Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Dynamic mapping from static labels: remote sensing dynamic sample generation with temporal-spectral embedding

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TasGen converts a single dated land-cover label into a full time-series of training labels by treating land-cover change as an anomaly in learned temporal-spectral patterns, reporting 91.64% average mapping accuracy across six regions.

desk verdict Worth a serious look because of the problem and the experiments, but the central VAE derivation and the Gibbs stopping rule are wrong as printed, so the theoretical support collapses until fixed. read the letter →

arxiv 2506.02574 v2 pith:5AWF6FXZ submitted 2025-06-03 eess.IV cs.CVcs.MM

Shuai Yuan , Shuang Chen , Tianwu Lin , Jincheng Yuan , Geng Tian , Yang Xu , Jie Wang , Peng Gong This is my paper
classification eess.IVcs.CVcs.MM
keywords dynamicmappingautomaticsamplegenerationtemporal-spectralembeddinghierarchicalvariationalautoencoderanomalydetectionremotesensingwetlandlandcoverchange
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Remote sensing maps go stale quickly: a training sample labeled from one date's imagery can misrepresent the same location months later, and re-labeling by hand across time is too labor-intensive to sustain. TasGen claims to break that bottleneck by turning each single-date static label into a full sequence of dynamic labels automatically. The method treats land-surface change as an anomaly in temporal-spectral sequences: it learns what "normal" looks like for a sample with a hierarchical temporal-spectral variational autoencoder, flags time steps that deviate from that normal pattern, and then uses a classifier trained on the stable time steps to relabel the anomalous ones. Across six study regions spanning land cover and wetland mapping, the relabeled samples support mapping with average overall accuracy 91.64% and Kappa 0.9071, exceeding the strongest compared methods by up to 7.97%. If the claim holds, the central practical consequence is that large-area dynamic mapping no longer requires repeated human annotation; one labeling pass plus learned temporal-spectral structure supplies the rest.

What carries the argument

The load-bearing object is the HTS-VAE (hierarchical temporal-spectral variational autoencoder) with a dual-dimension embedding, together with the Gibbs-sampling anomaly attribution and the relabeling classifier that ride on its latent codes. The encoder compresses the input along the temporal axis with stacked 1D convolutions, reconstructs a denoised curated signal, then encodes the spectral axis with a bidirectional GRU; the two stochastic latent variables $e_t$ (temporal) and $e_s$ (spectral) are organized hierarchically so that $e_t$ acts as a contextual prior for $e_s$, and a Glow normalizing flow makes the spectral posterior flexible enough to capture multimodal real-world distributions. This design simultaneously supplies the anomaly score (negative reconstruction log-likelihood), the features for relabeling (fused $e_t$ and $e_s$), and the target for Gibbs-based attribution (imputing one band-timepoint at a time and measuring reconstruction-score improvement). The paper's operating principle is the identity that class change implies detectable temporal-spectral deviation: "If there is a type change, temporal-spectral changes must occur."

What would settle it

Run TasGen on time-series samples whose class is independently verified to never change across the full four-year window, such as permanent deep water or stable closed forest confirmed by high-resolution imagery at multiple dates, and measure how many stable time steps are flagged as anomalies and relabeled; if the false-relabeling rate is not near zero, the core identity between reconstruction deviation and class change is not holding. A complementary check is to compare TasGen's relabeled transition times against independently recorded change dates in a region with known history, and to see whether the reported anomaly-detection F1 gap widens when seasonal variation is isolated from real transitions.

Watch

Extended reading notes

Core claim

The paper's central claim is that land-surface dynamics can be captured by a reconstruction-anomaly principle: if a location's land-cover or wetland type changes, its temporal-spectral signature must deviate from the pattern the location normally follows, so those deviations can be located as anomalies in a learned latent space and then explained and labeled. TasGen operationalizes this in two stages. First, a hierarchical temporal-spectral variational autoencoder (HTS-VAE) compresses each time-series sample through a dual-dimension embedding, temporal compression along the observation axis and a spectral-wise recurrent encoding along the band axis, with the temporal embedding acting as context for the spectral one, and learns the low-dimensional normal pattern of stable samples; reconstruction likelihood then serves as the anomaly score. Second, a lightweight classifier is trained on the assumed-stable time steps using the fused temporal-spectral embeddings as features, and is applied only to time steps flagged as significant anomalies, propagating the single known label across the whole observation window. A Gibbs-sampling attribution step imputes suspected anomalous bands and time points one dimension at a time to identify which spectral-temporal components drove each change. The paper reports that this pipeline yields anomaly-detection F1 above 92%, relabeling F1 above 92% (96% or higher in land cover regions), and final mapping accuracy that beats CCDC, LandTrendr, TSSCD, and static-sample training in almost every region, with the largest margins in seasonally dynamic wetlands.

Load-bearing premise

The load-bearing premise is that statistically significant deviations from a sample's learned temporal-spectral pattern are caused by genuine land-cover or wetland type changes, so a classifier trained on the presumed-stable time steps can correctly relabel the flagged ones; if seasonal phenology, clouds, or sensor noise masquerade as anomalies, or true slow changes stay below the detection threshold, the self-generated labels quietly corrupt the training pool.

