The hydrological streamline detection using an attention U-net model

Data description: Training, validation, and testing dataset (LiDAR derived feature maps), scripts, pretrained models, and prediction results for the hydrological streamline detection of Rowan Creek, NC, watershed area. The dataset is generated and provided by USGS (author: Larry Stanislawski ([email protected]) and Ethan Shavers ([email protected])). Contents:

• data.zip:

  • mask.npy: The mask of research area. 1 indicates research area. 0 indicates outside area.
  • reference.npy: numpy array format of reference data. 1 indicates streamlines, 0 indicates non-streamlines
  • prediction_data.npy: The prediction data of research area
  • train*.npy: training data with data augmentation.
  • vali*.npy: validation data for model tuning.
  • trainlabel.npy: the labels for training data.
  • valilabel.npy: the labels for validation data.

• script:

  • train_U_net.py: training script of the U_net model for streamline detection.
  • train_attention_U_net.py: training script of the attention U_net model for streamline detection.
  • accuracy_evaluation.py: accuracy evaluation for the prediction results.
  • generate_resultant_tifs.py: convert result to geo-referenced .tiffs.

• result (files to be saved in):

  • *.pickle: stores the intermediate training statistics and printouts.
  • preds_test_*.npy: predicted results orginzed as patches.
  • preds_test_*.tif: Organized.tiffs of predicted results.

• models (files to be saved in):

  • *.h5: stores the trained model.

Environment: Python3.7 Library dependencies: cuda10.0, pdal 1.2, sklearn 0.18.1, keras 2.0, tensorflow 1.0, osgeo(gdal) 3.0.4.

Steps:

1. Download the data from https://doi.org/10.6084/m9.figshare.12584975.v1, unzip, and put it in the same directory as models, scripts, and result folders.
2. Run train_U_net.py or train_attention_U_net.py for model training by specifying the input directory and training scenario ('up','down','left', or 'right').
3. Run generate_resultant_tifs.py to transfer predicted results to geo-referenced tiffs.
4. Run accuracy_evaluation.py to generate the evaluation statistics of the four scenarios' prediction maps using the U-net (0) or the attention U-net model(1).

Note: 1. If you don't want to train from scratch, you can simply download the pretrained model from: https://doi.org/10.6084/m9.figshare.12587318.v1, unzip it in the models folder, and load the pretrained models to generate your predictions. 2. You can also directly download, unzip, and visualize the generated prediction maps from: https://doi.org/10.6084/m9.figshare.12587324.v1.

Publication: