pith. sign in

arxiv: 2604.18853 · v1 · submitted 2026-04-20 · 💻 cs.CV

DDF2Pol: A Dual-Domain Feature Fusion Network for PolSAR Image Classification

Mohammed Q. Alkhatib This is my paper

Pith reviewed 2026-05-10 04:36 UTC · model grok-4.3

classification 💻 cs.CV
keywords PolSAR image classificationdual-domain CNNcomplex-valued networksfeature fusioncoordinate attentionlightweight model
0
0 comments X

The pith

DDF2Pol combines real and complex feature streams to classify PolSAR images more accurately with fewer parameters than existing models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a lightweight convolutional network that processes PolSAR data through two parallel paths, one handling real numbers and one handling complex numbers, to capture both spatial patterns and polarimetric details. It adds depth-wise convolutions and coordinate attention to sharpen these features before classification. On standard test images from Flevoland and San Francisco, the network reaches overall accuracies above 96 percent while using under 100,000 parameters, beating several earlier real-valued and complex-valued approaches. A sympathetic reader would care because PolSAR classification supports land-use mapping and environmental monitoring, and a simpler yet stronger model makes these tasks more accessible when labeled data is scarce.

Core claim

DDF2Pol integrates two parallel feature extraction streams—one real-valued and one complex-valued—to capture complementary spatial and polarimetric information from PolSAR data, followed by depth-wise convolution for spatial enhancement and coordinate attention to focus on informative regions, achieving overall accuracies of 98.16% on Flevoland and 96.12% on San Francisco with only 91,371 parameters.

What carries the argument

The dual-domain feature fusion with parallel real and complex streams, enhanced by depth-wise convolution and coordinate attention.

Load-bearing premise

The dual real-complex streams together with depth-wise convolution and coordinate attention actually supply complementary information that accounts for the accuracy improvements over existing models.

What would settle it

Training the network without one of the two streams and checking whether accuracy drops below the reported levels on the same datasets would test if the dual-domain design is necessary for the gains.

Figures

Figures reproduced from arXiv: 2604.18853 by Mohammed Q. Alkhatib.

Figure 1
Figure 1. Figure 1: Overall Architecture of the Proposed DDF2Pol Model. layers and filters are employed, but with 3D complex-valued convolutional filters of the same size (3 × 3 × 3). These filters enable the simultaneous extraction of spatial and polarimetric features by capturing local spatial dependencies while preserving interactions across the three polarimetric channels. To integrate complex-valued features into the sub… view at source ↗
Figure 2
Figure 2. Figure 2: CNNs Feature Extractor. M.Q. Alkhatib: Preprint submitted to Elsevier Page 8 of 7 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Classification results of the Flevoland dataset. (a) Pauli RGB; (b) Reference Class Map; (c) 3D-CNN; (d) WaveletCNN; (e) PolSARFormer; (f) 3D-CV-CNN; (g) CV-2D-3D; (h) HybridCVNet; (i) Proposed DDF2Pol [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Classification results of the San Francisco dataset. (a) Pauli RGB; (b) Reference Class Map; (c) 3D-CNN; (d) WaveletCNN; (e) PolSARFormer; (f) 3D CV-CNN; (g) CV-2D-3D; (h) HybridCVNet; (i) Proposed DDF2Pol [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overall accuracy at different percentages of training data (a) Flevoland (b) San Francisco. M.Q. Alkhatib: Preprint submitted to Elsevier Page 9 of 7 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
read the original abstract

This paper presents DDF2Pol, a lightweight dual-domain convolutional neural network for PolSAR image classification. The proposed architecture integrates two parallel feature extraction streams, one real-valued and one complex-valued, designed to capture complementary spatial and polarimetric information from PolSAR data. To further refine the extracted features, a depth-wise convolution layer is employed for spatial enhancement, followed by a coordinate attention mechanism to focus on the most informative regions. Experimental evaluations conducted on two benchmark datasets, Flevoland and San Francisco, demonstrate that DDF2Pol achieves superior classification performance while maintaining low model complexity. Specifically, it attains an Overall Accuracy (OA) of 98.16% on the Flevoland dataset and 96.12% on the San Francisco dataset, outperforming several state-of-the-art real- and complex-valued models. With only 91,371 parameters, DDF2Pol offers a practical and efficient solution for accurate PolSAR image analysis, even when training data is limited. The source code is publicly available at https://github.com/mqalkhatib/DDF2Pol

