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).
Important
This readme is intended for the general public.
For 510-specific instructions during emergencies, see the ADA SOP.
- 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
- 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)
- [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
└── ...
- Install Docker.
- Download the latest Docker Image
docker pull jmargutti/ada-collection
- 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
- Access the container
docker exec -it ada-collection bash
conda create --name abdenv python=3.7
conda activate abdenv
- From root directory, move to
ada_tools
and install
cd ada_tools
pip install .
Note: Make sure libboost/boost is installed.
- Move to
abd_model
and install
cd ../abd_model
pip install .
- 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
- 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
andCONTAINER_NAME
environmental variables corresponding to your Azure account
- 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
- 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
- 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
- Prepare images for building damage classification
prepare-data --data <workspace>/images --buildings <workspace>/abd/buildings-clean.geojson --dest <workspace>/caladrius
- 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
- 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
- Take your favorite GIS application and visualize
<workspace>/buildings-predictions.geojson
in a nice map