Code and Trained Models for Benchmarking Parameter-Efficient Fine-Tuning Strategies for SAM-based SAR Flood Mapping

Abstract

This repository provides the source code, trained model checkpoints, and computational environment used in the study “Benchmarking Parameter-Efficient Fine-Tuning Strategies for Adapting Vision Foundation Models to SAR Flood Mapping”. The study evaluates parameter-efficient fine-tuning (PEFT) strategies for adapting the Segment Anything Model (SAM) to SAR-based flood mapping using the UrbanSARFloods dataset. The repository includes: training_code.zip: source code used to train the evaluated PEFT configurations. testing_code.zip: source code used for model evaluation and generation of the reported segmentation results. ckp_batch_2_seed_26.zip, ckp_batch_2_seed_42.zip, and ckp_batch_2_seed_2026.zip: trained model checkpoints obtained using a batch size of 2 and random seeds 26, 42, and 2026, respectively. ckp_batch_4_seed_26.zip, ckp_batch_4_seed_42.zip, and ckp_batch_4_seed_2026.zip: trained model checkpoints obtained using a batch size of 4 and random seeds 26, 42, and 2026, respectively. environment.yml: Conda environment specification containing the software dependencies used for the experiments. sam_vit_b_01ec64.pth: pretrained SAM ViT-B checkpoint used as the foundation model for the experiments. The six checkpoint archives correspond to the repeated experimental configurations used in the study, comprising three random seeds (26, 42, and 2026) and two batch sizes (2 and 4). These files are provided together with the training and testing code to facilitate reproduction of the experiments and evaluation reported in the associated manuscript. The SAR flood data used in this study were derived from the publicly available UrbanSARFloods dataset. Processed data used for training and independent testing are provided separately [https://doi.org/10.5281/zenodo.22305555].