Pith. sign in

REVIEW 3 major objections 4 minor 24 references

Adding a local-consistency regularizer to the TIM++ objective yields state-of-the-art few-shot remote sensing scene classification across ten datasets and two vision-language backbones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 11:53 UTC pith:2CUZNWVW

load-bearing objection Useful empirical extension with a valuable open benchmark, but the closed-form q-update is asserted as KKT without stating the objective—treat the theoretical framing as heuristic until that is fixed. the 3 major comments →

arxiv 2607.29192 v1 pith:2CUZNWVW submitted 2026-07-31 cs.CV

Locally Consistent Transductive Information Maximization for Few-Shot Remote Sensing Scene Classification

Karim El Khoury , Beno\^it G\'erin , Beno\^it Macq , Christophe De Vleeschouwer This is my paper
classification cs.CV
keywords transductive few-shot learningremote sensing scene classificationlocal consistencyinformation maximizationvision-language modelsnearest-neighbor affinity graphmulti-source fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Few-shot remote sensing scene classification can be improved by a simple, nearly-free regularizer: after each alternate update, each query's prediction is pulled toward the average prediction of its k nearest feature-space neighbors. This paper introduces LC-TIM, which adds this local-consistency term to the TIM++ transductive objective, whose q-update becomes a single multiplicative factor involving the neighborhood-averaged predictions. On ten remote sensing datasets and with both CLIP and GeoRSCLIP backbones, LC-TIM outperforms TIM++ and other transductive baselines at every shot setting, with the largest gains in the low-shot regime (e.g., 90.3 vs 87.7 average accuracy at 2 shots with GeoRSCLIP). A multi-source extension fuses the affinity graphs of two complementary encoders (the VLM's CLS token and DINOv3 patch embeddings), further boosting accuracy without adding tunable parameters.

Core claim

The paper's central claim is that enforcing prediction agreement within local feature-space neighborhoods improves transductive few-shot scene classification for remote sensing. Specifically, the objective augments TIM++ with L_LC(p) = (1/|Q|) Σ_i D_KL(p_i || pbar_i), where pbar_i is the mean soft prediction of query i's k nearest neighbors. Solving the constrained KKT conditions yields the closed-form q-update q_ik ∝ (p_ik)^(1+α) * yhat_ik^γ * (pbar_ik)^(λ_LC), which adds a single multiplicative factor to TIM++'s update. The multi-source variant redefines the neighbor graph using the product of min-max normalized cosine similarities from two encoders, so a pair is a neighbor only if both so

What carries the argument

The central mechanism is the local-consistency regularizer L_LC and its closed-form influence on the q-update. For each query sample, the neighborhood-averaged prediction pbar_i is computed from the k nearest query features under a cosine-similarity affinity; the regularizer penalizes the KL divergence between each query's distribution and this average. In the alternating-direction solver, the regularizer manifests as the multiplicative factor (pbar_ik)^(λ_LC) in the q-update, so enforcing local agreement adds negligible computational cost. The multi-source extension changes only the neighbor definition: affinities from two encoders are min-max normalized and multiplied (a_ij = s^v_ij · s^g_

Load-bearing premise

The method assumes that each query sample's k nearest neighbors in the frozen vision-language feature space are likely to share its true class, so averaging their soft predictions is a trustworthy target; if the feature space is not locally class-consistent — for fine-grained, spectrally atypical, or mixed-patch scenes — the regularizer can reinforce incorrect neighborhoods and hurt accuracy.

