Overview

This repository contains the codebase for BTC (Be The Change) and the accompanying analysis from our paper.

The work systematically investigates the impact of fundamental design choices in deep learning-based change detection, like pretraining strategies, data augmentations, loss functions, and learning rate schedulers. Building on these insights, we introduce BTC, a simple yet strong baseline that leverages these core components to achieve state-of-the-art performance across multiple benchmark datasets.

BTC architecture

Get Started

Environment setup

Create a Python environment and install required packages:

conda create -n btc_env python=3.10
conda activate btc_env

pip install -r requirements.txt

Training and evaluation

Run the training and evaluation with:

python train.py --config configs/exp/BTC-B.yaml

Run evaluation only:

python train.py --config configs/exp/BTC-B.yaml --eval_only --ckpt_path 

Checkpoints:

The checkpoints are available on Hugging Face and Google drive.

If you want to use the checkpoints from Huggingface, pass their model id like so: --ckpt_path blaz-r/BTC-B_<dataset_name>", e.g. --ckpt_path blaz-r/BTC-B_oscd96.

Pass the path to downloaded checkpoint if you use .ckpt files, e.g. --ckpt_path /path/to/ckpt_file.ckpt

Data

By default, datasets are automatically downloaded from Hugging Face and saved into ./datasets directory. We thank Weikang Yu for making these splits publicly available.

Possible already integrated dataset options are: oscd96, levir, sysu, egybcd, clcd, gvlm.

Note: The OSCD dataset cropped into 96x96 tiles is now available on Hugging Face. We also provide zip and HDF5 files here.

Dataset Configuration Options

• data.use_hf=True (default): Use HuggingFace datasets.

• data.use_hf=False: For custom datasets in directory-based format.

When you have use_hf set to False, you can additionally set the following flags:

• data.load_in_mem=None (default): Read images from the disk on-the-fly.

• data.load_in_mem="direct": For custom datasets in directory-based format that is read into RAM first.

• data.load_in_mem="hdf5": For HDF5-formatted datasets (including our version of OSCD) that is read into RAM first.

Custom directory format

Use the following structure for a custom directory dataset.

data_root/
    dataset_name/
       train/
            A/
                0.png
                ...
            B/
                0.png
                ...
            label/
                0.png
                ...
       test/
            A/
            B/
            label/
       val/
            A/
            B/
            label/

Custom HDF5 data

Set data.load_in_mem to "hdf5" and data.use_hf=False if you have the dataset in a format with hdf5 files:

data_root/
    dataset_name/
        train.h5
        test.h5
        val.h5

Each HDF5 file must contain the following collections:

• imageA
• imageB
• label
• img_idx

Configuration files

All experiment configurations are available in the configs directory.

The directory also includes configurations for related remote sensing foundation models here

For foundation models, please manually download weights and place them in weights/, or update paths in the config files.

Custom configuration

The config file contains the configuration for the whole setup, including data, architecture and training parameters. Refer to BTC-B config file for more details on each parameter.

You can manually override any argument in the command line without changing the config file. For example, to keep all the parameters from config but change the dataset name, simple do the following.

python train.py --config configs/BTC-B.yaml --data.dataset oscd96

The architecture follows the pipeline:

encoder -> diff -> decoder

Additional modules like in_proc, pre_diff, and out_proc are supported but unused in this work (see models/finetune_framework.py).

Alternatively, you can also list all arguments with:

python train.py -h

Performance benchmarking

The benchmarking code and results are inside the perf directory.

Run model inference speed and efficiency evaluation with:

python perf.py

Slurm

Slurm job scripts are included for both training and evaluation workflows.

Results

All evaluation metrics (F1, Precision, Recall, cIoU) across seeds and methods are stored in the results directory. This includes:

• Design choice analysis
• BTC-T / BTC-B models
• Remote sensing foundation models
• Change detection-specific baselines

Reference

If you found this work useful, consider citing our paper and giving this repo a ⭐ 😃

@ARTICLE{rolih2025btc,
  author={Rolih, Blaž and Fučka, Matic and Wolf, Filip and Zajc, Luka Čehovin},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Be the Change You Want to See: Revisiting Remote Sensing Change Detection Practices}, 
  year={2025},
  volume={63},
  number={},
  pages={1-11},
  doi={10.1109/TGRS.2025.3585342}
}

Questions

For issues or questions, please open a GitHub issue or email the author directly.

🚀 Quick Start

conda create -n btc_env python=3.10
conda activate btc_env

pip install -r requirements.txt