SpaceNet-8 Baseline

Detecting flooded roads and buildings.

Setup

1. Download this repo

2. Build docker image
   nvidia-docker build -t sn8/baseline:1.0 /path/to/sn8_baseline/docker

3. Create and run the container (mount volumes to access your data, etc. see https://docs.docker.com/engine/reference/commandline/run/ for options)
   nvidia-docker run -it --rm sn8/baseline:1.0 bash

Data Preparation

Follow these steps to prepare data for training and validation.

1. Clean the geojson labels and add speed values to the roads based on road type, number of lanes, and surface type.
   python baseline/data_prep/geojson_prep.py --root_dir /path/to/spacenet8/aws/data/download --aoi_dirs Germany_Training_Public Louisiana-East_Training_Public

   The cleaning step here also catches geometry problems, makes a single commom schema, and moves roads and buildings to seperate geojsons/shps. It will output a few additional files that are used in the subsequent step for creating image masks. The following new files will be written to the AOI annotations directory:
   - {AOI}/annotations/prepped_cleaned/roads_cleaned_{x}_{y}_{id}.geojson
   - {AOI}/annotations/prepped_cleaned/buildings_cleaned_{x}_{y}_{id}.geojson
   - {AOI}/annotations/prepped_cleaned/roads_speed_{x}_{y}_{id}.geojson
   - {AOI}/annotations/prepped_cleaned/roads_speed_{x}_{y}_{id}.shp
   - {AOI}/annotations/prepped_cleaned/buildings_{x}_{y}_{id}.shp

2. Create training and validation masks from the geojsons to use during training and validation.
   python baseline/data_prep/create_masks.py --root_dir /path/to/spacenet8/aws/data/download --aoi_dirs Germany_Training_Public Louisiana-East_Training_Public

   Four masks are generated during this process:
   - binary building mask (0 non-building, 1 building)
   - binary road mask (0 non-building, 1 building)
   - 8-channel road speed mask
   - 4-channel flood mask

3. Create a random train/val split to train the models on python baseline/data_prep/generate_train_val_test_csvs.py --root_dir /path/to/spacenet8/aws/data/download --aoi_dirs Germany_Training_Public Louisiana-East_Training_Public --out_csv_basename sn8_data --val_percent 0.15 --out_dir /path/to/output/folder/for/train/val/csvs

   This will create a csv file with filepaths to training images and labels and csv file with filepaths to validation images. It will do a random train/val split. These csvs are used by the dataloader during training.

Train/Validate Foundation Features Network

python baseline/train_foundation_features.py --train_csv /path/to/train.csv --val_csv /path/to/val.csv --save_dir /path/to/save/directory/foundation --model_name resnet34 --lr 0.0001 --batch_size 2 --n_epochs 50 --gpu 0

Inference with Foundation Features Network

1. Write prediction tiffs to be used for postprocessing and generating the submission .csv
   python baseline/foundation_eval.py --model_path /path/to/saved/foundation/best_model.pth --in_csv /path/to/val/or/test.csv --save_preds_dir /path/to/output/foundation/eval_test --gpu 0 --model_name resnet34

2. Write prediction .pngs for visual inspection of predictions
   python baseline/foundation_eval.py --model_path /path/to/saved/foundation/best_model.pth --in_csv /path/to/val/or/test.csv --save_fig_dir /path/to/output/foundation/eval_test/pngs --gpu 0 --model_name resnet34

Train/Validate Flood Features Network

python baseline/train_flood.py --train_csv /path/to/train.csv --val_csv /path/to/val.csv --save_dir /path/to/save/directory/flood --model_name resnet34_siamese --lr 0.0001 --batch_size 2 --n_epochs 50 --gpu 0

Inference With Flood Features Network

1. Write prediction tiffs to be used for postprocessing and generating the submission .csv
   python baseline/flood_eval.py --model_path /path/to/saved/flood/best_model.pth --in_csv /path/to/val/or/test.csv --save_preds_dir /path/to/output/flood/eval_test --gpu 0 --model_name resnet34_siamese

2. Write prediction .pngs for visual inspection of predictions
   python lib/flood_eval.py --model_path /path/to/saved/flood/best_model.pth --in_csv /path/to/val/or/test.csv --save_fig_dir /path/to/output/flood/eval_test/pngs --gpu 0 --model_name resnet34_siamese

Postprocessing

Any of the following postprocessing steps require that you have run inference with the flood features network and foundation features network. As input, the postprocessing requires the geotiff predictions.

