CODE-PIF: Convex/Deep (CODE) Multiview Pseudoinvariant Features for Multitemporal Multispectral Surface Reflectance Imagery
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "CODE-PIF: Convex/Deep (CODE) Multiview Pseudoinvariant Features for Multitemporal Multispectral Surface Reflectance Imagery".
Dev: Here's the gist of the paper. Multitemporal multispectral surface reflectance imagery exhibits radiometric and spectral variability due to atmospheric effects, sensor differences, and actual surface changes, complicating the identification of pseudoinvariant features (PIFs) The proposed CODE-PIF architecture is an unsupervised convex/deep (CODE) framework that refines the coarse PIF from DeepMGCCA using convex optimization Pruning redundant reasoning tokens The paper introduces CODE-PIF, a framework that refines the initial pseudoinvariant feature (PIF) selection derived from DeepMGCCA by integrating the deep mask with invariance scores calculated from spectral and temporal information through convex optimization This refinement is motivated by the fact that agreement in the learned latent space does not necessarily imply radiometric and spectral consistency in original reflectance observations The final PIF selection is guided by a monotone piecewise-linear transformation to harmonize multitemporal reflectance observations, which selects calibration references while adjusting reflectance values without reversing their ordering How the training signal is built The initial PIF mask, denoted as DeepMGCCA, is obtained by ranking pixels based on normalized alteration scores for each acquisition date and selecting the lowest-scoring fraction to construct the initial PIF mask The deep change prior is then defined as 1 minus this initial PIF mask, assigning zero to pixels selected as PIFs by DeepMGCCA and one to all remaining pixels The data-fitting term in the convex optimization problem relates these continuous change scores to spectral and temporal information derived from observations This term incorporates spectral angle measurements and reflectance residuals weighted by a temporal reference to promote radiometric and spectral consistency Where the gains come from The gains of CODE-PIF are demonstrated across four geographically distinct regions—Turkey, Taipei, Tainan, and Saudi Arabia—showing improved average radiometric and spectral agreement relative to DeepMGCCA Averaged across these regions, the RMSE decreases from 0.0192 to 0.0164, PSNR increases from 35.91 to 37.44 dB, and the between-date spectral angle decreases from 3.77° to 2.97° The qualitative comparison shows that CODE-PIF excludes open-water regions initially selected by DeepMGCCA, resulting in spatially dispersed selections over built-up and bare surfaces rather than large contiguous regions Quantitative Comparison of PIF Selection The quantitative results show that the proposed CODE-PIF achieves lower RMSE and between-date spectral angle (SAMh), together with higher PSNR, than DeepMGCCA in all four regions Averaged across the ROIs, RMSE decreases from 0.0192 to 0.0164, corresponding to an approximately 14.5% reduction, while SAMh decreases from 3.77◦ to 2.97◦ The regional improvements are consistent with the scatterplots, with the most pronounced gains observed in Turkey and Taipei Discussion The reliability of CODE-PIF depends on the quality of both the initial DeepMGCCA selection and reflectancederived change scores The pixelwise data-fitting map does not explicitly model local spatial relationships and may overlook variations in reflectance magnitude that preserve spectral shape The experimental protocol defines the scope of the reported improvements, as results quantify the contribution of PIF refinement within this calibration pipeline and do not independently establish the physical invariance of every selected pixel In conclusion, CODE-PIF is an unsupervised CODE framework that refines DeepMGCCA-derived PIFs through mathematically simple convex optimization for multitemporal surface reflectance harmonization The extensive experiments across four geographically distinct regions demonstrate improved average radiometric and spectral agreement compared with DeepMGCCA, thereby achieving state-of-the-art performance Future work will focus on developing an interpretable architecture for multitemporal surface reflectance harmonization, facilitating explicit analysis of how model updates affect radiometric and spectral consistency Motivated by recent advances in deep unfolding for remote sensing, the authors plan to investigate a convex formulation for multitemporal surface reflectance harmonization and unfold the associated iterative solver into a trainable network The paper concludes that PIF selection should be assessed jointly in terms of reflectance agreement, spectral stability, spatial structure, and linear correlation Acknowledgements This study was supported by the Emerging Young Scholar Program (the 2030 Cross-Generation Young Scholars Program) of National Science and Technology Council (NSTC), Taiwan, under Grant NSTC 115-2628-E-006-001 The authors thank the National Center for Theoretical Sciences (NCTS) and the National Center for High-performance Computing (NCHC) for providing the computing resources References Yuan, D., Elvidge, C.D., 1996.