What would settle it

Conduct an ablation where the kNN graph is replaced with random query-selected neighbors (or neighbors from a deliberately different class) while keeping all other components identical, and measure accuracy on the ten datasets; if LC-TIM does not degrade toward TIM++'s level, the regularizer's gains are not attributable to local consistency.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Transductive processing of the unlabeled query set yields large gains over zero-shot inference: all evaluated transductive methods beat the zero-shot baseline at every shot setting, with at least +17.8% average accuracy at 1 shot on GeoRSCLIP.
  • LC-TIM achieves the best average accuracy across ten remote sensing datasets at every shot count on both CLIP and GeoRSCLIP, e.g., 90.3 vs 87.7 at 2 shots and 94.2 vs 92.9 at 16 shots with GeoRSCLIP.
  • The regularizer adds minimal runtime (0.32s vs 0.22s for TIM++ on EuroSAT with 8,100 queries), making it suitable for batch-oriented operational pipelines.
  • Fusing affinity graphs from GeoRSCLIP and DINOv3 patch embeddings improves accuracy further, e.g., 91.0 vs 90.3 at 2 shots with GeoRSCLIP, without new hyperparameters.
  • Ablations show low sensitivity to the neighborhood size k (1,3,5,10) and the regularizer weight λ_LC, indicating the method is robust to hyperparameter choices near its operating point.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same local-consistency trick is likely to transfer to other transductive few-shot settings beyond remote sensing — natural images, medical imaging, or aerial video — whenever the frozen feature space is locally smooth; the paper itself notes the core formulation is domain-agnostic.
  • Because the gain is largest when query sets are large and densely packed (e.g., RSICB128 with 9,147 queries), the method's benefit may shrink on small query batches; a stress test with downsampled query sets would quantify this.
  • The multiplicative form suggests an interpretation as a temperature-scaled re-weighting of assignments toward neighborhood consensus; iterating with an annealing schedule for λ_LC could potentially squeeze out additional gains.
  • The fused-affinity product could be extended to more than two encoders, with adaptive weights that down-weight an encoder when its affinities are noisy on a per-dataset basis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes LC-TIM, a transductive few-shot remote sensing scene classification method that extends the TIM++ objective with a local consistency regularizer. The regularizer encourages each query sample's predicted distribution to agree with the average prediction of its κ nearest feature-space neighbors, using a kNN graph built from frozen VLM embeddings. The claimed contribution is that this regularizer enters as a single multiplicative factor in the closed-form q-update, adding negligible overhead. The paper also introduces a multi-source variant that fuses GeoRSCLIP and DINOv3 affinities for the neighborhood graph. The authors establish a benchmark of ten RS datasets and evaluate CLIP ViT-B/32 and GeoRSCLIP ViT-B/32 backbones at 0,1,2,4,8,16 shots, comparing against LP++, TransCLIP, TIM++, and zero-shot baselines. They report consistent average improvements over TIM++ (e.g., 90.3 vs 87.7 at 2-shot with GeoRSCLIP; 94.2 vs 92.9 at 16-shot) and additional gains from the DINOv3 extension, with ablation studies on κ, λ_LC, and graph construction.

Significance. If the empirical results are robust, the paper makes a useful practical contribution: a simple, computationally cheap modification of a strong transductive baseline that consistently improves few-shot RS classification across a broad benchmark. The proposed benchmark is itself a contribution, as transductive few-shot evaluation on RS VLMs is underexplored. The paper's methodology has notable strengths: hyperparameters are fixed across all datasets and shot settings, the code is publicly available, ablations probe the key design choices, and the runtime comparison shows negligible overhead. The central theoretical claim, however, is not fully supported: the closed-form q-update is stated as the KKT solution of an unreported 'reformulated objective,' and the natural derivation from Eq. (8) leads to different update terms. This gap weakens the 'principled extension' narrative, even if the heuristic update may still be effective. The absence of variance/statistical significance reporting also makes it difficult to assess the robustness of the reported average gains.

major comments (3)
  1. [§3.5, Eq. (10)] The q-update is asserted to be the KKT solution of a 'reformulated objective,' but that objective is never stated. Solving a natural TIM++-style relaxation of Eq. (8) with pbar_i fixed from the previous iterate gives a stationarity condition containing a class-marginal term qbar_k and rescaled exponents, not Eq. (10). If pbar_i is instead expressed in terms of q_j, the neighborhood graph couples all q_i. Thus Eq. (10) is not the exact KKT solution of any clearly stated objective; it is at best a heuristic multiplicative update. Since the paper's theoretical narrative is that adding L_LC to TIM++ yields a principled objective, this derivation gap is load-bearing and must be fixed.
  2. [Tables 1–3] The tables report only mean top-1 accuracy over 10 seeds, with no standard deviations, confidence intervals, or paired significance tests. Several average improvements over TIM++ are under 1 percentage point at higher shot settings (e.g., Table 1, 16-shot: 90.7 vs 89.0; per-dataset differences are often ~0.5 pp), and there are per-dataset degradations (e.g., Table 1, 1-shot RSC11: LC-TIM 72.8 vs TIM++ 74.4; Table 2, 1-shot RSC11: 85.9 vs 86.7). Without variance or significance testing, the claim that LC-TIM 'consistently surpasses' TIM++ is not supported. Please report error bars and pairwise tests, or temper the claim.
  3. [§4.4, Table 4(b)] The ablation for λ_LC shows sensitivity at the high end (λ_LC=0.5 degrades AID, EuroSAT, MLRSNet relative to 0.3), and the authors acknowledge dataset-specific tuning could improve results. While the parameter is fixed across all experiments, this raises a concern that the chosen λ_LC=0.3 may not be universally safe, especially on datasets or feature sources where the manifold-smoothness premise is weaker. The paper should discuss this limitation explicitly and perhaps evaluate the variance of gains across datasets where the premise may fail (e.g., fine-grained classes).
minor comments (4)
  1. [Tables 1–2] The table headers contain typos ('EuroSA T') and the bold/underline formatting is not visually distinct in the supplied text. Please ensure the final PDF renders bold and underlined entries clearly.
  2. [§3.5] The phrase 'reformulated objective' is undefined. If Eq. (10) is indeed an approximation, state this explicitly and provide the surrogate objective that is actually solved.
  3. [§3.2, Eq. (4)] The mutual information term is written as İ_α(X_Q;Y_Q) but the definition in the text uses İ_α = αH(Y)−H(Y|X). Please make the notation consistent or clarify whether İ_α is a scaled estimator.
  4. [§4.3, Table 3] The multi-source extension table reports only LC-TIM vs LC-TIM+DINO. It would strengthen the comparison to also include standalone DINOv3-based kNN in the ablation (partially covered in Table 4(c)), but the current single comparison is sufficient for the main claim.

