SF$^2$Bench: Evaluating Data-Driven Models for Compound Flood Forecasting in South Florida

This repository is the official implementation of SF$^2$Bench. A benchmark paper of compound flood in the South Florida area. In this paper, we consider seveal key factors for compound flood forecasting, including sea level, rainfall, groundwater and human control/management activities.

Dataset

Please download the dataset from https://doi.org/10.7910/DVN/TU5UXE and Unzip to dataset folder. The detail information of each file is provided in Dataset.md

Requirements

To install requirements:

pip install -r requirements.txt

Training

To train the model(s) in the paper, run main.py(load the whole dataset) or main_three_parts.py(load three interest parts). We provide the explanations of each argument. We provide the detailed hyperparameters in run.sh.

python main.py --dataset_path [PATH] --cache_dir [cache_PATH] --dataset [split]  --length_input [Lookback]  --length_output [Prediction] \
 --method [model-name] --lr [learning_rate] --weight_decay [Regularization] --epoches [train-epoch] --bachsize [B] --store [Store-model-dict] \
 --device [device] --seed [seed]

• dataset_path: the dir path of dataset
• cache_dir: the dir for save the model state dict
• length_input: the length of lookback window
• length_output: the length of prediction
• method: which architecture, “mlp,tcn,gcn,lstm” for the whole dataset, "mlp,tcn,gcn,lstm,NLinear,TSMixer,GPT4TS,AutoTimes,DeepAR,DilatedRNN,stemGNN,FourierGNN,ModernTCN,Timesnet,PatchTST,iTransformer" for the three parts dataset
• lr: learning rate
• weight_decay: regularization weight
• epoches: training epoches
• bachsize: batch size
• store: store the model state dict or not
• device: cpu or cuda
• seed: random seeds

Evaluation

To evaluate the model, the eval.py can be used. Notably, in training main.py/main_three_parts.py, the evaluation results are also available. We provide some pretrain state dict for reproduce our results.

python eval.py --dataset_path [PATH] --cache_dir [cache_PATH] --dataset [split]  --length_input [Lookback]  --length_output [Prediction] \
 --bachsize [B] --method [model-name] --device [device] --seed [seed]

Pre-trained Models

We use this repo to reporduce the results, the model state dicts are provided in google drive.

Results

The detailed results are provided in our paper. The part benchmark results on three interest parts:

Acknowledgement

Thanks for the help of the South Florida Water Management District(SFWMD) in collecting the raw dataset. Thanks for the support and help of coauthors. This repo borrowed some code from third-party repo and library. Thanks for the contributors. Here is the detailed link of these repos.

• NeuralForecast: https://github.com/Nixtla/neuralforecast
• Time-Series-Library: https://github.com/thuml/Time-Series-Library
• TCN: https://github.com/locuslab/TCN/tree/master
• ModernTCN: https://github.com/luodhhh/ModernTCN
• FourierGNN: https://github.com/aikunyi/FourierGNN
• StemGNN: https://github.com/microsoft/StemGNN
• One Fits All: https://github.com/DAMO-DI-ML/NeurIPS2023-One-Fits-All
• AutoTimes: https://github.com/thuml/AutoTimes/tree/main
• Neural-Flood-Forecasting: https://github.com/nkirschi/neural-flood-forecasting

Citation

If this repo is useful for you, please consider citing our paper:

@misc{zheng2025sf2benchevaluatingdatadrivenmodels,
      title={SF$^2$Bench: Evaluating Data-Driven Models for Compound Flood Forecasting in South Florida}, 
      author={Xu Zheng and Chaohao Lin and Sipeng Chen and Zhuomin Chen and Jimeng Shi and Wei Cheng and Jayantha Obeysekera and Jason Liu and Dongsheng Luo},
      year={2025},
      eprint={2506.04281},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.04281}, 
}