Rosa: Comparison of relative radiometric normalization techniques Rahman, M., Hay, G., Couloigner, I., Hemachandran, B., Bailin, J., 2015.
Dev: A comparison of four relative radiometric normalization (RRN) techniques for mosaicing H-res multi-temporal thermal infrared (TIR) flight-lines of a complex urban scene Guindon, B., 1997.
Rosa: Assessing the radiometric fidelity of high resolution satellite image mosaics Lin, C.H., Lin, B.Y., Lee, K.Y., Chen, Y.C., 2015.
Paper summary: Rosa: So we're looking at this paper called "CODE-PIF: Convex/Deep Multiview Pseudoinvariant Features for Multitemporal Multispectral Surface Reflectance Imagery". We're dealing with multitemporal multispectral surface reflectance imagery where the readings change because of things like the atmosphere, different sensors, or actual changes on the ground.
Dev: Which makes finding these pseudoinvariant features, or PIFs, really tricky because those readings are not consistent across time and space.
Taro: So what's the main thing this paper is trying to do? It claims they have a framework called CODE-PIF that refines an initial set of PIFs that they got from something called DeepMGCCA using convex optimization.
Rosa: Right, so DeepMGCCA gives you this coarse initial PIF mask, and then CODE-PIF takes that and refines it by adding invariance scores calculated from spectral and temporal information through convex optimization. It's not just taking the first set of features; they are refining the selection process itself.
Dev: That refinement is motivated by the idea that just having similar features in some learned latent space doesn't mean those original reflectance observations are actually consistent in terms of radiance or spectrum. They need something more rigorous than what DeepMGCCA provides on its own, which is why they introduce this optimization step.
Taro: So how do they build that initial signal? They start by ranking pixels based on normalized alteration scores for each acquisition date and then selecting the lowest scoring fraction to create the initial PIF mask. Then, the deep change prior is defined as one minus that initial mask, assigning zero to those pixels selected as PIFs by DeepMGCCA and one to everything else.
Rosa: And that sets up this data-fitting term in their convex optimization problem which relates these continuous change scores to spectral and temporal information derived from observations. It incorporates spectral angle measurements and reflectance residuals weighted by a temporal reference, which is supposed to promote radiometric and spectral consistency.
Dev: That sounds like they are using the structure of the measurements themselves, like how much the angles are changing over time or how much the reflectance values are off from what we expect based on time, to guide this continuous change score mapping. It's linking the observed data directly into the optimization loop.
Taro: The gains they show come from testing it across four different places—Turkey, Taipei, Tainan, and Saudi Arabia—and they claim improved average radiometric and spectral agreement compared to DeepMGCCA.
Rosa: And what are these specific improvements? They say that averaged across these four regions, the root mean square error drops from zero point zero one nine two down to zero point zero one six four, and the peak signal-to-noise ratio increases from thirty-five point nine one dB up to thirty-seven point four four dB.
Dev: Plus they report that the between-date spectral angle decreases from three point seven seven degrees down to two point nine seven degrees across those same regions, which is a pretty significant reduction in that angular difference.
Taro: Qualitatively, they found that CODE-PIF excludes open-water regions initially selected by DeepMGCCA. This means the selections are more spread out over built-up and bare surfaces instead of being clustered in large contiguous areas.
Rosa: That's interesting because it suggests the refinement process is more discriminating about where it places its invariant targets, focusing on stable ground rather than just following the initial deep mask blindly. So, if you're driving somewhere, this means the system is trying to find features that are actually physically stable across time and space.
Dev: It also shows that quantitatively, CODE-PIF achieves lower root mean square error and a smaller between-date spectral angle in all four regions when compared to DeepMGCCA. The regional improvements are consistent with scatterplots, with the biggest gains seen in Turkey and Taipei.
