Source-linked AI summary
Convolutional Neural Networks for Global Human Settlements Mapping from Sentinel-2 Satellite Imagery
Christina Corbane, Vasileios Syrris, Filip Sabo, Panagiotis Politis, Michele Melchiorri, Martino Pesaresi, Pierre Soille, Thomas Kemper
TL;DR
Reliable, up-to-date global built-up maps are needed for urbanization and sustainability research and policy, while existing products face accuracy, scale, and coverage limitations. The paper develops GHS-S2Net, a multi-model CNN framework applied to a 10 m, 2018 cloud-free Sentinel-2 composite. Validation across 277 sites supports the reliability and robustness of the resulting global built-up layer, including in remote African and Asian areas.
Problem
Global built-up mapping needs spatially consistent, up-to-date information, but existing approaches face accuracy, sensor-scale, settlement-density, and continuous-monitoring limitations.
Method
GHS-S2Net uses a multi-model CNN framework with Sentinel-2 imagery and detailed public built-up-area training data for global pixel-wise mapping.
Results
Independent building-footprint validation across 277 sites establishes the reliability and robustness of the produced built-up layer, including classification in remote areas of Africa and Asia.
Takeaways & Limitations
The model outputs can serve as proxy measures for building densities without additional calibration or modeling.
Takeaways & Limitations
The applicability of the CNN models is constrained by the variable quality, spatial resolution, consistency, and accuracy of available training data.
Abstract
from arXiv · showhide
Spatially consistent and up-to-date maps of human settlements are crucial for addressing policies related to urbanization and sustainability, especially in the era of an increasingly urbanized world.The availability of open and free Sentinel-2 data of the Copernicus Earth Observation program offers a new opportunity for wall-to-wall mapping of human settlements at a global scale.This paper presents a deep-learning-based framework for a fully automated extraction of built-up areas at a spatial resolution of 10 m from a global composite of Sentinel-2 imagery.A multi-neuro modeling methodology building on a simple Convolution Neural Networks architecture for pixel-wise image classification of built-up areas is developed.The core features of the proposed model are the image patch of size 5 x 5 pixels adequate for describing built-up areas from Sentinel-2 imagery and the lightweight topology with a total number of 1,448,578 trainable parameters and 4 2D convolutional layers and 2 flattened layers.The deployment of the model on the global Sentinel-2 image composite provides the most detailed and complete map reporting about built-up areas for reference year 2018. The validation of the results with an independent reference data-set of building footprints covering 277 sites across the world establishes the reliability of the built-up layer produced by the proposed framework and the model robustness.
1 Introduction
Global built-up-area mapping remains constrained by accuracy, sensor dependence, settlement-density extremes, and the lack of consistent automated monitoring. The paper motivates GHS-S2Net as a CNN-based, Sentinel-2 framework for 10 m global mapping.
- 1 Introduction: Existing built-up products and methods remain limited by accuracy errors, sensor-scale dependency, poor representation of settlement-density extremes, and inconsistent monitoring across time and sensors.The stated problems include confusion with non-urban surfaces and disagreement among sensor-derived land-cover estimates.
- 1 Introduction: Sentinel-2 provides a five-day revisit cycle, 13 spectral bands, and spatial resolution as high as 10 m, creating opportunities for global built-up-area mapping.The paper identifies novel methods as necessary to exploit these spatial and temporal characteristics.
- 1 Introduction: CNNs can perform pixel-wise image classification while naturally modeling hierarchical image features, supporting built-up-feature extraction from satellite imagery.The paper relates this capability to CNN adaptation from image classification to segmentation.
- 1 Introduction: Prior CNN work has addressed built-up mapping across several sensors, but large-scale mapping from lower-resolution Sentinel-2 data remains comparatively underexplored.A previous 20 m global-scale framework is described as failing to implement its CNN model in rural areas.
- 1.2 Challenges addressed in this work: The proposed framework targets global 10 m built-up mapping with a probability-grid output and combines UTM-zone multi-modeling with close- and far-range transfer learning.The contributions also include systematic two-stage sampling and independent fine-scale validation.
2 Input data and methods
The method uses a global cloud-free Sentinel-2 composite for 2018, generated through pixel-based compositing, together with detailed public built-up-area datasets for model training.
- 2.1 Sentinel-2 cloud-free image composite: The input is a global cloud-free Sentinel-2 composite for reference year 2018 covering the landmass except Antarctica.Sentinel-2 is selected for fine-scale mapping and monitoring because of its five-day revisit time and decametric resolution.
- 2.1 Sentinel-2 cloud-free image composite: Pixel-based compositing uses the 25th percentile of pixel time series to mitigate cloud, aerosol, and data-gap contamination in large-area optical imagery.The composite was generated and exported from Google Earth Engine.
- 2.1 Sentinel-2 cloud-free image composite: The composite is organized by UTM grid zones, matching the spatial organization used for the global mapping framework.The figure presents the cloud-free Sentinel-2 composite in this UTM-zone arrangement.
2.2 Model input data: learning sets
The learning sets combine publicly available built-up-area products with different resolutions, coverage, dates, and quality issues to train global Sentinel-2 classification models. A hierarchical selection process prioritizes the most detailed and locally suitable sources.