Circularity Check

0 steps flagged

No circular derivation; empirical claims are self-contained, with an unverified KKT update and non-load-bearing self-citations.

full rationale

I find no circular step. LC-TIM's contribution is adding L_LC (Eq. 7) to the TIM++ objective (Eq. 8); the resulting q-update (Eq. 10) is a fixed-point iteration using the previous iterate p^(t), the fixed zero-shot prior yhat, and the neighborhood average pbar^(t), none of which is fitted to the reported accuracies. The accuracy claims are tested against external baselines (LP++, TransCLIP, TIM++, CLIP/GeoRSCLIP zero-shot) on ten public datasets with hyperparameters fixed across datasets (Sec. 4.1), so the improvements are externally checkable rather than encoded by construction. I explicitly flag one missing proof: Sec. 3.5 says Eq. (10) follows by 'solving the Karush–Kuhn–Tucker (KKT) conditions of the reformulated objective,' but the reformulated objective is never stated; this makes the closed-form update unverifiable as an exact KKT solution. That is a rigor/correctness gap, not a circular reduction, and it does not make the experimental comparison circular. The self-citations [4,5] supply dataset splits and an earlier RS-VLM benchmark, but they are not load-bearing for the algorithm's derivation or for the reported improvements over the external TIM++ baseline. This is an honest non-finding of circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The claims rest on standard transductive few-shot assumptions plus a new manifold-smoothness premise. There are two hand-set hyperparameters (κ, λ_LC) and an unstated reformulated objective behind Eq. (10), but no new physical or conceptual entities are postulated.

free parameters (3)
  • κ (neighborhood size) = 5
    Chosen by hand and fixed across all datasets and shot settings. Ablation (Table 4a) shows mild sensitivity; no validation-selection protocol is described.
  • λ_LC (local consistency strength) = 0.3
    Chosen by hand and fixed. Ablation (Table 4b) shows it matters: λ_LC=0 reduces to TIM++ and drops performance. The selection basis is not reported.
  • TIM++ inherited hyperparameters (τ, λ, α, γ, T) = τ=120, λ=0.4, α=1.0, γ=0.05, T=150
    Taken unchanged from TIM++ [9]; the LC-TIM result depends on these values just as much as on κ and λ_LC.
axioms (5)
  • domain assumption Query set is closed-set and drawn from the same K classes as the support set
    Sec. 3.1 states Q is drawn from the same K classes; all transductive methods rely on this, and violations would break the mutual-information terms.
  • domain assumption Feature-space manifold smoothness: κ nearest neighbors of a query tend to share its class
    Sec. 3.3 motivation: 'VLM encoders produce tightly clustered representations... nearest neighbors very likely to share its true class.' This is the load-bearing premise for the regularizer.
  • domain assumption Frozen VLM features and zero-shot text priors are informative for RS scene classification
    The whole pipeline builds on CLIP/GeoRSCLIP features and text prompts; if the foundation model is unreliable, all evaluated methods fail.
  • domain assumption DINOv3 mean patch embeddings capture complementary structural cues and the product affinity is meaningful
    Sec. 3.4 assumes the product of normalized affinities acts as a soft AND, which is a modeling choice rather than a proven fact.
  • ad hoc to paper ADM with T=150 iterations and the stated q-update solves Eq. (8)
    The reformulated objective is not given, so the KKT solution is asserted rather than demonstrated; this is a core algorithmic assumption.

