Leveraging Concept Relevance Propagation to Understand The Natural Areas in Satellite Imagery

Master's Thesis: CMS

TU Bergakademie Freiberg and Rheinische Friedrich-Wilhelms Universitat Bonn

Installation

Follow these steps to set up your environment to run the code from this repository.

Prerequisites

Python (recommended version 3.8 or higher)
pip (latest version)

Setup Environment

Clone the repository to your local machine:

git clone https://github.com/viswambhar-yasa/ExplainingWilderness.git
cd ExplainingWilderness

Create a virtual environment to keep dependencies required by different projects separate. Execute the following command in your terminal:
```
python -m venv venv
```
Activate the virtual environment:
```
venv\Scripts\activate
```
Install the required packages by running
```
pip install -r requirements.txt
```

Background: Untouched natural areas, such as wild and protected regions, play a crucial role as vital ecosystems, supporting various species and essential ecological processes. Their preservation is critical for maintaining biodiversity, mitigating climate change impacts, and the well-being of future generations.

Challenge: While satellite imagery and machine learning contribute to monitoring efforts, comprehending the characteristics of these areas remains challenging. In this context, explainable machine learning methods offer a promising approach to interpret these traits and uncover the underlying geo-ecological patterns.

Problem Statement: Commonly used explainable machine learning methods often fall short of providing valid explanations for the spatial and spectral patterns observed in protected regions.

Objective: This thesis aims to investigate the potential of concept relevance propagation and relevance maximization techniques in explaining the natural authenticity of protected natural areas.

Methodology: This research will involve developing and implementing the necessary models and methods to apply these techniques effectively for to identify concept by relevance maximization an perform concept disentanglement.

Global Reference: Also generate global reference image from which the channel was able to identify or learn the feature.

Evaluation: A crucial aspect of this thesis will be the evaluation of the explanations generated by these techniques and their comparison to state-of-the-art explainable machine learning methods.

Sensitivity Analysis Performing randomization of layer weights

Notebooks and python file to perform the above operations are given in experiment folder