- 2.2 Model input data: learning sets: The training data combine GHSL_BU, ESM_BU, FB_HRS, and MS_BFP products describing built-up areas at different spatial resolutions and geographic coverages.GHSL_BU is derived from 30 m Landsat data, ESM_BU provides a 2 m European layer, FB_HRS supplies 30 m settlement grids, and MS_BFP provides detailed building footprints.
- 2.2 Model input data: learning sets: The sources differ in temporal coverage and error patterns, including occasional rural commission errors and systematic omissions in large urban areas for FB_HRS.FB_HRS imagery spans 2002–2017 with a surface-weighted average around 2013, and its rural precision contrasts with omission errors in large urban areas.
- 2.2 Model input data: learning sets: Training-set characteristics include spatial resolution, geographic coverage, source-image dates, identified issues, and total and relative numbers of training pixels.These characteristics are summarized in Table 1, while Figure 3 shows the geographic distribution of selected information sources.
- 2.2 Model input data: learning sets: The hierarchical selection prioritizes MS_BFP and ESM_BU, then FB_HRS, and finally GHSL_BU as the least detailed built-up representation.The ordering reflects the goal of selecting the best locally available proxies for built-up areas at 10 m resolution.
2.3 GHS-S2Net building blocks
GHS-S2Net is a pixel-wise CNN framework for 10 m built-up-area classification using local image patches and geographically distributed models. Its lightweight architecture and two-stage training strategy are designed to capture fine settlement details while adapting to regional variation.
- 2.3 GHS-S2Net building blocks: GHS-S2Net classifies each 10 m pixel as built-up or non-built-up using the paper’s GHSL-consistent definition of built-up area.Pixel-wise grouping is treated as image segmentation, with each output corresponding to an individual central pixel.
- 2.3 GHS-S2Net building blocks: The network processes 5 x 5 pixel image patches to capture settlement structures ranging from single residences to contiguous building blocks.The design targets fine information in individual pixels and small homogeneous pixel groups rather than large natural-image regions.
- 2.3 GHS-S2Net building blocks: The lightweight topology contains 4 2D convolutional layers and 2 flattened layers, with 1,448,578 total parameters and 1,447,042 trainable parameters.The architecture uses a high number of filters to capture subtle details while remaining smaller than indicative CNNs such as VGGNet and GoogleNet.
- 2.3 GHS-S2Net building blocks: The convolutional blocks use linear and tanh activations, while the final sigmoid maps the output to a [0,1] probability that the central pixel is built-up.The authors report that tanh captured Sentinel-2 feature complexity better in their experiments with a smaller number of layers.
- 2.3.1 Two-stage training approach: Two-stage training fits one model per UTM grid zone, using geographically selected patches and transfer learning to model local Sentinel-2 and settlement variation.The approach combines close-range and far-range parameter transfer, while training samples include built-up patches and randomly sampled non-built-up patches.
2.4 Processing infrastructure
The computing-intensive training and prediction workflow ran on the JEODPP infrastructure, which combines large-scale storage with heterogeneous GPU resources. Experiments used two GPU nodes with Quadro RTX 6000 and Tesla V100 modules.
- 2.4 Processing infrastructure: JEODPP provides multi-petabyte storage alongside computational resources using commodity hardware and open-source software.The infrastructure is described as having 14 PiB of storage and an EOS-based software stack.
- 2.4 Processing infrastructure: Training and prediction used two GPU nodes equipped with four Quadro RTX 6000 GPUs and two Tesla V100-PCIE GPUs.The Quadro modules had 24.2 GB of memory each, while the V100 modules had 32.5 GB each.
3.1 Training phase of CNN models per UTM grid zone
The CNN models were trained with held-out validation data, normalized inputs, Adam optimization, and binary cross-entropy. Across 485 UTM-zone models, training and validation losses converged rapidly with little separation, supporting the reported generalization behavior.
- 3.1 Training phase of CNN models per UTM grid zone: Ten percent of each training set was reserved for validation to evaluate classification performance and help prevent overfitting.The same validation fraction was used during training across the UTM-zone models.
- 3.1 Training phase of CNN models per UTM grid zone: The models used Sentinel-2 inputs rescaled to [0,1], 25 training epochs, Adam optimization with learning rate 0.0001, and binary cross-entropy loss.Weights were initialized from a uniform distribution with bounds [-0.1065, 0.1065].
- 3.1 Training phase of CNN models per UTM grid zone: Training and validation losses for 485 UTM-zone models converged around 12 epochs, with the gap between curves disappearing after approximately 12 iterations.The models were trained for 25 epochs, and the reported curve behavior indicates stable convergence with limited overfitting.
- 3.1 Training phase of CNN models per UTM grid zone: The 485 trained models correspond to land-containing UTM zones with built-up presence in the learning sets, out of 615 total landmass zones.The learning curves are interpreted by the authors as evidence that the two-stage training-set sizes provide good generalization capacity.
3.2 Computational performance of the GHS-S2Net models during the training and prediction phases
The GHS-S2Net multi-modeling approach was computationally scaled across UTM grid zones using distributed GPUs, with processing times varying between inland- and water-dominated zones.
