Post-event building damage assessment

This repository is part of a lecture series at the University of Bonn and is designed as an introduction to deep learning methods in the remote sensing domain. It contains code and resources for the development of an AI model that performs rapid building damage assessment. Deep learning techniques are utilized to analyze satellite or aerial imagery and predict the level of damage to buildings after natural disasters or other events. After training, the model can be used to analyze large-scale disaster areas and help prioritize rescue and recovery efforts. Additional materials on the workflow and further background information can be found under res/background.pdf.

Table of Contents

• Installation
• Dataset
• Workflow

Installation

Software

The following software packages or equivalents have to be installed.

• Integrated development environment (IDE)
  An IDE is a software application that provides comprehensive tools and features to aid programmers in developing, testing, and debugging software. While all IDEs for Python can be used, PyCharm offers a user-friendly, powerful, and free option.

• conda package manager
  The conda package manager is an open-source package management system primarily used for managing software environments. Conda simplifies package management and environment creation, avoiding compatibility issues and ensuring that all the necessarey dependencies are properly installed. It is widely used in the data science and machine learning communities due to its robustness and extensive package ecosystem. To install the conda package manager, either Miniforge or Anaconda can be downloaded.

• Git
  The distributed version control system (VCS) Git is widely used for tracking changes in source code during software development. It allows multiple developers to collaborate on a project, managing and merging their code efficiently.

Repository

After successful installation of the software above, follow the instructions to download and set up the repository.

1. Clone the repository using the git CMD application:

cd path-to-target-directory
git clone https://github.com/vhertel/bda-session-bonn.git

2. Install the repository environment using miniforge prompt or Anaconda application:

cd path-to-repository
conda env create -f environment.yml

3. Open the project and select the conda environment bda as interpreter. In PyCharm, this can be done via File -> Settings... -> Project: bda-session-bonn -> Project Interpreter -> Add Interpreter -> Add Local Interpreter... -> Conda Environment . Click on the ... button next to the Interpreter field to open the Select Python Interpreter dialog. Select the python.exe of the desired environment. PyCharm will now use the selected Conda environment as the interpreter for your project.

Dataset

The model is trained on the xBD dataset, which is a valuable resource consisting of pairs of pre-disaster and post-disaster satellite imagery with corresponding damage labels on building level. By utilizing the xBD dataset, researchers and developers can advance the field of building damage assessment by leveraging machine learning techniques to automatically analyze satellite imagery and provide accurate and timely information about the extent of damage after natural disasters. The xBD dataset can be downloaded here and must be prepared in tiles of 256x256 pixels.

Workflow

This repository hosts a U-Net model that has been pre-trained on the xBD dataset. The model underwent 25 epochs of training, excluding volcanic events. The main purpose of this session is to train a model for assessing building damage.

For doing so, follow the instructions:

1. Open train.py, adjust the CONFIG parameters and run the script. During the training process, model weights are stored under res/models/ in the format epoch-f1_macro-loss.pth.tar. Note: if memory errors occur, reduce the batch size.
2. Open inference.py, adjust the CONFIG parameters and run the script. Make sure to select the model with highest macro F1 score of the training in step 3. Note and compare the macro and micro F1 scores after the process.

During the training process, the evaluation of the loss as well as macro and micro F1 scores are stored under res/logs/. Example tiles with the size of 256x256 pixels are plotted and stored under res/output/ during inference.