Taro: So the core idea is that this convex/deep framework, CODE-PIF, refines those initial DeepMGCCA PIFs using mathematically simple convex optimization for harmonization across multitemporal surface reflectance imagery.
Rosa: And what does this mean for us? It suggests a way to harmonize these complex time series data sets—the atmospheric effects and sensor differences—in a way that actually preserves the physical consistency of the surface features, rather than just getting a visually consistent map.
Dev: The paper's authors note that the reliability of CODE-PIF really depends on the quality of both that initial DeepMGCCA selection and those change scores derived from reflectance. They also point out a limitation: this pixelwise data-fitting map doesn't explicitly model local spatial relationships, which means it might miss variations in reflectance magnitude that still maintain the spectral shape.
Taro: And they stress that the experimental protocol defines the scope of these improvements; they aren't claiming physical invariance for every single selected pixel independently, just within this specific calibration pipeline.
Rosa: So, to wrap up this summary of "CODE-PIF: Convex/Deep Multiview Pseudoinvariant Features for Multitemporal Multispectral Surface Reflectance Imagery", it’s an unsupervised convex framework that refines DeepMGCCA-derived PIFs through mathematically simple convex optimization for harmonizing multitemporal surface reflectance imagery.
Dev: The extensive experiments across those four geographically distinct regions demonstrate improved average radiometric and spectral agreement compared to DeepMGCCA, achieving state-of-the-art performance in some metrics like the root mean square error dropping from zero point zero one nine two to zero point zero one six four and the between-date spectral angle decreasing from three point seven seven degrees to two point nine seven degrees.
Taro: Future work mentioned is developing an interpretable architecture for multitemporal surface reflectance harmonization, so they plan to investigate a convex formulation for this harmonization and unfold the associated iterative solver into a trainable network.
Rosa: So, the authors conclude that when you select pseudoinvariant features, you need to check them not just for agreement but also for spectral stability and how they sit spatially.
Dev: It’s about making sure these PIFs are reliable tools for change detection by checking their performance across all those different factors.
Taro: That leads us into how this method might be used in other areas where you're trying to compare data from different sensors or different times.
Conclusion: Rosa: So, we're wrapping up our look at CODE-PIF, which is that Convex/Deep Multiview Pseudoinvariant Features for Multitemporal Multispectral Surface Reflectance Imagery paper.
Dev: Yeah, it’s an unsupervised framework that takes the initial features from DeepMGCCA and refines them using convex optimization with spectral and temporal scores.
Taro: It’s basically making sure the pixels you pick are actually consistent in their raw readings across different times and sensors, not just similar in some abstract mathematical space.
Rosa: That refinement is about ensuring that when you look at those surface features, they actually match up physically, rather than just looking the same on a computer screen.
Dev: The gains they showed across four regions—Turkey, Taipei, Tainan and Saudi Arabia—is a measurable improvement in radiometric and spectral agreement.
Taro: They cut the root mean square error from about zero point one nine two down to zero point one six four and the between-date spectral angle goes down from three point seven seven degrees to two point nine seven degrees.
Rosa: That’s a noticeable drop in those angular differences, which suggests better harmonization of the data when you're dealing with time series imagery.
Dev: It means when you’re trying to look at how surfaces change over time, this method helps you get a more physically consistent picture without having to manually adjust every single pixel.
Taro: Qualitatively, it also seems to exclude open-water areas that DeepMGCCA picked initially, focusing the selection on built-up and bare ground instead.
Rosa: So the authors are saying that for harmonizing these complex time series data sets, this convex method provides a way to select features that are genuinely stable across space and time.
Dev: It’s about using mathematical structure to guide the feature selection so it aligns with real-world physical consistency, which is what matters when you're building a safety case for remote sensing data.
Taro: The paper points out their limitation is that this pixelwise mapping doesn't explicitly model local spatial relationships, so it might miss variations in reflectance magnitude that still keep the spectral shape consistent.
Rosa: It means you have to be careful; the method works within its defined scope, but you can’t assume it captures every little detail of the local surface variation.
Dev: So while the numbers look good across those four regions, we need to keep an eye on how this refinement handles those sharp local changes where things don't look smooth.
