{"id":"c723b63a-f2fa-405b-bcec-73c066c4879e","arxiv_id":"2502.08137","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HPDNet processes PolSAR covariance matrices directly on the complex HPD manifold and reports state-of-the-art classification accuracy on three datasets.","lead":"This paper introduces HPDNet, a deep learning model that works directly with complex Hermitian positive definite matrices on a curved manifold instead of flattening them into vectors, and applies it to classify polarimetric SAR images. It reports higher classification accuracy than six existing methods on three real-world radar datasets, especially in hard-to-classify regions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CM-ASQRT cannot yield the eigendecomposition needed for ReEig/LogEig; the paper never specifies how eigenvectors and eigenvalues are recovered from the iterative square-root, so the claimed speedup and the described network operations are unsupported.","rationale":"The reader correctly identifies the CM-ASQRT description as the weakest assumption. My stress test sharpens the concern: the Newton–Schulz iteration computes square roots, not eigendecompositions, and the paper gives no mapping from (X_k, Z_k) to (U, Λ). This is not merely an accuracy-of-approximation issue; it is an internal incompleteness. The ReEig and LogEig layers are mathematically defined via eigendecomposition, so without U and Λ the network as specified cannot be implemented. The speedup table (Table VI) claims a 4.35x reduction in training time, which would be misleading if exact eigendecomposition is still used under the hood. The rest of the paper has independent value: the high-level architecture (manifold mapping to a tangent space followed by CV-3DCNN) is coherent, and the reported accuracies are consistent across three datasets. However, the gap in the fast-eigenvalue method directly undermines the reproducibility of both the accuracy and the efficiency claims. Since the reader's verdict is already CONDITIONAL and my concern aligns with the identified weak point, no verdict change is needed; the condition should be that the authors clarify or correct the CM-ASQRT implementation and provide code or a precise derivation.","tokens_in":18623,"tokens_out":5684,"duration_ms":45888,"concrete_test":"Inspect the released code (or request it) to determine whether any full eigendecomposition (e.g., torch.linalg.eigh) remains in the forward paths of ReEig and LogEig. Then run the Xi'an experiment in two configurations: (A) exact eigh for ReEig/LogEig with CM-ASQRT disabled, and (B) the described CM-ASQRT with an explicit, author-specified procedure for recovering U and Λ. Compare OA, Kappa, and training time to Tables I and VI; if variant B cannot reproduce variant A's 96.90% OA within sampling noise, the fast-method claim and the reported speedups are not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.B.4 (CM-ASQRT) claims that after the coupled Newton–Schulz iteration (Eqs. 14–18), 'the SVD decomposition can be approximately obtained.' This is the load-bearing step: the ReEig layer (Eq. 11) needs U and Λ to threshold eigenvalues, and the LogEig layer (Eq. 13) needs U and log(Λ). The iteration in [51] computes only C^{1/2} and C^{-1/2}; it provides no direct route to individual eigenvalues and eigenvectors. No formula is given that maps (X_k, Z_k) to (U, Λ). Without U and Λ, neither fr(Λ) nor log(Λ) can be applied, so the HPD network cannot be built as described. If the authors instead use an exact torch.linalg.eigh in their code, then CM-ASQRT is not actually used in ReEig/LogEig, contradicting the speedup claim in Table VI (training time drops from 663.24s to 152.35s). If they use an approximation, its effect on accuracy is unknown and unreported. This is an internal gap, independent of the plausibility of the Riemannian idea.