Roads

1. Skeletonize road raster predictions and convert to vector data.
   python baseline/postprocessing/roads/vectorize_roads.py --im_dir /path/to/road/prediction/geotiffs --out_dir /path/to/road/prediction/geotiffs --write_shps --write_graphs --write_csvs --write_skeletons
2. Generate road network graph from vector data
   python baseline/postprocessing/roads/wkt_to_G.py --wkt_submission /path/to/output/csv/from/last/step --graph_dir /output/path/to/save/simplified/graphs --log_file /path/to/log
3. Assign road speed predictions to road network graph linestrings
   python baseline/postprocessing/roads/infer_speed.py
4. Generate the road submission .csv
   python baseline/postprocessing/roads/create_submission.py

To run all these steps, see the script /baseline/postprocessing/roads/road_post.sh. Change only the variables:
EVAL_CSV
ROAD_PRED_DIR
FLOOD_PRED_DIR

Buildings

1. Merge the building predictions from the foundation features network with the flooded predictions from the flood features network to attribute buildings as flooded or non-flooded. Polygonize the building prediction raster, remove polygons below a certain area threshold and simplify polygon geometries.
   python baseline/postprocessing/buildings/building_postprocessing.py --root_dir /path/to/foundation/features/building/prediction/geotiffs --flood_dir /path/to/flood/prediction/geotiffs --out_submission_csv /path/to/output/building_submission.csv --out_shapefile_dir /path/to/output/building/pred_shps --square_size 5 --simplify_tolerance 0.75 --min_area 5 --perc_positive 0.5

Submission

Your submission files should be in the following format.

ImageId	Object	Flooded	Wkt_Pix	Wkt_Geo

ImageId is the reference image for the prediction. Object should be either "Building" or "Road". Flooded should be set to "True" or "False". Wkt_Pix is well-known-text format for coordinates as (row, col) of the predictions in reference to the image. Wkt_Geo is well-known-text coordinates of the predictions in WGS84 (x, y).

Example for buildings:

ImageId	Object	Flooded	Wkt_Pix	Wkt_Geo
104001006504F400_0_11_20	Building	Null	POLYGON EMPTY	POLYGON EMPTY
104001006504F400_0_10_29	Building	FALSE	POLYGON ((6 706, 0 708, 0 715, 9 719, 12 706, 6 706))	POLYGON ((-90.4706194788189 29.8296995794957,-90.4706358281571 29.8296930397604,-90.4706390980247 29.8296668808194,-90.4706063993484 29.8296570712164,-90.4705965897455 29.8296963096281,-90.4706194788189 29.8296995794957))

References and Resources

1. SpaceNet-8 workshop paper: https://openaccess.thecvf.com/content/CVPR2022W/EarthVision/papers/Hansch_SpaceNet_8_-_The_Detection_of_Flooded_Roads_and_Buildings_CVPRW_2022_paper.pdf

2. The preprocessing for road speed labels, road speed training, and road speed post-processing leverages: https://github.com/avanetten/cresi

3. See the following paper for more information on the road speed segmentation: https://openaccess.thecvf.com/content_WACV_2020/html/Van_Etten_City-Scale_Road_Extraction_from_Satellite_Imagery_v2_Road_Speeds_and_WACV_2020_paper.html

4. SpaceNet 5 blogs for road speed estimation: https://spacenet.ai/sn5-challenge/