Source-linked AI summary
Outlining where humans live -- The World Settlement Footprint 2015
Mattia Marconcini, Annekatrin Metz-Marconcini, Soner Üreyen, Daniela Palacios-Lopez, Wiebke Hanke, Felix Bachofer, Julian Zeidler, Thomas Esch, Noel Gorelick, Ashwin Kakarla, Emanuele Strano
TL;DR
Accurate global settlement extent remains important but existing maps suffer from sensor-specific errors, single-date effects, and update constraints. The paper constructs WSF2015, a 10m global map from open multitemporal Sentinel-1 and Landsat-8 imagery, and validates it against 900,000 crowdsourced samples. WSF2015 outperforms comparable layers and improves detection of small and scattered settlements, while very small structures remain difficult to detect.
Problem
Accurate settlement location and extent information is needed for environmental and societal analyses, while existing layers have accuracy, update, and single-sensor limitations.
Method
WSF2015 jointly uses open-and-free multitemporal Sentinel-1 radar and Landsat-8 optical imagery with temporal statistics and a statistically robust validation protocol.
Results
WSF2015 outperforms comparable settlement layers, achieving average AA% of 86.37 and average Kappa of 0.6885 across the reported experiments.
Takeaways & Limitations
WSF2015 provides detailed settlement information for applications involving human presence, including socioeconomic development, population distribution, and risk assessment.
Takeaways & Limitations
Very small structures may not be consistently detected because of their scale, materials, temporariness, or dense vegetation.
Abstract
from arXiv · showhide
Human settlements are the cause and consequence of most environmental and societal changes on Earth; however, their location and extent is still under debate. We provide here a new 10m resolution (0.32 arc sec) global map of human settlements on Earth for the year 2015, namely the World Settlement Footprint 2015 (WSF2015). The raster dataset has been generated by means of an advanced classification system which, for the first time, jointly exploits open-and-free optical and radar satellite imagery. The WSF2015 has been validated against 900,000 samples labelled by crowdsourcing photointerpretation of very high resolution Google Earth imagery and outperforms all other similar existing layers; in particular, it considerably improves the detection of very small settlements in rural regions and better outlines scattered suburban areas. The dataset can be used at any scale of observation in support to all applications requiring detailed and accurate information on human presence (e.g., socioeconomic development, population distribution, risks assessment, etc.).
BACKGROUND & SUMMARY
Global settlement maps support research on urbanization, demography, and land-use change, but existing products have important accuracy, update, and sensor-specific limitations. The paper introduces WSF2015, a global 10m settlement map generated from open, multitemporal optical and radar imagery and evaluated with 900,000 crowdsourced reference samples.
- Settlement extent maps are used to define urban areas, model population displacement, and calibrate land-use change models.
- Existing global layers include GUF at 12m for 2012, GHSL at 30m for 2014, and GLC30 at 30m for 2010.
- GUF outperforms GHSL and GLC30 but remains limited by single-date imagery, costly commercial data, and exclusive reliance on one sensor type.
- WSF2015 jointly exploits open-and-free multitemporal optical and radar imagery, using temporal statistics because settlement dynamics differ from non-settlement classes.
- WSF2015 is a global 10m binary settlement mask derived from 2014–2015 Sentinel-1 and Landsat-8 imagery and validated against 900,000 crowdsourced samples.
METHODS
The method combines independently processed Sentinel-1 radar and Landsat-8 optical imagery, temporal features, climate-aware training samples, and post-classification editing. Joint radar–optical features help distinguish settlements across contrasting environments and settlement patterns.
- Preprocessing and Feature Extraction: The workflow preprocesses Sentinel-1 and Landsat-8 data, computes temporal statistics and texture features, separately classifies radar and optical inputs, and combines the outputs.
- Preprocessing and Feature Extraction: Sentinel-1 processing includes orbit correction, thermal-noise removal, radiometric calibration, terrain correction, and conversion to decibels.
- Preprocessing and Feature Extraction: Ascending and descending radar passes are treated separately, while VH data are discarded because adding VV/VH provided no considerable improvement over VV alone.
- Preprocessing and Feature Extraction: Texture features, including the coefficient of variation of temporal-mean backscatter, support detection of low-density residential areas surrounded by vegetation.
- Preprocessing and Feature Extraction: Landsat-8 features use cloud-screened spectral indices and temporal statistics to characterize settlement dynamics and improve rural and suburban detection.