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HPDNet, a two-module framework for PolSAR image classification. The first module operates on complex Hermitian positive definite (HPD) matrices directly in the Riemannian manifold, applying complex HPD mapping layers, a rectifying (ReEig) layer, and a LogEig layer that projects the manifold representation to a tangent space. The second module is a complex-valued 3D CNN that learns contextual features in Euclidean space. To reduce the cost of eigendecomposition, the authors introduce CM-ASQRT, an iterative complex matrix square-root method based on coupled Newton-Schulz iterations, and report that it accelerates training by a factor of roughly four. Experiments on the Xi'an, Oberpfaffenhofen, and Flevoland datasets report the highest overall accuracy and Kappa coefficient against six baselines, along with an ablation study and runtime comparison.","tokens_in":18923,"tokens_out":4483,"duration_ms":65595,"significance":"If the technical gaps described below are resolved, the paper would introduce a novel and useful direction: to the best of the authors' claim, it is the first complex HPD manifold network for PolSAR data, and the Riemannian-to-Euclidean design is a sensible way to combine manifold geometry with standard contextual deep learning. The reported accuracy gains are substantial on Xi'an and Oberpfaffenhofen, and the ablation study supports the contribution of the HPD module. The paper does not ship code or data, and the accuracy and runtime claims rest on single-run point estimates, so the empirical claims are not yet fully reproducible as written.","major_comments":[{"comment":"The definition of the rectifying function is internally contradictory: both branches of max(τI, Λ_{k-1})_{ii} return (Λ_{k-1})_{ii}. The second branch should return τ, otherwise the ReEig layer is the identity on the eigenvalues and cannot perform the intended nonlinear thresholding. Please correct Eq. (12) and verify that the implemented layer matches the corrected formula.","section":"III.B.2, Eq. (12)"},{"comment":"The claim that 'After several iterations, the SVD decomposition can be approximately obtained' is not supported. The coupled Newton-Schulz iteration computes only approximate C^{1/2} and C^{-1/2}; it does not directly provide the eigenvector matrix U and eigenvalue vector Λ required by the ReEig layer in Eq. (11) and the LogEig layer in Eq. (13). No formula is given that maps (X_k, Z_k) to (U, Λ). Please specify how the eigendecomposition is recovered, or revise the architecture so that the square-root approximation can be used where the eigendecomposition appears. Without this, the described network operations and the Table VI speedup claim are unsupported.","section":"III.B.4, Eqs. (14)-(18)"},{"comment":"All accuracy, Kappa, and runtime results are reported as single-run point estimates without error bars, confidence intervals, or repeated trials. This is particularly problematic for the Flevoland dataset, where the margin over PolMPCNN is only 0.64% OA and over HybridCVnet is 0.76% OA; these differences may be within run-to-run variability. Please provide multiple runs with means and standard deviations, or a statistical significance test, to support the claim of consistent superiority.","section":"IV, Tables I-III and Table VI"},{"comment":"In the Z update, the variable Y_{k-1} appears but is never defined. If this is a typo for Z_{k-1}X_{k-1} or X_{k-1}Z_{k-1}, the equation should be corrected and the iteration restated. As written, the coupled iteration is incomplete and cannot be implemented.","section":"III.B.4, Eq. (14)"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'Hermit' for 'Hermitian', 'Remianian' for 'Riemannian', 'non-dialog' for 'non-diagonal', 'buidlings' for 'buildings', and 'dateset' for 'dataset'. A thorough language edit is needed.","section":"Throughout"},{"comment":"The text refers to figures as 'Figs. 1(b)-(h)', 'Fig. 2(b)', and 'Figs. 2(b)-(h)', but the actual figures are numbered 5, 6, and 7. The figure cross-references should be corrected.","section":"IV.B-IV.D"},{"comment":"The training ratio (10%) and patch size (13×13) are selected using the Xi'an dataset and then applied to the other two datasets. Please justify this transferability or report per-dataset tuning, since such selection can inflate reported performance.","section":"IV.A.2 and IV.E"},{"comment":"The MRF smoothing operation used to refine the classification result is mentioned but not described; no parameters, neighborhood size, or implementation details are provided, which hinders reproducibility.","section":"IV.C.4"},{"comment":"References [22] and [44] are the same paper, as are [19] and [56]. Please consolidate or otherwise disambiguate them.","section":"References"},{"comment":"No code or trained models are released. Given the reliance on the CM-ASQRT implementation and the exact handling of complex HPD operations, providing code would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central Riemannian HPD idea is plausible and the reported accuracy improvements are promising, but the manuscript