ada-collection

Tools for Automated Mapping and Damage Assessment.

This repo wraps two packages:

• abd_model - forked from automated-building-detection
• ada_tools - tools for pre- and post-processing of remote sensing images and vector data

The damage assessment framework & model is at caladrius:ada-0.1 automatically installed via Docker (caladrius.Dockerfile).

Getting started

Get pre-trained models and images

1. Download pre-trained building detection model: neat-fullxview-epoch75.pth:

• architecture: AlbuNet (U-Net-like encoder-decoder with a ResNet, ResNext or WideResNet encoder)
• training: xBD dataset, 75 epochs
• performance: IoU 0.79, MCC 0.75

2. Download pre-trained building damage classification model: caladrius_att_effnet4_v1.pkl

• architecture: pseudo-siamese network with two ImageNet pre-trained EffNet-B4 models + attention
• training: xBD dataset, 75 epochs
• performance: up to F1 score 0.79, AUC 0.984 (see performance paper)

3. [OPTIONAL] pre- and post-disaster satellite images

Your workspace should then look like

    <workspace>
    ├── ...
    ├── neat-fullxview-epoch75.pth    # pre-trained building detection model
    ├── caladrius_att_effnet4_v1.pkl  # pre-trained building damage classification model
    ├── images                        # satellite images
    │   ├── pre-event                 # before the disaster
    │   └── post-event                # after the disaster
    └── ...

Using Docker

1. Install Docker.
2. Download the latest Docker Image

docker pull jmargutti/ada-collection

3. Create a docker container and connect it to a local directory (<path-to-your-workspace>)

docker run --name ada-collection -dit -v <path-to-your-workspace>:/workdir --ipc=host --gpus all -p 5000:5000 rodekruis/ada-collection

4. Access the container

docker exec -it ada-collection bash

Manual Setup

1. Install Python 3.7 and pip.
2. Install Anaconda.
3. Create and activate a new Anaconda environment.

conda create --name abdenv python=3.7 
conda activate abdenv

4. From root directory, move to ada_tools and install

cd ada_tools
pip install .

Note: Make sure libboost/boost is installed.

5. Move to abd_model and install

cd ../abd_model
pip install .

6. Get caladrius:ada-0.1 (damage assessment framework) and install

git clone --branch ada-0.1 https://github.com/rodekruis/caladrius.git
cd caladrius
./caladrius_install.sh

End-to-end example

1. Get satellite images of typhoon Mangkhut from Maxar Open Data

load-images --disaster typhoon-mangkhut --dest <workspace>/images

• Alternatively, load images from Azure blob storage

load-images-azure --disaster typhoon-mangkhut --dest <workspace>/images

• Set the CONNECTION_STRING and CONTAINER_NAME environmental variables corresponding to your Azure account

2. Prepare images for building detection

abd cover --raster <workspace>/images/pre-event/*.tif --zoom 17 --out <workspace>/abd/cover.csv
abd tile --raster <workspace>/images/pre-event/*.tif --zoom 17 --cover <workspace>/abd/cover.csv --out <workspace>/abd/images --format tif --no_web_ui --config ada-tools/config.toml

3. Detect buildings

abd predict --dataset <workspace>/abd --cover <workspace>/abd/cover.csv --checkpoint <workspace>/neat-fullxview-epoch75.pth --out <workspace>/abd/predictions --metatiles --keep_borders --config ada-tools/config.toml

4. Generate vector file with buildings and filter noise

abd vectorize --masks <workspace>/abd/predictions --type Building --out <workspace>/abd/buildings.geojson --config ada-tools/config.toml
filter-buildings --data <workspace>/abd/buildings.geojson --dest <workspace>/abd/buildings-clean.geojson

5. Prepare images for building damage classification

prepare-data --data <workspace>/images --buildings <workspace>/abd/buildings-clean.geojson --dest <workspace>/caladrius

6. Classify building damage

CUDA_VISIBLE_DEVICES="0" python caladrius/caladrius/run.py --run-name run --data-path <workspace>/caladrius --model-type attentive --model-path <workspace>/caladrius_att_effnet4_v1.pkl --checkpoint-path <workspace>/caladrius/runs --batch-size 2 --classification-loss-type f1 --output-type classification --inference

7. Generate vector file with buildings and damage labels

final-layer --builds <workspace>/abd/buildings-clean.geojson --damage <workspace>/caladrius/runs/run-input_size_32-learning_rate_0.001-batch_size_32/predictions/run-split_inference-epoch_001-model_inception-predictions.txt --out <workspace>/buildings-predictions.geojson --thresh 1

8. Take your favorite GIS application and visualize <workspace>/buildings-predictions.geojson in a nice map