Keras with TensorFlow.
- 80% of the data was used for training and 20% for validation, while separately sampled ground-truth raster grids were reserved for testing.
- The networks were implemented in Keras with a TensorFlow backend and optimized using stochastic gradient descent with Adadelta on multi-class cross-entropy.
- Input images were 224 × 224 × 3 RGB pixels and underwent flipping, shearing, scaling, and rotation augmentation.