Taro: The future work they mentioned is focusing on developing a more interpretable architecture for this harmonization process, so they plan to unfold the iterative solver into a trainable network.
Rosa: So, the authors conclude that when you select pseudoinvariant features, you need to check them not just for agreement but also for spectral stability and how they sit spatially.
Dev: It’s about making sure these PIFs are reliable tools for change detection by checking their performance across all those different factors.
Taro: That leads us into how this method might be used in other areas where you're trying to compare data from different sensors or different times.
Chia-Hsiang Lina, Lino Garda Denaroa, Jian-Kai Huanga
Department of Electrical Engineering, National Cheng Kung University · Research Center for Geoinformatics, National Research and Innovation Agency (BRIN)
eess.IV
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: 14 pages, 6 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 88/100
The gist: Multitemporal multispectral surface reflectance imagery exhibits radiometric and spectral variability due to atmospheric effects, sensor differences, and actual surface changes…
Key concepts
- Pseudoinvariant Features (PIFs)
- These are pixels in a multispectral image that remain consistent across different time points or acquisitions, despite atmospheric effects or surface changes. The paper seeks to select these features robustly using deep learning and optimization.
- DeepMGCCA
- This is the initial method used to select coarse PIFs by ranking pixels based on normalized alteration scores for each acquisition date. It provides a starting point, but it lacks the fine-grained spectral consistency achieved by CODE-PIF.
- Convex Optimization
- This mathematical technique is used to refine the initial PIF mask. It guides the selection process by fitting continuous change scores to spectral and temporal information, ensuring radiometric and spectral consistency in a mathematically constrained way.
Terminology
Summary
Multitemporal multispectral surface reflectance imagery exhibits radiometric and spectral variability due to atmospheric effects, sensor differences, and actual surface changes, complicating the identification of pseudoinvariant features (PIFs) The proposed CODE-PIF architecture is an unsupervised convex/deep (CODE) framework that refines the coarse PIF from DeepMGCCA using convex optimization
Pruning redundant reasoning tokens
The paper introduces CODE-PIF, a framework that refines the initial pseudoinvariant feature (PIF) selection derived from DeepMGCCA by integrating the deep mask with invariance scores calculated from spectral and temporal information through convex optimization This refinement is motivated by the fact that agreement in the learned latent space does not necessarily imply radiometric and spectral consistency in original reflectance observations The final PIF selection is guided by a monotone piecewise-linear transformation to harmonize multitemporal reflectance observations, which selects calibration references while adjusting reflectance values without reversing their ordering
How the training signal is built
The initial PIF mask, denoted as DeepMGCCA, is obtained by ranking pixels based on normalized alteration scores for each acquisition date and selecting the lowest-scoring fraction to construct the initial PIF mask The deep change prior is then defined as 1 minus this initial PIF mask, assigning zero to pixels selected as PIFs by DeepMGCCA and one to all remaining pixels The data-fitting term in the convex optimization problem relates these continuous change scores to spectral and temporal information derived from observations This term incorporates spectral angle measurements and reflectance residuals weighted by a temporal reference to promote radiometric and spectral consistency
Where the gains come from
