Standardized Comparison of Urban Green Space Mapping Through Remote Sensing for Frankfurt, Germany

Challenge Background

Monitoring urban vegetation is a costly and time-consuming task — hence there is an increased interest in automated mapping techniques. The use of remote sensing imagery to distinguish different land cover and land use types in an urban environment is a mature sub-discipline of remote sensing research.

However, the amount of literature dedicated to green space segmentation and classification remains notably smaller. Still, standardized comparisons of different mapping techniques are rare. For example, for green spaces that lie below the resolution limit of the concerned sensor - which is for example 10m x 10m for freely accessible sentinel-2 imagery - need to be detected on sub-pixel-level. 

Here, it can be aimed to calculate the percentage of pixel to which green spaces contribute, based on either single bands or indices. So called “unmixing” of pixels can then be performed through, e.g. with multiple endmember spectra mixture analysis, based on, e. g., linear or Gaussian models. 

For these techniques open questions remain regarding the ultimate resolution limit. Likewise, it has to be investigated whether alternative techniques, such as e.g. input amplification, can push resolution limits further. 

Also, a plethora of further information exists in remote sensing such as e. g. seasonal change analysis, or temperature mapping that provide further information about the presence/absence of urban greenery. Albeit computationally more costly, deep learning strategies such as U-nets or the recently proposed selective state space models (SSM) can enhance predictive power.

The Problem

Albeit different methods exist for satellite-based urban green space mapping - many of which were introduced within the last 3-4 years - systematic comparisons of resolution limits between different methods (especially different categories of methods) are rare.

This project aims to compare different unmixing techniques to deep learning architectures (in particular a U-net and SSM) in combination with or without change detection as well as potentially input amplification in their ability to detect and map small patches of urban greenery.

The project is focused on the geographical area of Frankfurt am Main, Germany. For this city, high-resolution, summer air photographs in the RGB-spectra as well as near infrared spectra are available.

Goal of the Project

1. Train project members in python-based geospatial methodology and basic knowledge in remote sensing
2. Use publicly available air photographs and maps of Frankfurt/ Main, Germany, to create fine-grained ground truth data segmented for urban green space mapping
3. Compare methodologically highly distinct algorithmic methods for ability to segment even small urban green spaces
4. Evaluate whether two deep learning techniques, U-net and mamba, can enhance detection of small urban green spaces further and publish results

Project Timeline

What you'll learn

1. Geospatial data handling in Python
2. Image segmentation
3. Image scaling and georeferencing
4. Building Machine Learning Models
5. Spatial error analysis
6. Building deep learning models
7. Scientific communication