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes DDF2Pol, a lightweight dual-domain convolutional neural network for PolSAR image classification. It consists of parallel real-valued and complex-valued feature extraction streams to capture complementary spatial and polarimetric information, followed by depth-wise convolution for spatial enhancement and coordinate attention to emphasize informative regions. On the Flevoland and San Francisco benchmark datasets, the model is reported to achieve overall accuracies of 98.16% and 96.12% respectively while using only 91,371 parameters, outperforming several state-of-the-art real- and complex-valued models; the source code is made publicly available.

Significance. If the reported performance gains prove reproducible under matched experimental conditions, the work would contribute a practical, low-complexity architecture suitable for PolSAR classification tasks with limited training data. The public release of the implementation code is a clear strength that supports verification and potential adoption.

major comments (2)
  1. [Experiments] The Experiments section provides no description of the experimental protocol, including data splits, patch extraction sizes, preprocessing steps, training hyperparameters, or the number of runs. Without this information it is impossible to determine whether the baseline models were re-implemented and evaluated under identical conditions to DDF2Pol, which directly affects the validity of the superiority claims (OA 98.16% and 96.12%).
  2. [Results] No error bars, standard deviations, or statistical significance tests are reported for the accuracy figures in the results tables. This omission makes it difficult to assess whether the observed improvements over competing real- and complex-valued models are robust or could arise from random variation in a single training run.
minor comments (1)
  1. [Abstract] The abstract states that the model 'outperforms several state-of-the-art' models but does not name the specific baselines; adding the names would improve immediate readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We appreciate the emphasis on experimental reproducibility and statistical rigor. We address each major comment below and will revise the manuscript accordingly to incorporate the requested details.

read point-by-point responses

  1. Referee: [Experiments] The Experiments section provides no description of the experimental protocol, including data splits, patch extraction sizes, preprocessing steps, training hyperparameters, or the number of runs. Without this information it is impossible to determine whether the baseline models were re-implemented and evaluated under identical conditions to DDF2Pol, which directly affects the validity of the superiority claims (OA 98.16% and 96.12%).

    Authors: We agree that the experimental protocol was not described in sufficient detail. In the revised manuscript we will insert a dedicated 'Experimental Setup' subsection that specifies the data splitting strategy used for the Flevoland and San Francisco benchmarks, the patch extraction sizes, all preprocessing operations, the complete set of training hyperparameters (optimizer, learning rate, batch size, epochs, weight decay, etc.), and the number of runs performed. We confirm that every baseline was re-implemented from its original publication and evaluated under identical data splits, preprocessing, and training conditions as DDF2Pol; the released source code will be updated to expose these exact settings for verification. revision: yes

  2. Referee: [Results] No error bars, standard deviations, or statistical significance tests are reported for the accuracy figures in the results tables. This omission makes it difficult to assess whether the observed improvements over competing real- and complex-valued models are robust or could arise from random variation in a single training run.

    Authors: We acknowledge the validity of this concern. The revised manuscript will report results from multiple independent training runs (different random seeds) and will include mean overall accuracy together with standard deviations in all tables. In addition, we will add statistical significance tests (paired t-test or Wilcoxon signed-rank test, as appropriate) comparing DDF2Pol against each baseline to establish that the reported gains are statistically meaningful rather than attributable to run-to-run variation. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical accuracies on public benchmarks are direct experimental outputs

full rationale

The paper's central claims consist of reported Overall Accuracy values (98.16% on Flevoland, 96.12% on San Francisco) obtained by training and evaluating the proposed DDF2Pol network on standard PolSAR benchmark datasets. These are measured performance numbers, not quantities derived from equations or parameters that reduce to the paper's own inputs by construction. The architecture description (dual real/complex streams, depth-wise convolution, coordinate attention) is a design choice whose effectiveness is validated externally via cross-validation on held-out test patches; no self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain exists in the reported results. The public code release further supports independent verification against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests entirely on empirical classification accuracies obtained by training the proposed network on two standard PolSAR datasets. No mathematical axioms, free parameters in the derivation sense, or invented physical entities are invoked.

