motokimura / spacenet_building_detection

Project to train/test convolutional neural networks to extract buildings from SpaceNet satellite imageries.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

SpaceNet Building Detection

This repository privides some python scripts and jupyter notebooks to train and evaluate convolutional neural networks which extract buildings from SpaceNet satellite images.

Dependency

Usage

0. Clone this repo

$ PROJ_DIR=~/spacenet_building_detection  # assuming you clone this repo to your home directory

$ git clone --recursive https://github.com/motokimura/spacenet_building_detection.git $PROJ_DIR

1. Download SpaceNet dataset

Download satellite images and ground-truth building footprints (of SpaceNet AOI-1 Rio). Assuming you already configured AWS CLI to download the data from AWS S3.

$ cd $PROJ_DIR/data

# Download processed 200mx200m tiles of AOI 1 (3.4 GB) with associated building footprints
$ aws s3api get-object --bucket spacenet-dataset --key AOI_1_Rio/processedData/processedBuildingLabels.tar.gz --request-payer requester processedBuildingLabels.tar.gz

$ tar -xvf processedBuildingLabels.tar.gz
$ cd processedBuildingLabels
$ tar -xvf 3band.tar.gz

$ cd vectordata 
$ tar -xvf geojson.tar.gz

# Download the Source Imagery Mosaic (3-band = 2.3 GB and 8-band = 6.5 GB)
$ cd $PROJ_DIR/data
$ aws s3api get-object --bucket spacenet-dataset --key AOI_1_Rio/srcData/rasterData/3-Band.tar.gz --request-payer requester 3-Band.tar.gz
$ tar -xvf 3-Band.tar.gz

You can find more details here.

2. Build Docker image

Build docker image to setup the environment to preprocess SpaceNet dataset and train/evaluate convolutional neural networks.

$ cd $PROJ_DIR/docker
$ bash build.sh

I strongly reccomend you to use docker because you may have some troubles to install some geo-spatial data processing libraries. In case you don't want to use docker, you have to install additional dependencies described in Dockerfile.

3. Preprocess SpaceNet dataset

Convert SpaceNet ground-truth building footprints from GeoJSON into building mask images.

Run docker container by following:

$ cd $PROJ_DIR/docker
$ bash run.sh

Now you should be inside the docker container you ran. Convert SpaceNet dataset by following:

$(docker) cd /workspace/src/features
$(docker) python build_labels.py ../../data/processedBuildingLabels/3band ../../data/processedBuildingLabels/vectordata/geojson ../../data/buildingMaskImages

Now you will find many pairs of satellite images and building mask images in $PROJ_DIR/data/processedBuildingLabels/3band and $PROJ_DIR/data/buildingMaskImages respectively like below:

4. Train U-Net

Train U-Net, a convolutional neural network originaly developed for medical image segmentation.

Train U-Net with SpaceNet dataset by following:

$(docker) cd /workspace/src/models
$(docker) python train_model.py ../../data/dataSplit ../../data/processedBuildingLabels/3band ../../data/buildingMaskImages

You can check training status and validation accuracy from TensorBoard:

# Open another terminal window outside the container and type:
$ cd $PROJ_DIR/docker
$ bash exec.sh

# Now you should be inside the container already running. Start TensorBoard by following:
$(docker) tensorboard --logdir /workspace/models

Then, open http://localhost:6006 from your browser.

5. Evaluate U-Net

Evaluate U-Net with jupyter notebook.

Launch jupyter notebook by flollowing:

$(docker) cd /workspace/notebooks
$(docker) jupyter notebook

Then, open http://localhost:8888 from your browser.

Default password is passw0rd as defined in Dockerfile.

Note that you may need to modify the path to pre-trained model defined in the notebooks below.

5.1 Quantitative evaluation

Open this notebook.

5.2 Qualitative evaluation on test tile-images

Open this notebook to see segmentation results on tile images in test-plit.

Output examples:

5.3 Qualitative evaluation on mosaic images

Open this notebook to see segmentation results on source mosaic images.

Output examples:

License

MIT License

About

Project to train/test convolutional neural networks to extract buildings from SpaceNet satellite imageries.

License:MIT License


Languages

Language:Jupyter Notebook 99.6%Language:Python 0.3%Language:Dockerfile 0.0%Language:Shell 0.0%