- 3900 s for training and 16,400 s for prediction were required on average for inland zones, compared with 3100 s and 15,000 s for water zones.The differences reflect the larger number of training samples typically available in inland zones.
- Global processing scaled across 485 land-containing UTM grid zones despite limited concurrent GPU availability.Six GPU modules were used for training and two for prediction.
- Efficient UTM partitioning, two-stage sampling, and 100 x 100 km tiles enabled global 10 m processing, while more GPUs and early stopping could reduce runtime.
3.3 Qualitative assessment of the models predictions
Qualitative assessment showed that GHS-S2Net refined built-up-area detection relative to its training sets, reducing errors and revealing settlement structure and open spaces.
- GHS-S2Net reduced commission and omission errors and artifacts relative to the training sets while refining built-up areas and open spaces.
- The model detected new settlements absent from the training sets and other global datasets, including fine settlement patterns omitted by earlier products.
- Visual comparisons showed reduced overdetection over bare rocks, agricultural fields, sand dunes, rocky beaches, bare soils, and riverbeds.
- In Sassari, probabilistic outputs nearly identified individual buildings and open spaces, suggesting sensitivity to built-up density patterns.
- The models used 1,447,042 trainable parameters and 511,502,073 built-up patches while remaining robust to noise or missing training-set data.
3.4 Validation of the model predictions and assessment of generalization performance
Independent building-footprint validation assessed GHS-S2Net both as a continuous predictor of built-up density and as a thresholded binary classifier across 277 global areas of interest.
- 3.4 Validation of the model predictions and assessment of generalization performance: Validation combined regression of model probabilities against building densities with binary confusion-matrix assessment after probability thresholding.
- 3.4 Validation of the model predictions and assessment of generalization performance: The reference database contained more than 40 million building polygons from 277 areas, with reference years spanning 2012–2018.
- 3.4.1 Continuous assessment: validation of the model output as predictor of built-up densities: Threshold selection depended on settlement density: 0.2 better represented scattered and rural areas, whereas 0.5 suited high-density city centers.
- 3.4.2 Binary accuracy assessment: A 0.2 threshold consistently outperformed 0.5 across continents, with average Balanced Accuracy above 0.7 and Kappa above 0.5 overall.At threshold 0.2, average Balanced Accuracy reached 0.91 in Oceania and Asia, 0.86 in North America, and 0.85 in Africa.
3.5 Comparison between the results of close range and far range transfer learning
Far-range transfer learning was used where detailed local training data were scarce and produced stronger mapping results than close-range transfer learning in the evaluated Moscow comparison.
- The approach addressed zones where only low-resolution GHS_BU training data were available by generalizing from nearby zones with detailed samples.
- In Moscow, transferring from a detailed-data zone reproduced fine settlement structures, reduced road and impervious-feature overdetection, and sharpened building and open-space boundaries.
- Far-range transfer learning produced higher overall and balanced accuracy than close-range transfer learning in Moscow.The comparison used binary outputs with probability cutoffs of 0.2 and 0.5.
- Far-range transfer learning was applied to 28 UTM grid zones, selected mainly by spatial adjacency or landscape and settlement similarities.
4 Discussion and future work
GHS-S2Net combines multiple lightweight CNN models, transfer learning, and high-throughput processing to produce a global 10 m built-up-area map. Validation indicates robust classification and a relationship between output probabilities and building density, while training-data quality and patch-size choices constrain future updates.
- Framework: The framework combines UTM-zone multi-model training, systematic sampling, far-range transfer learning, and high-throughput multi-GPU processing for global-scale mapping.The infrastructure stores 15 TB of inputs and 1.5 TB of outputs while parallelizing model training and inference.
- Results: Validation across 277 worldwide sites establishes GHS-S2Net reliability and robustness across variable imagery, landscapes, and built-up characteristics.The model also classified remote built-up areas in Africa and Asia that were not reported in several existing global products.
- Results: GHS-S2Net outputs provide pixel-level built-up probabilities, with a strong relationship to building densities that supports their use as density proxies without additional calibration or modeling.
- Limitations: Testing a 10 × 10 patch produced a deeper topology without improving loss, while prediction accuracy worsened, constraining the benefit of larger contextual inputs.
- Limitations: Variable training-data quality produces outputs with inconsistent quality, limiting CNN applicability and restricting the effectiveness of far-range transfer learning at global scale.Reference sources differ in spatial resolution, completeness, consistency, accuracy, and sample trustworthiness.
- Future applications: The global Sentinel-2 deployment produces probability outputs for reference year 2018 and supports frequent updates and continuous monitoring of built-up areas.
Open Access
The article is openly licensed under Creative Commons Attribution 4.0 International, permitting reuse with attribution and specified licence conditions.
- Open Access: The Creative Commons Attribution 4.0 licence permits use, sharing, adaptation, and reproduction in any medium or format.
- Open Access: Reuse requires crediting the original authors and source and providing a link to the Creative Commons licence.
- Open Access: Users must indicate whether changes were made when adapting or reproducing the article or its covered materials.