currently has a load-bearing gap: the fast eigendecomposition method does not, as described, produce the eigenvectors and eigenvalues needed by the ReEig and LogEig layers. This is fixable, but it requires either a derivation of how (U, Λ) are recovered from the Newton-Schulz iteration or a redesign of the layers to operate directly with the matrix square root. Combined with the ReEig typo and the lack of error bars, the paper needs a substantive revision before it can be accepted. The paper fits the journal's scope; the issues are technical rather than editorial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is the first complex HPD manifold network for PolSAR and the accuracy claims are plausible, but the fast eigendecomposition module (CM-ASQRT) as written cannot produce the eigenvectors and eigenvalues the ReEig and LogEig layers need, so the speedup numbers are unsupported.\n\nWhat's actually new: applying the SPDNet-style BiMap/ReEig/LogEig stack to complex Hermitian PD matrices, with the real/imaginary decomposition, and then projecting to a tangent space and feeding the complex vector to a CV-3DCNN. The ablation on Xi'an shows the HPD module alone (80.53%) plus CV-3DCNN alone (93.36%) gives 96.90% combined, so the manifold part contributes. The gains over six baselines are consistent across datasets, with the largest on Oberpfaffenhofen (90.94% vs 88.65% for the best baseline).\n\nThe soft spots are real but concentrated. The CM-ASQRT section (III.B.4) claims the Newton-Schulz iteration 'approximately' yields the SVD, but the iteration in [51] converges to C^{1/2} and C^{-1/2}, not to U and Lambda. No formula is given that maps the iterates to an eigendecomposition. If the authors are actually calling torch.linalg.eigh in the ReEig/LogEig layers, then CM-ASQRT is not used there and the reported 4.4x training speedup (663s to 152s) is unexplained. If they are using the approximation, its effect on accuracy is unknown. Either way, the paper needs to clarify this. Eq. (12) is also a typo: both branches return the same eigenvalue, so the rectifying function as printed is the identity. And the experimental tables have no error bars or repeated runs; the 0.64-2.16 point gains on Flevoland could easily be within run-to-run variation.\n\nWho this is for: the PolSAR classification community, and anyone working on manifold deep learning for complex matrices. It deserves a serious referee, not a desk reject, because the core idea is novel and the accuracy evidence is consistent. A referee should ask for code, a corrected CM-ASQRT description (or removal of the speedup claim), and error bars. The central empirical finding will probably survive; the speedup claim needs to be verified.","headline":"The first complex HPD manifold network for PolSAR is a plausible new idea with consistent accuracy gains, but the fast eigendecomposition module (CM-ASQRT) as written cannot produce the eigenvectors and eigenvalues the ReEig/LogEig layers need, so the speedup claim is unsupported.","tokens_in":19411,"tokens_out":4037,"would_cite":false,"duration_ms":41910,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that HPDNet, which processes PolSAR covariance matrices directly as complex Hermitian positive definite matrices on a Riemannian manifold before projecting them to Euclidean space, reaches the best reported accuracy on…","keywords":["PolSAR image classification","Hermitian positive definite matrix","Riemannian manifold learning","complex-valued convolutional network","LogEig layer","iterative matrix square root","covariance matrix geometry","remote sensing"],"falsifier":"On a PolSAR dataset, replace the CM-ASQRT approximation in the rectifying and LogEig layers with exact complex Hermitian eigendecomposition, retrain with identical settings, and compare overall accuracy and Kappa; if exact decomposition does not match or beat the reported values, the approximation is not a faithful substitute. A direct check would compute the relative error between eigenvalues and eigenvectors recovered from the five CM-ASQRT iterations and those from exact decomposition on random complex HPD matrices to see whether the iteration converges at all.","tokens_in":18449,"feed_emoji":"🛰️","tokens_out":6750,"duration_ms":47021,"temperature":0.7,"pith_summary":"The paper tries to show that vectorizing PolSAR covariance matrices before feeding them to a neural network destroys exactly the structure that separates terrain classes. It proposes HPDNet, which treats