pith-pipeline@v1.3.0-daily-deepseek · 11921 in / 16132 out tokens · 166504 ms · 2026-08-03T11:53:43.472038+00:00 · methodology

0 comments
read the original abstract

Remote sensing scene classification is increasingly relying on foundation models pre-trained on large-scale Earth-observation data. Moreover, transductive inference, which exploits the collective statistical structure of the entire unlabeled query set, appears to naturally match remote sensing pipelines where large images are routinely split into patches and inferred as a batch. In this work, we introduce LC-TIM (Locally Consistent Transductive Information Maximization), which extends the state-of-the-art Transductive Information Maximization for Few-Shot CLIP (TIM++) objective with a local consistency regularizer that enforces prediction agreement between each query sample and its $\kappa$ nearest feature-space neighbors. The regularizer enters as a single multiplicative factor in the closed-form $q$-update, adding negligible computational overhead. We further propose a multi-source extension that fuses the affinity graph from multiple remote sensing foundation model, further boosting classification accuracy. To assess these methods, we establish the first comprehensive, open-source benchmark for transductive few-shot RS scene classification, evaluating LP++, TransCLIP, TIM++, and LC-TIM across ten diverse datasets, two remote sensing vision-language models, and across various few-shot settings. Our experiments show that transductive methods consistently outperform zero-shot baselines, and that LC-TIM achieves state-of-the-art accuracy, with the largest gains in the low-shot regime where neighborhood cues are most informative. Code is publicly available at: https://github.com/elkhouryk/LC-TIM

Figures

Figures reproduced from arXiv: 2607.29192 by Beno\^it G\'erin, Beno\^it Macq, Christophe De Vleeschouwer, Karim El Khoury.

Figure 1
Figure 1. Figure 1: Qualitative neighborhoods comparison between GeoRSCLIP, DINOv3 mean patch embedding and our proposed fused approach on 3 queries from OPTIMAL31. For each query, the 3 rows show the κ = 5 closest neighbors according to each kNN graph N . LC-TIM and LC-TIM+DINO use a single shot for their prediction. Green outline indicates that the neighbor has the same ground truth label as the query image [PITH_FULL_IMAG… view at source ↗

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

24 extracted references · 1 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems33, 2445–2457 (2020)

    Boudiaf, M., Ziko, I., Rony, J., Dolz, J., Piantanida, P., Ben Ayed, I.: Information maximization for few-shot learning. Advances in Neural Information Processing Systems33, 2445–2457 (2020)

    2020

  2. [2]

    Proceedings of the IEEE105(10), 1865–1883 (2017)

    Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE105(10), 1865–1883 (2017)

    2017

  3. [3]

    El Khoury, K., Godelaine, T., Delvaux, S., Lugan, S., Macq, B.: Streamlined hybrid annotation framework using scalable codestream for bandwidth-restricted uav ob- jectdetection.In:2024IEEEInternationalConferenceonImageProcessing(ICIP). pp. 1581–1587 (2024).https://doi.org/10.1109/ICIP51287.2024.10647448

    arXiv 2024

  4. [4]

    In: IGARSS 2026-2026 IEEE International Geoscience and Remote Sensing Symposium

    El Khoury, K., Zanella, M., De Vleeschouwer, C., Macq, B.: Few-shot adapta- tion benchmark for remote sensing vision-language models. In: IGARSS 2026-2026 IEEE International Geoscience and Remote Sensing Symposium. pp. 1–5. IEEE (2026)

    2026

  5. [5]

    In: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    El Khoury, K., Zanella, M., Gérin, B., Godelaine, T., Macq, B., Mahmoudi, S., De Vleeschouwer, C., Ayed, I.B.: Enhancing remote sensing vision-language mod- els for zero-shot scene classification. In: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2025)

    2025

  6. [6]

    Helber, P., Bischke, B., Dengel, A., Borth, D.: Introducing eurosat: A novel dataset anddeeplearningbenchmarkforlanduseandlandcoverclassification.In:IGARSS 2018-2018 IEEE international geoscience and remote sensing symposium. pp. 204–