The gains of CODE-PIF are demonstrated across four geographically distinct regions—Turkey, Taipei, Tainan, and Saudi Arabia—showing improved average radiometric and spectral agreement relative to DeepMGCCA Averaged across these regions, the RMSE decreases from 0.0192 to 0.0164, PSNR increases from 35. p. 091 to 37.44 dB, and the between-date spectral angle decreases from 3. p. 077° to 2.97° The qualitative comparison shows that CODE-PIF excludes open-water regions initially selected by DeepMGCCA, resulting in spatially dispersed selections over built-up and bare surfaces rather than large contiguous regions
Quantitative Comparison of PIF Selection
The quantitative results show that the proposed CODE-PIF achieves lower RMSE and between-date spectral angle (SAMh), together with higher PSNR, than DeepMGCCA in all four regions Averaged across the ROIs, RMSE decreases from 0.0192 to 0.0164, corresponding to an approximately 14. p. 65% reduction, while SAMh decreases from 3. p. 977◦ to 2.97◦ The regional improvements are consistent with the scatterplots, with the most pronounced gains observed in Turkey and Taipei
Discussion
The reliability of CODE-PIF depends on the quality of both the initial DeepMGCCA selection and reflectancederived change scores The pixelwise data-fitting map does not explicitly model local spatial relationships and may overlook variations in reflectance magnitude that preserve spectral shape The experimental protocol defines the scope of the reported improvements, as results quantify the contribution of PIF refinement within this calibration pipeline and do not independently establish the physical invariance of every selected pixel
Conclusions and Future Works
In conclusion, CODE-PIF is an unsupervised CODE framework that refines DeepMGCCA-derived PIFs through mathematically simple convex optimization for multitemporal surface reflectance harmonization The extensive experiments across four geographically distinct regions demonstrate improved average radiometric and spectral agreement compared with DeepMGCCA, thereby achieving state-of-the-art performance Future work will focus on developing an interpretable architecture for multitemporal surface reflectance harmonization, facilitating explicit analysis of how model updates affect radiometric and spectral consistency Motivated by recent advances in deep unfolding for remote sensing, the authors plan to investigate a convex formulation for multitemporal surface reflectance harmonization and unfold the associated iterative solver into a trainable network The paper concludes that PIF selection should be assessed jointly in terms of reflectance agreement, spectral stability, spatial structure, and linear correlation
Acknowledgements
This study was supported by the Emerging Young Scholar Program (the 2030 Cross-Generation Young Scholars Program) of National Science and Technology Council (NSTC), Taiwan, under Grant NSTC 115-2628-E-006-001 The authors thank the National Center for Theoretical Sciences (NCTS) and the National Center for High-performance Computing (NCHC) for providing the computing resources
References
Yuan, D., Elvidge, C. p. 0D., 1996. Comparison of relative radiometric normalization techniques Rahman, M. p. 0, Hay, G., Couloigner, I., Hemachandran, B., Bailin, J., 2015. A comparison of four relative radiometric normalization (RRN) techniques for mosaicing H-res multi-temporal thermal infrared (TIR) flight-lines of a complex urban scene Guindon, B., 1997. Assessing the radiometric fidelity of high resolution satellite image mosaics Lin, C. p. 6H., Lin, B.Y., Lee, K.Y., Chen, Y.C., 2015. Radiometric normalization and cloud detection of optical satellite images using invariant pixels Syariz, M. p. 0A., Lin, B.Y., Denaro, L. p. 0 p. 0G., Jaelani, L.M., Van Nguyen, M., Lin, C.H., 2019. Spectral-consistent relative radiometric normalization for multitemporal Landsat 8 imagery Clark, B. p. 0, Suomalainen, J., Pellikka, P., 2011. The selection of appropriate spectrally bright pseudo-invariant ground targets for use in empirical line calibration of SPOT satellite imagery Mei, A. p. 0, Bassani, C., Fontinovo, G., Salvatori, R., Allegrini, A., 2016. The use of suitable pseudo-invariant targets for MIVIS data calibration by the empirical line method Xu, Q. p. 0, Hou, Z., Tokola, T., 2012. Relative radiometric correction of multitemporal ALOS AVNIR-2 data