Caladrius - Assessing Building Damage caused by Natural Disasters using Satellite Images
Created by: Artificial Incompetence for the Red Cross #1 Challenge in the 2018 Hackathon for Peace, Justice and Security
Note: Parts of this project are not made public for privacy and operational reasons. If you would like to access any restricted content, please send an email to grahman@rodekruis.nl with the relevant details (content url, purpose of request, et cetera).
The network architecture is a pseudo-siamese network with two ImageNet pre-trained Inception_v3 models.
Install Docker.
Download the Caladrius Docker Image using,
docker pull gulfaraz/caladrius
Create a data folder in your local machine.
Create a docker container using,
docker run --name caladrius -dit -v <path/to/data>:/workspace/data -p 5000:5000 gulfaraz/caladrius
Access the container using,
docker exec -it caladrius bash
- Python 3.6.5
- Anaconda or Miniconda 2019.07
- NodeJS v10
- Run the following script,
./caladrius_install.shThe Sint Maarten 2017 dataset can be downloaded from here.
To extract the contents to the data folder execute,
tar -xvzf rc.tgz
Transform the raw dataset to a training dataset using,
python caladrius/dataset/sint_maarten_2017.py --version 1.0.0 --create-image-stamps --query-address-api --address-api openmapquest --address-api-key <ADDRESS_API_KEY> --create-report-info-file
The above command will create the dataset as per the specifications.
sint_maarten_2017.py accepts the command line arguments described below,
usage: sint_maarten_2017.py [-h] --version VERSION [--create-image-stamps]
[--query-address-api] [--address-api ADDRESS_API]
[--address-api-key ADDRESS_API_KEY]
[--create-report-info-file]
[--label-type label_type]
optional arguments:
-h, --help show this help message and exit
--version VERSION set a version number to identify dataset (default:
None)
--create-image-stamps
For each building shape, creates a before and after
image stamp for the learning model, and places them in
the approriate directory (train, validation, or test)
(default: False)
--query-address-api For each building centroid, preforms a reverse geocode
query and stores the address in a cache file (default:
False)
--address-api ADDRESS_API
Which API to use for the address query (default:
openmapquest)
--address-api-key ADDRESS_API_KEY
Some APIs (like OpenMapQuest) require an API key
(default: None)
--create-report-info-file
Creates a geojson file that contains the locations and
shapes of the buildings, their respective
administrative regions and addresses (if --query-
address-api has been run) (default: False)
--label-type label_type
Sets whether the damage label should be produced on a
continuous scale or in classes. (default: regression)
From the caladrius/interface directory execute,
npm start
The interface should be accessible at http://localhost:5000.
python caladrius/run.py --run-name caladrius_2019
python caladrius/run.py --run-name caladrius_2019 --test
Click here to download the trained model.
run.py accepts the command line arguments described below,
usage: run.py [-h] [--checkpoint-path CHECKPOINT_PATH] [--data-path DATA_PATH]
[--run-name RUN_NAME] [--log-step LOG_STEP]
[--number-of-workers NUMBER_OF_WORKERS]
[--model-type {quasi-siamese,random,average}] [--disable-cuda]
[--cuda-device CUDA_DEVICE] [--torch-seed TORCH_SEED]
[--input-size INPUT_SIZE] [--number-of-epochs NUMBER_OF_EPOCHS]
[--batch-size BATCH_SIZE] [--learning-rate LEARNING_RATE]
[--test] [--max-data-points MAX_DATA_POINTS]
[--train-accuracy-threshold TRAIN_ACCURACY_THRESHOLD]
[--test-accuracy-threshold TEST_ACCURACY_THRESHOLD]
[--output-type {regression,classification}]
optional arguments:
-h, --help show this help message and exit
--checkpoint-path CHECKPOINT_PATH
output path (default: ./runs)
--data-path DATA_PATH
data path (default: ./data/Sint-Maarten-2017)
--run-name RUN_NAME name to identify execution (default: <timestamp>)
--log-step LOG_STEP batch step size for logging information (default: 100)
--number-of-workers NUMBER_OF_WORKERS
number of threads used by data loader (default: 8)
--model-type {quasi-siamese,random,average}
type of model (default: quasi-siamese)
--disable-cuda disable the use of CUDA (default: False)
--cuda-device CUDA_DEVICE
specify which GPU to use (default: 0)
--torch-seed TORCH_SEED
set a torch seed (default: 42)
--input-size INPUT_SIZE
extent of input layer in the network (default: 32)
--number-of-epochs NUMBER_OF_EPOCHS
number of epochs for training (default: 100)
--batch-size BATCH_SIZE
batch size for training (default: 32)
--learning-rate LEARNING_RATE
learning rate for training (default: 0.001)
--test test the model on the test set instead of training
(default: False)
--max-data-points MAX_DATA_POINTS
limit the total number of data points used, for
debugging on GPU-less laptops (default: None)
--train-accuracy-threshold TRAIN_ACCURACY_THRESHOLD
window size to calculate regression accuracy (default:
0.1)
--test-accuracy-threshold TEST_ACCURACY_THRESHOLD
window size to calculate regression accuracy (default:
0.3)
--output-type {regression,classification}
choose if want regression or classification model
(default: regression)
After cloning the repo, run pre-commit install to enable format checking when committing changes.
When making changes, increment version number in VERSION, package.json, the badge in README.md and package.json according to PEP 440 and update CHANGES.md.
To build and tag the Docker image with VERSION,
make build_productionFor development and tagging with the latest commit version,
make build_fast