Environmental Agency2024 Government

National Land-Cover Mapping Pipeline

An automated pipeline producing nationwide land-cover maps from Sentinel-2 imagery, refreshed each season.

Python
Rasterio
PyTorch
PostGIS
Airflow

Challenge: The agency relied on a manual, multi-week process to produce land-cover maps, making seasonal updates impractical and error-prone.
Solution: We built an orchestrated pipeline that ingests Sentinel-2 scenes, applies cloud masking and a segmentation model, and publishes analysis-ready COGs and vector summaries to PostGIS.
Outcome: Map production dropped from weeks to hours, enabling quarterly national updates with consistent, documented quality metrics.

Background

The agency needed timely, consistent land-cover information across the entire country, but their existing workflow combined manual scene selection, desktop classification, and ad-hoc QA. Each update cycle took weeks and was difficult to reproduce.

What we built

We designed an end-to-end pipeline orchestrated with Airflow:

Automated discovery and download of Sentinel-2 imagery for each tile and season.
Cloud and shadow masking, atmospheric normalization, and mosaicking into analysis-ready stacks.
A PyTorch segmentation model producing per-pixel land-cover classes.
Vectorization and aggregation into PostGIS, with Cloud-Optimized GeoTIFFs for raster delivery.

Results

The agency now refreshes national land-cover maps every quarter. Each run produces documented accuracy metrics, and the whole process is reproducible from source imagery to published product.

Working on something similar?

Tell us about it