Project aim

Build a continuously updated picture of ground deformation across Accra, and a lightweight model that can flag deforming areas on small edge devices.

  • 2020+Sentinel-1 archive used
  • 8Input channels
  • SBASTime-series method
  • ONNXEdge deployment
Illustrative map for Urban Subsidence Watcher Illustrative graphic

Overview

Urban Subsidence Watcher retrieves Sentinel-1 acquisitions from 2020 onward through the ASF Vertex API and processes them through an SBAS InSAR chain. A Tiny U-Net is trained on an 8-channel stack and exported to ONNX so inference can run on edge hardware rather than in the cloud.

The challenge

Urban deformation is slow, spatially patchy and easy to miss until damage appears. InSAR can measure it to millimetres, but full processing is heavy, and coherence in a tropical city varies with vegetation and construction. The project pairs the physical measurement with a small learned model that runs cheaply.

Data & inputs

  • Sentinel-1 SLC (2020 – present)
  • ASF Vertex API
  • DEM
  • Slope
  • Water proximity

Approach

  1. 01

    Data retrieval

    Sentinel-1 acquisitions from 2020 onward are searched and retrieved through the ASF Vertex API.

  2. 02

    SBAS processing

    Interferograms are formed and inverted with an SBAS chain into deformation time series.

  3. 03

    8-channel feature stack

    Coherence, amplitude, DEM, slope, temporal baseline and proximity to water are stacked as model inputs.

  4. 04

    Tiny U-Net

    A deliberately lightweight U-Net learns to segment deforming areas from the stack.

  5. 05

    Edge export

    The trained model is exported to ONNX for deployment on edge devices.

Results

The processing chain and model pipeline are in place and running on the 2020-to-present Sentinel-1 archive.

Validation

Work in progress. Model evaluation is part of the ongoing project.

Status & caveats

This is an ongoing project. Results will be published here as validation is completed.