SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening
Authors: Yu Zhong, Xiao Wu, Liang-Jian Deng, ZiHan Cao, Hong-Xia Dou
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, extensive experiments on four commonly used datasets, i.e., World View-3, World View-2, Gao Fen-2, and Quick Bird, demonstrate the superiority of SSDiff both visually and quantitatively. |
| Researcher Affiliation | Academia | Yu Zhong University of Electronic Science and Technology of China yuuzhong1011@gmail.com Xiao Wu University of Electronic Science and Technology of China wxwsx1997@gmail.com Zihan Cao University of Electronic Science and Technology of China iamzihan666@gmail.com Hong-Xia Dou Xihua University hongxiadou1991@126.com Liang-Jian Deng University of Electronic Science and Technology of China liangjian.deng@uestc.edu.cn |
| Pseudocode | Yes | Algorithm 1: Training stage of the proposed method. |
| Open Source Code | Yes | The code is available at https://github.com/Z-ypnos/SSdiff_main. |
| Open Datasets | Yes | The Pancollection1 dataset for pansharpening consists of data from four satellites: World View-3 (8 bands), World View-2 (8 bands), Quick Bird (4 bands), and Gao Fen-2 (4 bands). 1https://liangjiandeng.github.io/PanCollection.html. |
| Dataset Splits | No | The paper mentions using 'training PAN and MS images' and conducting experiments on '20 test images' for various datasets, but it does not specify explicit percentages or sample counts for training, validation, and test splits needed for reproduction, beyond referring to a data simulation protocol. |
| Hardware Specification | Yes | Our SSDiff is implemented in Py Torch 1.7.0 and Python 3.8.5 using Adam W optimizer with an initial learning rate of 0.001 to minimize Lsimple on a Linux operating system with an Intel 12th Gen i7-12700K processor and two NVIDIA Ge Force RTX3090 GPUs. |
| Software Dependencies | Yes | Our SSDiff is implemented in Py Torch 1.7.0 and Python 3.8.5... |
| Experiment Setup | Yes | Our SSDiff is implemented in Py Torch 1.7.0 and Python 3.8.5 using Adam W optimizer with an initial learning rate of 0.001 to minimize Lsimple... For the diffusion denoising model, the initial number of model channels is 32, the diffusion time step used for training in the pansharpening is set to 1000, while the diffusion time step for sampling is set to 10. The exponential moving average (EMA) ratio is set to 0.9999. The total training iterations for the WV3, GF2, and QB datasets are set to 150k, 100k, and 200k iterations, respectively. During the model fine-tuning, the learning rate is set to 0.0001, and the total fine-tune training iterations are set to 30k. |