BSRLC-U - The first tri-annual 10-m maps of urban Residential, Industrial and Open spaces for the Baltic Sea region over two decades (2000 – 2021)

This post summarises my presentation at the Living Planet Symposium 2025 (Vienna, 25 June 2025).

Landsat and Sentinel-2 are the principal sources of Earth-observation data for large-scale, long-term land-cover time-series analyses.
The fine spatial resolution (10 m in the visible bands) of Sentinel-2 is particularly advantageous for distinguishing built-up classes such as residential buildings, industrial buildings, and artificial open spaces (e.g. roads, railways).
Existing 10 m urban products—e.g. the Global Human Settlement Characteristics layer—depend on Sentinel-2 imagery and therefore cover only recent years.
How can we obtain equally detailed maps for earlier periods, when only Landsat imagery was available?

Deploying deep-learning single-image super-resolution (SR) on Landsat data is not a new idea.
Most studies train SR models on temporally aligned Landsat/Sentinel-2 pairs and then apply the models to other Landsat scenes acquired under similar conditions.
In practice, applying SR models to historical Landsat imagery is complicated by issues such as the Landsat-7 SLC-off gaps, inter-scene spectral inconsistencies, and cloud contamination.

Spectral–temporal metrics (STMs)—statistics such as the minimum, maximum, median, or selected percentiles—aggregate information over time.
STM composites are spatially consistent across scenes.
In our recent study we showed that impervious surfaces exhibit resilience to temporal variability, an asset for long-term mapping.

By generating synthetic STM composites, we can harmonise Landsat and Sentinel-2 datasets in terms of both completeness and spectral characteristics.
This harmonisation enables an SR workflow that is suitable for multi-decadal mapping.

We employ a single-image SRGAN architecture, training it on STM composites instead of raw imagery.

Traditional pixel-based classification often struggles with urban scenes because different built-up types share similar spectral signatures at the pixel level.
Patch-based approaches leverage convolutional neural networks (CNNs) and spatial context, but at the cost of reduced output resolution.
Deep-learning semantic-segmentation models (e.g. U-Net) exploit spatial context while preserving resolution, but require exhaustive pixel-wise labelling—a labour-intensive task.

Have only pixel-level reference data but need CNN-level spatial context without sacrificing resolution? Centre-patch classification is the answer.
It combines the strengths of pixel-based and patch-based approaches, functioning as a “lightweight” alternative to full semantic segmentation.
Training is simple: provide a sufficient number of pixel samples, and the network learns to classify the centre pixel of each patch.
We optionally add spatial-attention layers to guide the network, although they are unnecessary when ample training data are available.

The examples below illustrate STM-based SR applied to Landsat imagery.
In many urban areas, 30 m Landsat scenes are realistically up-scaled to 10 m.
In the third panel, road surfaces—largely absent in native Landsat—are convincingly synthesised in the 10 m output.