DamageMap

This repository contains the code used to create and evaluate the model described in the paper DamageMap: A post-wildfire damaged buildings classifier.

Codes

In the folder "code files" one can find the following jupyter notebooks:

"create_input_images.py": This script creates square cropped training images from large GeoTiffs (xBD, Camp Fire or Carr Fire).
"Data Augmentation.ipynb": This notebook can be used on an imbalanced dataset to perform the data augmentation techniques discussed in the paper.
"Train Models.ipynb": This is the notebook used to train the models discussed in the paper. One can use it to train their own models on any dataset of their choice.
"Use Model on New Dataset.ipynb": With this notebook, one can load a model and use it to classify the images of a new dataset.
"Visualize the performance of the model.ipynb": This notebook can help someone understand better the quality (and potential problems) of the predictions of a model on a given dataset.

Models

In the folder "models and checkpoints" one can find some of the models discussed in the paper ready for use (due to space limitations some other models were not uploaded). Moreover, this folder contains some checkpoint files ("Train Models.ipynb" can illustrate how to use them), that contain useful information about the models, like the optimizer used during their training, their optimal parameters for the classification task discussed in the paper, and their validation accuracy on the Xbd validation set.

How to use our model

If the user has a single GeoTiff image covering a large land area (town, city, etc.) then they need to start by running the "create_input_images.py" script, which will cropp square images around the buildings of the area contained in the GeoTiff image. If the user already has separated images depicting individual buildings, then they can skip step 1.
The user can follow the instructions listed in "Use Model on New Dataset.ipynb" to classify the images of their dataset and they can optionally use "Visualize the performance of the model.ipynb" to evaluate the model performance.

How to train your own model

Once again, if the user has a single GeoTiff image covering a large land area (town, city, etc.) then they need to start by running the "create_input_images.py" script, which will cropp square images around the buildings of the area contained in the GeoTiff image. If the user already has separated images depicting individual buildings, then they can skip step 1.
I the dataset is imbalanced, the script "Data Augmentation.ipynb" can be used to perform data augmentation.
The user can follow the instruction listed in "Train Models.ipynb" to create their own model with their own hyper-parameter choices.
Use the model and evaluate its performance using "Use Model on New Dataset.ipynb" and "Visualize the performance of the model.ipynb" respectively.

Model Visualization

To visualize better how the model works one can visit the devpost page or the repository that we made for "Big Earth Hackathon: Wildland Fire Challenge 2020".

Devpost page link: https://devpost.com/software/damagenet-post-wildfire-building-damage-detector