each pixel's 3x3 complex covariance matrix as a Hermitian positive definite matrix on a Riemannian manifold, applies manifold-preserving mapping and rectifying layers, and only then projects to Euclidean space with a complex LogEig layer for a complex-valued 3D CNN. The claim is that this Riemannian-to-Euclidean pipeline beats vector-based complex CNNs and other state-of-the-art PolSAR classifiers, with the largest gains in heterogeneous regions such as buildings and roads. On the Xi'an, Oberpfaffenhofen, and Flevoland datasets, the paper reports overall accuracies of 96.90%, 90.94%, and 98.73%, with Kappa coefficients of 95.56%, 87.30%, and 98.32%, each above the compared methods.","feed_headline":"98.73%: PolSAR classifier keeps covariance matrices on manifold","feed_subtitle":"Keeping radar covariance matrices on the manifold beats vectorizing them on three datasets.","key_machinery":"The load-bearing object is the complex HPD covariance matrix and its Riemannian geometry. The argument rests on three layers: the HPD mapping layer $X_k = W_k X_{k-1} W_k^H$ with a complex kernel, which moves matrices between HPD manifolds; the rectifying layer, which clamps eigenvalues below a threshold while preserving HPD structure; and the complex LogEig layer $X_k = U_{k-1}\\operatorname{diag}(\\log \\Lambda_{k-1}) U_{k-1}^H$, which sends manifold points to the tangent space so a Euclidean complex 3D CNN can follow. Speeding all of this is CM-ASQRT, a complex version of the Newton-Schulz iteration for matrix square roots that replaces costly exact eigendecomposition with GPU-friendly matrix multiplications.","core_discovery":"The central claim is that a complex Hermitian positive definite (HPD) matrix should be learned as a manifold-valued object, not flattened into a vector. The paper defines complex HPD mapping and rectifying layers that operate on the real and imaginary parts of the covariance matrix separately, keeping the output on the HPD manifold, and a complex LogEig layer that maps the manifold to a tangent space. Once in Euclidean space, a complex-valued 3D CNN extracts spatial and scattering context. The authors report that this Riemannian-to-Euclidean pipeline outperforms vector-based complex CNNs and other state-of-the-art PolSAR classifiers, and they attribute the gains to preserving the geometric structure of the covariance matrices rather than to any single Euclidean module.","pith_inferences":["The paper leaves implicit that the same HPD manifold layers could be applied to other matrix-valued remote sensing data, such as coherency matrices from PolInSAR, or to any signal whose natural representation is a complex HPD matrix.","The paper states that five Newton-Schulz iterations give good performance but does not specify how eigenvectors are recovered from the iteration; a careful reader should check whether the approximate square root is differentiable and whether its gradients remain stable during backpropagation.","The training-split experiment (1% to 15%) suggests the method's advantage may be largest in low-label regimes, but the paper does not report per-class errors under extremely scarce training data."],"forward_implications":["Treating PolSAR data as manifold-valued HPD matrices rather than as vectors should improve classification in heterogeneous regions such as buildings and roads, where vectorization loses phase and off-diagonal structure.","The Riemannian-to-Euclidean design means existing Euclidean deep networks can be reused after a LogEig projection, so manifold geometry and contextual features are combined rather than chosen between.","The CM-ASQRT iteration cuts training time on the Xi'an dataset from about 663 seconds to about 152 seconds, which makes the manifold layers practical for larger PolSAR scenes.","The ablation results imply that neither manifold learning alone (80.53% OA on Xi'an) nor a complex 3D CNN alone (93.36%) is sufficient; the reported gains come from the combination of both modules."],"supporting_citations":[{"why":"Establishes that PolSAR covariance matrices are HPD matrices endowed with a Riemannian manifold, motivating manifold processing.","marker":"[21]"},{"why":"SPD network that defines convolution, ReLU, and pooling on manifolds; the proposed HPD mapping and rectifying layers extend its BiMap and ReEig ideas to complex HPD matrices.","marker":"[22]"},{"why":"Newton-Schulz iterative matrix square root method that CM-ASQRT complexifies; supplies the fast eigenvalue approximation