for the estimation of forest attributes Eivazi, A. p. 0, Kolesnikov, A., Junttila, V., Kauranne, T., 2015. Variancepreserving mosaicing of multiple satellite images for forest parameter estimation: Radiometric normalization Volpi, M. p. 0, Camps-Valls, G., Tuia, D., 2015. Spectral alignment of multitemporal cross-sensor images with automated kernel canonical correlation analysis Andrew, G. p. 0, Arora, R., Bilmes, J., Livescu, K., 2013. Deep canonical correlation analysis, in: Dasgupta, S. p. 0, McAllester, D. (Eds.), Proc. International Conference on Machine Learning Canty, M. p. 0J., Nielsen, A. p. 0 p. 0A., 2008. Automatic radiometric normalization of multitemporal satellite imagery with the iteratively re-weighted mad transformation Canty, M. p. 0J., Nielsen, A.A., Schmidt, M., 2004. Automatic radiometric normalization of multitemporal satellite imagery Chambolle, A. p. 0, Jan. 2004. An algorithm for total variation minimization and applications Chen, H. p. 1, Song, J., Wu, C., Du, B., Yokoya, N., 2023a. Exchange means change: An unsupervised single-temporal change detection framework based on intra- and inter-image patch exchange Chen, H. p. 1, Yokoya, N., Chini, M., 2023b. Fourier domain structural relationship analysis for unsupervised multimodal change detection Chi, C. p. 1Y., Li, W.C., Lin, C.H., 2017. Convex Optimization for Signal Processing and Communications Denaro, L. p. 1G., Lin, C.H., 2020. Hybrid canonical correlation analysis and regression for radiometric normalization of cross-sensor satellite imagery Denaro, L. p. 0G., Lin, C.H., 2026a. Constrained harmonization of optical satellite imagery for surface reflectance bias removal Denaro, L. p. 6G., Lin, C.H., 2026b. Monotone piecewise-linear harmonization of multitemporal multispectral surface reflectance with unsupervised deep multiview pseudoinvariant features learning Fang, B. p. 0, Chen, G., Kou, R., Paoletti, M.E., Haut, J.M., Plaza, A., 2023. CIT: Content-invariant translation with hybrid attention mechanism for unsupervised change detection Gan, W. p. 1, Feng, Y., Miao, J., Li, X., Shen, H., 2026. WEGLA-NormGAN: wavelet-enhanced Cycle-GAN with global-local attention for radiometric normalization of remote sensing images Hall, F., Strebel, D., Nickeson, J., Goetz, S., 1991. Radiometric rectification: Toward a common radiometric response among multidate, multisensor images Horst, P., 1961. Relations among m sets of measures Hussain, M. p. 1, Chen, D., Cheng, A. p. 0, Wei, H., Stanley, D., 2013. Change detection from remotely sensed images: From pixel-based to objectbased approaches Nielsen, A. p. 1A., Conradsen, K., Simpson, J.J., 1998. Multivariate alteration detection (MAD) and MAF postprocessing in multispectral, bitemporal image data: New approaches to change detection studies Lin, B. p. 0Y., Wang, Z.J., Syariz, M. p. 0A., Denaro, L.G., Lin, C.H., 2019. Pseudoinvariant feature selection using multitemporal mad for optical satellite images Lin, C. p. 0H., Bioucas-Dias, J. p. 1M., Dec. 2019. An explicit and scene-adapted definition of convex self-similarity prior with application to unsupervised Sentinel-2 super-resolution Lin, C. p. 1H., Chen, J.T., Leng, Z.C., Lin, J.T., 2025a. COS2A: Conversion from sentinel-2 to AVIRIS hyperspectral data using interpretable algorithm with spectral-spatial duality Lin, C. p. 2H., Chen, Y.Y., 2023. HyperQUEEN: Hyperspectral quantum deep network for image restoration Lin, C. p. 1H., Chu, M.C., Tang, P.W., 2023. CODE-MM: Convex deep mangrove mapping algorithm based on optical satellite images Lin, C. p. 1H., Hsieh, C.Y., Lin, J.T., 2024a. CODE-IF: A convex/deep image fusion algorithm for efficient hyperspectral super-resolution Lin, C. p. 1H., Hsu, C.C., Young, S. p. 1 p. 1S., Hsieh, C.Y., Tai, S.C., Mar. 2024. QRCODE: Quasi-residual convex deep network for fusing misaligned hyperspectral and multispectral images Lin, C. p. 1H., Hsu, S.M., Liang, C.Y., Chanussot, J., Chen, J.Y., 2026a. Hyperspectral calibration detection: A novel concept for change detection with unsupervised incremental safe pseudo-labeling implementation Lin, C. p. 1H., Huang, J. p. 0 p. 0K., Young, S.S., Zheng, W. p. 