A-Study-on-Urban-Pluvial-Flood-Prediction-Using-a-Transformer-Based-Model

Data, codes, and model parameters in the study of Predicting Urban Pluvial Flood Water Depth Using a Transformer-Based Deep Learning Model with Data Augmentation

This is the instruction for utilizing the codes and data from the study of Predicting Urban Pluvial Flood Water Depth Using a Transformer-Based Deep Learning Model with Data Augmentation

1. Install python dependencies by running:
   pip install -r requirements.txt

2. Download the original data used in the study by Löwe et al. (2021) from https://data.dtu.dk/articles/software/U-FLOOD_-_computer_code_and_data_associated_with_the_article_U-FLOOD_topographic_deep_learning_for_predicting_urban_pluvial_flood_water_depth_/14206838/1
   After unzip the file, please run "A_LoadData.py" to get 2 datafile, "X_data.npz" and "wsheds.npz".
   Then, copy "X_data.npz", "wsheds.npz", and "Datafiles" into the folder "data" .

3. Download the published trained model parameters of MobileViTv2 (or ResU-Net) from Releases for your test (Please don't download the outdated code in the releases). Copy the model you downloaded into the folder "models".

4. "eval.py" under "code" folder can be employed for model evaluation. There are 3 evaluation functions for different experiments:
   (1) "eval_valid": experiments in Section 3.2 and Appendix A to assess model performance on validation dataset.
   (2) "eval_test14": experiments in Section 3.3 to assess model performance on test dataset 1-4.
   (3) "eval_test47": experiments in Section 3.3 to assess model performance on test dataset 4-7.
   In the main function in "eval.py", you can modify the model class, model path, and evaluation function to reproduce the certain results in our study. The evaluation results will be presetned under "results" fofolder.

5. If you want to trained your own model, you can run "train.py" and modify the training hyper-parameters. The meaning of hyper-parameters are annotated at the beginning of "train.py" when these variables are defined.
   In the main function in "train.py", you can modify the hyper-parameters for training. The saved models during training are under "models/start time of training" folder.
   Our models were distributed trained using 4 Nvidia Tesla V100.