Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Physics-informed Neural Operator for Pansharpening
Authors: Xinyang Liu, Junming Hou, Chenxu Wu, Xiaofeng Cong, zihao chen, Shangqi Deng, Junling Li, Liang-Jian Deng, Bo Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple benchmark datasets show that our method surpasses state-of-the-art fusion algorithms, achieving reduced spectral aberrations and finer spatial textures. Furthermore, extension to hyperspectral (HS) data demonstrates its generalizability and universality. The code is available at https://github.com/ez4lionky/PINO. 4 Experiments Datasets, Metrics and Implementaion Details. We assess the effectiveness of our method using data collected from the World View-3 (WV3), Gao Fen-2 (GF2), and World View-2 (WV2) satellites, which are publicly available through the Pan Collection dataset [10]. Reduced resolution evaluation is conducted using five established metrics: PSNR, SAM [80], ERGAS [68], Q2n and SCC [81]. Full resolution performance is assessed through three no-reference indicators: Dλ, and Ds, and HQNR [2]. |
| Researcher Affiliation | Academia | 1 School of Information Science and Engineering, Southeast University 2 School of Engineering Mathematics and Technology, University of Bristol 3 School of Cyber Science and Engineering, Southeast University 4 School of Mathematical Sciences, University of Electronic Science and Technology of China 5 Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University 6 Faculty of Computer Science and Control Engineering, Shenzhen University of Advanced Technology EMAIL,EMAIL EMAIL,EMAIL |
| Pseudocode | No | The paper describes the method in section 3 'Methodology' using prose and mathematical equations, without explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/ez4lionky/PINO. |
| Open Datasets | Yes | Datasets, Metrics and Implementaion Details. We assess the effectiveness of our method using data collected from the World View-3 (WV3), Gao Fen-2 (GF2), and World View-2 (WV2) satellites, which are publicly available through the Pan Collection dataset [10]. To evaluate the versatility of our model, we further conducted experiments on the multispectral and hyperspectral image fusion (MHIF) task by using the CAVE dataset. |
| Dataset Splits | Yes | Notably, due to the absence of ground-truth (GT) images, we generate reduced-resolution MS PAN training and testing pairs following Wald s protocol [69]. To evaluate the versatility of our model, we further conducted experiments on the multispectral and hyperspectral image fusion (MHIF) task by using the CAVE dataset. The CAVE dataset comprises 32 Hyperspectral Images (HSIs) with 31 spectral bands spanning from 400 nm to 700 nm at 10 nm intervals. We randomly selected 20 images for training and used the remaining 11 for testing (same as [39, 13]). |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA Ge Force GTX 4090 GPU, using the Py Torch framework. |
| Software Dependencies | Yes | The implementation of our proposed PINO uses Py Torch 2.1.0 and Python 3.10 on an Ubuntu 20.04.6 OS, with training conducted on an NVIDIA RTX 4090 GPU. |
| Experiment Setup | Yes | A two-stage training strategy is adopted, with optimization performed using the Adam algorithm (with parameters β1 = 0.9 and β2 = 0.999) [31]. In the first stage, the encoder is trained independently using a learning rate of 0.0002 to minimize a composite loss function consisting of the MAE loss and the structural similarity index loss (L1 + αs1Lssim, where αs1 = 0.1). The encoder is trained for 100, 600, and 2000 epochs on the WV3, GF2, and CAVE datasets, respectively. In the second stage, the entire framework is tuned with a learning rate halved from the first stage, optimizing a loss composed of MAE and histogram loss (L1 + αs2Lhist, where αs2 = 0.01). This stage is conducted over 200 epochs for WV3 and GF2, and 2000 epochs for CAVE. |