    2018

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, Y., Shakeri, F., Dolz, J., Boudiaf, M., Bahig, H., Ben Ayed, I.: Lp++: A surprisingly strong linear probe for few-shot clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23773–23782 (2024)

    2024

  8. [8]

    Sensors20(6), 1594 (2020)

    Li, H., Dou, X., Tao, C., Wu, Z., Chen, J., Peng, J., Deng, M., Zhao, L.: Rsi-cb: A large-scale remote sensing image classification benchmark using crowdsourced data. Sensors20(6), 1594 (2020)

    2020

  9. [9]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Li, Y., Zou, Y., Huang, Y., Jiao, C., Wang, X., Peng, S., Guo, Z., Gou, S.: Tim++: Transductive information maximization for few-shot clip. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 6671–6680 (2026)

    2026

  10. [10]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    Liu, F., Chen, D., Guan, Z., Zhou, X., Zhu, J., Ye, Q., Fu, L., Zhou, J.: Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    2024

  11. [11]

    Ieee Access11, 127057– 127076 (2023)

    Phang,S.K.,Chiang,T.H.A.,Happonen,A.,Chang,M.M.L.:Fromsatellitetouav- based remote sensing: A review on precision agriculture. Ieee Access11, 127057– 127076 (2023)

    2023

  12. [12]

    ISPRS Journal of Photogrammetry and Remote Sensing169, 337–350 (2020)

    Qi,X.,Zhu,P.,Wang,Y.,Zhang,L.,Peng,J.,Wu,M.,Chen,J.,Zhao,X.,Zang,N., Mathiopoulos, P.T.: Mlrsnet: A multi-label high spatial resolution remote sensing dataset for semantic scene understanding. ISPRS Journal of Photogrammetry and Remote Sensing169, 337–350 (2020)

    2020

  13. [13]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  14. [14]

    arXiv preprint arXiv:2508.10104 (2025) LC-TIM for RS Few-Shot Scene Classification 15

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025) LC-TIM for RS Few-Shot Scene Classification 15

    Pith/arXiv arXiv 2025

  15. [15]

    IEEE transactions on neural networks10(5), 988–999 (1999)

    Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks10(5), 988–999 (1999)

    1999

  16. [16]

    IEEE Transactions on Geoscience and Remote Sensing57(2), 1155–1167 (2018)

    Wang, Q., Liu, S., Chanussot, J., Li, X.: Scene classification with recurrent atten- tion of vhr remote sensing images. IEEE Transactions on Geoscience and Remote Sensing57(2), 1155–1167 (2018)

    2018

  17. [17]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5805–5813 (2024)

    2024

  18. [18]

    IEEE Transactions on Geoscience and Remote Sensing55(7), 3965–3981 (2017)

    Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing55(7), 3965–3981 (2017)

    2017

  19. [19]

    In: ISPRS TC VII Symposium-100 Years ISPRS

    Xia, G.S., Yang, W., Delon, J., Gousseau, Y., Sun, H., Maître, H.: Structural high-resolution satellite image indexing. In: ISPRS TC VII Symposium-100 Years ISPRS. vol. 38, pp. 298–303 (2010)

    2010

  20. [20]

    Remote sensing of Environment241, 111716 (2020)

    Yuan, Q., Shen, H., Li, T., Li, Z., Li, S., Jiang, Y., Xu, H., Tan, W., Yang, Q., Wang, J., et al.: Deep learning in environmental remote sensing: Achievements and challenges. Remote sensing of Environment241, 111716 (2020)

    2020

  21. [21]

    Advances in Neural Information Processing Systems37, 62223–62256 (2024)

    Zanella, M., Gérin, B., Ayed, I.B.: Boosting vision-language models with transduc- tion. Advances in Neural Information Processing Systems37, 62223–62256 (2024)

    2024

  22. [22]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

    Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m and georsclip: A large-scale vision- language dataset and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

    2024

  23. [23]

    Journal of Applied Remote Sensing10(3), 035004–035004 (2016)

    Zhao, L., Tang, P., Huo, L.: Feature significance-based multibag-of-visual-words model for remote sensing image scene classification. Journal of Applied Remote Sensing10(3), 035004–035004 (2016)

    2016

  24. [24]

    ISPRS journal of pho- togrammetry and remote sensing145, 197–209 (2018)

    Zhou, W., Newsam, S., Li, C., Shao, Z.: Patternnet: A benchmark dataset for performance evaluation of remote sensing image retrieval. ISPRS journal of pho- togrammetry and remote sensing145, 197–209 (2018)

    2018