1C., 2026b. Interpretable landsat-to-hyperspectral dual super-resolution without large matrix inversion Lin, C. p. 0H., Huang, S. p. 0H., Lin, T.H., Wu, P.C., Nov. 2023. Metasurfaceempowered snapshot hyperspectral imaging with convex/deep (CODE) small-data learning theory Miao, J. p. 1, Li, S., Bai, X., Gan, W., Wu, J., Li, X., 2025. RS-NormGAN: Enhancing change detection of multi-temporal optical remote sensing images through effective radiometric normalization Nielsen, A. p. 0A., Conradsen, K., Simpson, J.J., 1998. Multivariate alteration detection (MAD) and MAF postprocessing in multispectral, bitemporal image data: New approaches to change detection studies Obata, K. p. 0, Yoshioka, H. p. 0, 2024. Unmixing-based radiometric and spectral harmonization for consistency of multi-sensor reflectance time-series data Razzak, M. p. 2T., Mateo-García, G., Lecuyer, G., Gómez-Chova, L., Gal, Y., Kalaitzis, F., 2023. Multi-spectral multi-image super-resolution of Sentinel-2 with radiometric consistency losses and its effect on building delineation Sun, Y. p. 0, Lei, L., Li, Z., Kuang, G., 2024. Similarity and dissimilarity relationships based graphs for multimodal change detection Tang, P. p. 1W., Lin, C.H., Huang, J.K., Huete, A.R., 2025. A quantumempowered SPEI drought forecasting algorithm using spatially aware mamba network Trevisiol, F. p. 2, Mandanici, E., Pagliarani, A., Bitelli, G., 2024. Evaluation of Landsat-9 interoperability with Sentinel-2 and Landsat-8 over Europe and local comparison with field surveys Tu, L. p. 0, Huang, X. p. 0, Li, J., Yang, J., Gong, J., 2024. A multi-task learning method for extraction of newly constructed areas based on bi-temporal hyperspectral images Tuia, D., Marcos, D., Camps-Valls, G., 2016. Multi-temporal and multisource remote sensing image classification by nonlinear relative normalization Wang, W. p. 0, Arora, R., Livescu, K., Bilmes, J., 2015. On deep multi-view representation learning, in: Bach, F. p. 0, Blei, D. (Eds.), Proc. International Conference on Machine Learning Xie, R. p. 0, Xia, M., Yao, J., Li, L., 2018. Guided color consistency optimization for image mosaicking Xu, J. p. 0, Liu, T., Lei, T., Chen, H., Yokoya, N., Lv, Z., Gong, M., 2025. CGSL: Commonality graph structure learning for unsupervised multimodal change detection Zhang, C. p. 1, Yue, P., Tapete, D., Jiang, L., Shangguan, B., Huang, L. p. 0, Liu, G., 2020. A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images Zhang, E. p. 1, Li, X., Zong, H., Gao, Y., Ren, J., 2026. Disentangling representation and cross-domain adaptation for unsupervised change detection in multimodal remote sensing images Zhao, C. p. 1, Goshtasby, A. p. 6A., 2016. Registration of multitemporal aerial optical images using line features Zheng, Z. p. 6, Wan, Y., Zhang, Y. p. 1, Xiang, S., Peng, D., Zhang, B. p. 1, 2021. CLNet: Cross-layer convolutional neural network for change detection in optical remote sensing imagery Zheng, Z. p. 1, Zhong, Y., Zhao, J., Ma, A., Zhang, L. p. 1, 2024.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved PIF selection via CODE-PIF allows for
re-ranking of initially selected pixels and consideration of candidates excluded by the initial PIFs from DeepMGCCA
by integratingspectral angle measurements and temporal-reference change metrics.
-
The refined system can achieve improved radiometric agreement, as shown by an average RMSE decrease from 0.0192 to 0.0164 across four regions, and a SAMh decrease from 3.77 degrees to 2.97 degrees in the same comparison against DeepMGCCA.
-
The system can perform better on scenes with complex features by excluding
open-water regions initially selected by DeepMGCCA
and retaining onlyspatially dispersed selections over built-up and bare surfaces rather than large contiguous regions.
Sources
- Monotone Piecewise-Linear Harmonization of Multitemporal Multispectral Surface Reflectance With Unsupervised Deep Multiview Pseudoinvariant Features Learning
- COS2A: Conversion from Sentinel-2 to AVIRIS Hyperspectral Data Using Interpretable Algorithm With Spectral-Spatial Duality
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation