Integrative super-resolution of Sentinel-2 satellite data for 2D and 3D urban sprawl analysis

Finished
project graphic Shirley Sidharta, CC BY SA 4.0

In the project, geometric and radiometric 2D and 3D changes of urban structures are analyzed. Sentinel-2 satellite data serve as a basis. The aim is to gain information about urban sprawl.

Motivation

The project is divided into the following steps:

  1. development of a Deep Learning (DL)-based super-resolution for optical satellite imagery (e.g. Sentinel-2 data).
  2. development of a DL-based algorithm for direct height extraction (3D) from individual optical satellite images (2D) and generation of Digital Surface Model (DSM) data
  3. integration of super-resolution and height extraction networks to directly obtain high-resolution 2D and 3D information
  4. development of an algorithm to measure and display the statistical data of the degree of change (e.g. area, volume, height, etc.) 5.Propose directional map of changes in relation to urban residential areas.

Activities

A preliminary study on the extraction of an elevation model based on a single (monocular) Sentinel-2 satellite image (task 2 above) is in progress. It is based on a deep learning algorithm (Unet++ architecture) developed by Johannes Hassler in the context of his master thesis Image to Height - Estimation of Population Density in Central European Cities from Satellite Images using Machine Learning Methods. The diagram (Fig. 1) shows the structure and an example result for monocular depth (height) prediction.

The method was further analyzed to determine the height of each building to estimate the population size as a function of the area and volume of the buildings . Three scenarios were tested to find out the correlations between the RGB sentinel image, the estimated elevation model and the combined RGB and elevation model with the population values (Fig. 2).