for the speed-up claim.","marker":"[51]"},{"why":"Complex-valued 3D CNN used as the Euclidean backend after LogEig projection; supplies the contextual feature extraction module.","marker":"[38]"},{"why":"Baseline CV-CNN that vectorizes the complex matrix; the comparison target that motivates the manifold processing design.","marker":"[17]"},{"why":"PolMPCNN baseline, the strongest comparable method on Flevoland at 98.09% OA, against which the proposed method claims a 0.64% gain.","marker":"[55]"},{"why":"HybridCVnet baseline combining CV-CNN and CV-ViT; the closest competitor on Xi'an at 94.56% OA and on Oberpfaffenhofen at 88.65% OA.","marker":"[56]"}],"fun_headline_variants":["Manifold-based PolSAR classifier outperforms vectorization","Riemannian HPDNet keeps covariance matrices intact","No vectorization: PolSAR classifier on the manifold","Complex HPD matrices processed on manifold for PolSAR","HPDNet: preserving geometry in PolSAR classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the fast five-step approximation of the matrix square root behaving exactly like the true eigenvalue decomposition inside the network; if that approximation distorts the eigenvalues or eigenvectors, the reported accuracy and speed gains cannot be credited to the manifold design.","fun_headline_variants_meta":{"raw":{"variants":["Manifold-based PolSAR classifier outperforms vectorization","Riemannian HPDNet keeps covariance matrices intact","No vectorization: PolSAR classifier on the manifold","Complex HPD matrices processed on manifold for PolSAR","HPDNet: preserving geometry in PolSAR classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1497,"prompt_tokens":924,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":496}},"tokens_in":540,"tokens_out":573,"duration_ms":32372,"temperature":1.0,"reasoning_tokens":496,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:17:59.330558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a PolSAR dataset, replace the CM-ASQRT approximation in the rectifying and LogEig layers with exact complex Hermitian eigendecomposition, retrain with identical settings, and compare overall accuracy and Kappa; if exact decomposition does not match or beat the reported values, the approximation is not a faithful substitute. A direct check would compute the relative error between eigenvalues and eigenvectors recovered from the five CM-ASQRT iterations and those from exact decomposition on random complex HPD matrices to see whether the iteration converges at all.","supporting_citations":[{"cited_title":"Unsupervised classification of polarimetric sar images via riemannian sparse coding,","cited_arxiv_id":null,"evidence_quote":"Establishes that PolSAR covariance matrices are HPD matrices endowed with a Riemannian manifold, motivating manifold processing."},{"cited_title":"A riemannian network for spd matrix learning,","cited_arxiv_id":null,"evidence_quote":"SPD network that defines convolution, ReLU, and pooling on manifolds; the proposed HPD mapping and rectifying layers extend its BiMap and ReEig ideas to complex HPD matrices."},{"cited_title":"Towards faster training of global covariance pooling networks by iterative matrix square root normalization,","cited_arxiv_id":null,"evidence_quote":"Newton-Schulz iterative matrix square root method that CM-ASQRT complexifies; supplies the fast eigenvalue approximation for the speed-up claim."},{"cited_title":"Complex-valued 3- d convolutional neural network for polsar image classification,","cited_arxiv_id":null,"evidence_quote":"Complex-valued 3D CNN used as the Euclidean backend after LogEig projection; supplies the contextual feature extraction module."},{"cited_title":"Polsar image classification via complex-valued multi-scale convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Baseline CV-CNN that vectorizes the complex matrix; the comparison target that motivates the manifold processing design."},{"cited_title":"Polarimetric multipath convolutional neural network for polsar image classification,","cited_arxiv_id":null,"evidence_quote":"PolMPCNN baseline, the strongest comparable method on Flevoland at 98.09% OA, against which the proposed method claims a 0.64% gain."},{"cited_title":"Polsar image classification using a hybrid complex- valued network (hybridcvnet),","cited_arxiv_id":null,"evidence_quote":"HybridCVnet baseline combining CV-CNN and CV-ViT; the closest competitor on Xi'an at 94.56% OA and on Oberpfaffenhofen at 88.65% OA."}],"review_version":1}