pith-pipeline@v0.9.0 · 5489 in / 1109 out tokens · 42748 ms · 2026-05-10T04:36:28.550159+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    Zhang, H

    H. Zhang, H. Lin, Y. Li, Y. Zhang, C. Fang, Mapping urban impervious surface with dual-polarimetric SAR data: An improved method, Landscape and urban planning 151 (2016) 55–63

    work page 2016

  2. [2]

    Ponnurangam, Y

    G. Ponnurangam, Y. Rao, The application of compact polarimetric decomposition algorithms to L-band PolSAR data in agricultural areas, InternationalJournalofRemoteSensing39(2018)8337–8360

    work page 2018

  3. [3]

    E. G. Tebege, Z. M. Birara, S. G. Takele, M. Jothimani, Geospatial mapping and multi-criteria analysis of groundwater potential in Libo Kemkemwatershed,upperblueNileRiverbasin,Ethiopia, Scientific African 27 (2025) e02549

    work page 2025

  4. [4]

    Rudant, Support vector machine for multifrequency SAR po- larimetric data classification, IEEE Trans

    C.Lardeux,P.-L.Frison,C.Tison,J.-C.Souyris,B.Stoll,B.Fruneau, J.-P. Rudant, Support vector machine for multifrequency SAR po- larimetric data classification, IEEE Trans. Geosci. Remote Sens. 47 (2009) 4143–4152

    work page 2009

  5. [5]

    Z. Qi, A. G.-O. Yeh, X. Li, Z. Lin, A novel algorithm for land use and land cover classification using RADARSAT-2 polarimetric SAR data, Remote Sensing of Environment 118 (2012) 21–39

    work page 2012

  6. [6]

    Y. Zhou, H. Wang, F. Xu, Y.-Q. Jin, Polarimetric SAR image classi- fication using deep convolutional neural networks, IEEE Geoscience and Remote Sensing Letters 13 (2016) 1935–1939

    work page 2016

  7. [7]

    Zhang, Z

    L. Zhang, Z. Chen, B. Zou, Y. Gao, Polarimetric SAR terrain classification using 3D convolutional neural network, in: IGARSS 2018-2018IEEEInternationalGeoscienceandRemoteSensingSym- posium, IEEE, 2018, pp. 4551–4554

    work page 2018

  8. [8]

    Zhang, H

    Z. Zhang, H. Wang, F. Xu, Y.-Q. Jin, Complex-valued convolutional neural network and its application in polarimetric SAR image classi- fication, IEEE Trans. Geosci. Remote Sens. 55 (2017) 7177–7188

    work page 2017

  9. [9]

    X. Tan, M. Li, P. Zhang, Y. Wu, W. Song, Complex-valued 3-D convolutional neural network for PolSAR image classification, IEEE Geoscience and Remote Sensing Letters 17 (2019) 1022–1026

    work page 2019

  10. [10]

    J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    work page 2018

  11. [11]

    Q. Hou, D. Zhou, J. Feng, Coordinate attention for efficient mobile network design, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 13713–13722

    work page 2021

  12. [12]

    Y. Kong, J. Li, K. Zhang, J. Wu, Multi-scale self-attention mixup for graphclassification,PatternRecognitionLetters168(2023)100–106

    work page 2023

  13. [13]

    R. Yang, X. Xu, R. Gui, Z. Xu, F. Pu, Composite sequential network with POA attention for PolSAR image analysis, IEEE Trans. Geosci. Remote Sens. 60 (2021) 1–15

    work page 2021

  14. [14]

    K.Karachristos,G.Koukiou,V.Anastassopoulos, Fusionofcoherent and non-coherent pol-sar features for land cover classification, Elec- tronics 13 (2024) 634

    work page 2024

  15. [15]

    Papadopoulos, V

    S. Papadopoulos, V. Anastassopoulos, G. Koukiou, Pixel-level deci- sion fusion for land cover classification using polsar data and local pattern differences., Electronics (2079-9292) 13 (2024)

    work page 2079

  16. [16]