- Preprocessing and Feature Extraction: Joint radar–optical temporal features address complementary errors: optical data help in dense Lagos settlements, while radar data reduce bare-soil confusion in arid Karachi.
- Training Points Selection: Optical training thresholds are climate-specific under the Köppen–Geiger scheme, and candidate samples are selected using jointly evaluated radar, optical, and ancillary conditions.
- The WSF2015: WSF2015 maps small settlements across diverse regions and estimates a global settlement surface of ~1.28 MKm², or ~0.95% of emerged surfaces.
DATA RECORDS
WSF2015 is distributed as tiled GeoTIFF data, with multiple coarser-resolution versions reporting settlement surface coverage for broader-scale modeling.
- Distribution: 306 GeoTIFF files divide WSF2015 into 10×10-degree portions in EPSG4326 projection.Each file name specifies its upper-left and lower-right corner coordinates.
- Distribution: Five resampled versions are provided at 100m, 250m, 500m, 1km, and 10km resolutions.
- Distribution: The resampled products report the ground percentage covered by settlements in each pixel.
- Distribution: The coarser products can serve as inputs to regional, continental, or global models.They are distributed as individual GeoTIFF files with embedded overviews.
TECHNICAL VALIDATION
The WSF2015 validation combines response, sampling, and accuracy-analysis protocols with 900,000 reference samples and comparisons against GUF, GHSL, and GLC30. It reports strong performance across settlement definitions, while revealing resolution-dependent differences and improved detection of small and peripheral settlements.
- Validation framework: Validation separates response design, sampling design, and analysis to define agreement, select representative samples, and quantify map accuracy.The study explicitly presents protocols for each component before discussing validation results.
- Response design: Reference labels use Google Earth very-high-resolution imagery from 2014–2015, assessed over 3×3 blocks containing nine 10×10m cells.The imagery resolution varies by source, from approximately 0.15m airborne data to about 1.5m SPOT imagery.
- Response design: Four reference labels distinguish buildings, building lots, roads or paved surfaces, and none of these classes.Settlement is evaluated under three definitions: buildings; buildings or building lots; and buildings, building lots, or roads/paved surfaces.
- Sampling design: Stratified random sampling divides map pixels into mutually exclusive strata, while 50 representative 1×1-degree tiles are selected from approximately 14,000 tiles.Tile selection is based on the ratio between the number of settlement clusters and overall tile area.
- Quantitative results: 86.37 average accuracy and 0.6885 average Kappa make WSF2015 the best-performing layer across experiments, exceeding GUF, GHSL, and GLC30 on both measures.Mean average-accuracy increases are +6.24, +15.28, and +18.58 over GUF, GHSL, and GLC30; mean Kappa increases are +0.0754, +0.2338, and +0.2975.
- Quantitative results: 0.7646 average Kappa for criteria 3–4 versus 0.6123 for criteria 1–2 reflects compatibility between 30m Landsat-derived mapping and block-level assessment.Criteria 3–4 label a whole 30×30m reference block as settlement when at least one cell is marked settlement.
- Quantitative results: WSF2015 reaches 89.33 average accuracy and 0.7822 Kappa under Landsat-compatible criteria, outperforming GHSL and GLC30 by more than 17 and 0.32, respectively.Criteria 1–2 are considered more suitable for comparison with the 12m GUF, whose results are described as broadly in line with WSF2015.
- Qualitative results: WSF2015 detects more small villages and better outlines major urban fringes than GUF, GHSL, and GLC30 across Igboland, Kampala, and Bangalore.GHSL and GLC30 show severe underestimation at all three sites, while GUF performs comparably only in Igboland.
USAGE NOTES
WSF2015 is intended to support analyses requiring detailed, accurate information on human presence, including research and decision making. However, data limitations prevent consistent detection of some very small structures.
- WSF2015 supports applications requiring detailed and accurate information on human presence.
- Very small structures such as huts, shacks, tents, and temporary camps cannot be consistently detected.Limitations include reduced scale, building materials, temporal nature, and dense vegetation.
CODE AVAILABILITY
WSF2015 processing uses multiple architectures, software systems, and proprietary tools. Because proprietary tools are involved, the code cannot be openly released.
- WSF2015 processing involves tens of sub-modules running across multiple architectures and software systems.
- The workflow uses Google Earth Engine, DLR proprietary software, GDAL, Pktools, Calvalus, and Python scripts for different processing stages.
- The code cannot be openly released because the workflow uses proprietary tools.