    Papadopoulos, G

    S. Papadopoulos, G. Koukiou, V. Anastassopoulos, Correlated de- cision fusion accompanied with quality information on a multi-band pixelbasisforlandcoverclassification, JournalofImaging10(2024) 91

    work page 2024

  17. [17]

    H. Dong, L. Zhang, B. Zou, Polsar image classification with lightweight 3d convolutional networks, Remote Sensing 12 (2020) 396

    work page 2020

  18. [18]

    Zhang, L

    S. Zhang, L. Cui, Z. Dong, W. An, A Deep Learning Classification Scheme for PolSAR Image Based on Polarimetric Features, Remote Sensing 16 (2024) 1676

    work page 2024

  19. [19]

    F. Han, L. Zhang, H. Dong, Trusted polarimetric feature fusion for polsar image classification, in: 2023 SAR in Big Data Era (BIGSARDATA), IEEE, 2023, pp. 1–4

    work page 2023

  20. [20]

    Homayouni, PolSAR image classification based on deep con- volutional neural networks using wavelet transformation, IEEE Geoscience and Remote Sensing Letters 19 (2022) 1–5

    A.Jamali,M.Mahdianpari,F.Mohammadimanesh,A.Bhattacharya, S. Homayouni, PolSAR image classification based on deep con- volutional neural networks using wavelet transformation, IEEE Geoscience and Remote Sensing Letters 19 (2022) 1–5

    work page 2022

  21. [21]

    Jamali, S

    A. Jamali, S. K. Roy, A. Bhattacharya, P. Ghamisi, Local window attentiontransformerforpolarimetricSARimageclassification,IEEE Geoscience and Remote Sensing Letters 20 (2023) 1–5

    work page 2023

  22. [22]

    M. Q. Alkhatib, PolSAR image classification using complex-valued multiscaleattentionvisiontransformer(CV-MsAtViT), International JournalofAppliedEarthObservationandGeoinformation137(2025) 104412

    work page 2025

  23. [23]

    W.Li,H.Xia,J.Zhang,Y.Wang,Y.Jia,Y.He, Complex-valued2D- 3Dhybridconvolutionalneuralnetworkwithattentionmechanismfor POLSAR image classification, Remote Sensing 16 (2024) 2908

    work page 2024

  24. [24]

    S. Pei, H. Song, Y. Lu, Small sample hyperspectral image classifi- cation method based on dual-channel spectral enhancement network, Electronics 11 (2022) 2540

    work page 2022

  25. [25]

    H. Yang, D. Zhang, P. Xie, X. Chen, Dsc-grunet: A lightweight neural network model for multimodal gesture recognition based on depthwise separable convolutions and gru, Pattern Recognition Letters (2025)

    work page 2025

  26. [26]

    P. Yu, A. K. Qin, D. A. Clausi, Unsupervised polarimetric SAR imagesegmentationandclassificationusingregiongrowingwithedge penalty, IEEE Transactions on Geoscience and Remote Sensing 50 (2011) 1302–1317

    work page 2011

  27. [27]

    Y. Cao, Y. Wu, M. Li, W. Liang, P. Zhang, PolSAR image classi- fication using a superpixel-based composite kernel and elastic net, Remote Sensing 13 (2021) 380

    work page 2021

  28. [28]

    X. Xing, Q. Chen, S. Yang, X. Liu, Feature-based nonlocal polari- metric SAR filtering, Remote Sensing 9 (2017) 1043

    work page 2017

  29. [29]

    X. Liu, L. Jiao, F. Liu, D. Zhang, X. Tang, PolSF: PolSAR image datasetsonsanFrancisco,in:InternationalConferenceonIntelligence Science, Springer, 2022, pp. 214–219

    work page 2022

  30. [30]

    M. Q. Alkhatib, PolSAR Image Classification using a Hybrid Complex-Valued Network (HybridCVNet), IEEE Geoscience and Remote Sensing Letters (2024). M.Q. Alkhatib:Preprint submitted to ElsevierPage 7 of 7 PolSAR Image Classification Using DDF2Pol Figure 2:CNN s Feature Extractor. M.Q. Alkhatib:Preprint submitted to ElsevierPage 8 of 7 PolSAR Image Classifi...

    work page 2024