Editorial extensions

If this is right

  • Static samples labeled for a single date can be propagated into dynamic training pools covering every observation in the window, removing the repeated manual-annotation cost that limits time-series mapping.
  • Because change detection and label assignment are separated, the first stage can deliberately favor high recall and the second-stage relabeling can absorb its false positives, giving the pipeline more fault tolerance than end-to-end change detectors.
  • Reliable dynamic labels can be produced in both abrupt-change regimes (urban expansion in Shenzhen and Cairo) and cyclical regimes (wetland water-marsh-mudflat transitions), with final mapping overall accuracy above 89% in all six regions.
  • Anomaly attribution identifies which spectral band and which time point drive each transition, making the generated samples explainable rather than a black box.
  • Ablation results indicate that the disentangle-then-couple design carries the gain: removing either the dual-dimension embedding or the HTS-VAE lowers anomaly-detection F1 by several points in every region.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The detect-then-relabel recipe is generic: any single-date-labeled time-series classification problem whose class changes disturb spectral-temporal regularities, such as forest disturbance, crop-type rotation, or burned-area recovery, could reuse the same pipeline without modification.
  • The experiments use cloud-free, analysis-ready synthetic composites; on raw satellite archives with cloud gaps and variable viewing geometry the anomaly prior will be noisier, so the practical ceiling depends on how well curation plus reconstruction filters that noise, and the missing-observation tests are not a substitute for real cloudy-season validation.
  • The Gibbs attribution scores could be reused directly as a change-driver diagnostic, identifying which bands and seasons carry wetland loss or urban gain, without fitting any additional model, since the scores are already dimension-wise.
  • The sharpest unresolved boundary is intra-class phenology versus class change: explicitly modeling a seasonal baseline as normal would tighten the anomaly threshold, and is a natural next test, since seasonal wetlands are exactly where the reported F1 is lowest.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper addresses the problem of updating training samples for dynamic remote sensing mapping when only single-date static labels are available. It proposes TasGen, a two-stage pipeline: first, a hierarchical temporal-spectral variational autoencoder (HTS-VAE) with a dual-dimension embedding is trained to reconstruct time-series windows, and reconstruction likelihood is used as an anomaly score; second, a classifier trained on time steps considered stable relabels anomalous time steps, producing dynamic training samples. A Gibbs-sampling procedure is proposed to attribute anomalies to spectral bands and time points. The method is evaluated on land cover and wetland mapping in six regions using the SDC30 dataset, reporting an average mapping OA of 91.64% and outperforming CCDC, LandTrendr, TSSCD, and static-sample baselines. The paper also includes ablation studies and robustness tests to reduced observation frequency and missing data.

Significance. If the proposed method works as described, it would reduce the manual labeling burden for time-series remote sensing mapping and provide an interpretable intermediate sample-level representation. The paper's strengths are its extensive evaluation across six geographically diverse regions, independent manual labels at every timestamp for validation, comparison with several established baselines, and systematic ablation and robustness experiments. However, the central theoretical derivation of the HTS-VAE objective is not valid as written, and the Gibbs attribution algorithm contains an inverted stopping rule. These issues directly affect the anomaly detection and relabeling stages that support the paper's main empirical claims, so the current manuscript does not provide a sound basis for the reported results.

major comments (4)
  1. [Section 2.4, Eqs. (8)-(11)] The derivation of the HTS-VAE training objective is not valid as written. The posterior factorization in Eq. (6) omits q_phi(e_t | x) and is dimensionally inconsistent, and Eq. (9) introduces an undefined variable z while also containing the typo p_theta(e_s, e | e_t). More importantly, the claim that sharing Deconv1D parameters makes q(d_{1:T} | e_t) = delta(d - D(e_t)) equal to p(z_{1:T} | e_s) = delta(z - D(e_s)) does not follow: equality of the two delta distributions would require D(e_t) = D(e_s) for every input, which is not established and is generically false. The subsequent identity q(d_{1:T} | e_t) = p(e_{1:W} | z_2) in Eq. (11) uses symbols that are not defined anywhere in the paper. Consequently, Eq. (8) is not a valid ELBO for the generative model in Eq. (5), and the claim that the model learns normal temporal-spectral patterns by variational inference is unsupported. Since the same embeddings feed the relabeling classifier, this affects both stages of TasGen. The authors should either provide a correct derivation of the actual objective or explicitly state that the training objective is a heuristic surrogate rather than a variational bound.
  2. [Algorithm 1 and Section 2.6] The stopping condition of the Gibbs attribution algorithm is inverted relative to the definition of the anomaly score. Equation (13) defines S as a negative log-likelihood, so a higher score means lower reconstruction likelihood. The text correctly states that after imputation the model should have recovered a latent embedding corresponding to normal conditions, which should make S_r smaller, not larger, than the baseline b. However, Algorithm 1 (line 12) breaks when mean(S_r) >= b, and Section 2.6 says the iteration proceeds until the updated reconstruction score 'exceeds' the baseline b. As printed, the algorithm stops exactly when the sample remains as anomalous as or more anomalous than baseline, so the attribution score AS = S_0 - S_r and the top-10% significance filter in Section 2.7 are not computed as intended. This should be corrected to mean(S_r) <= b (with a tolerance) and the surrounding text updated accordingly.
  3. [Sections 2.3-2.4 and 4.1] The construction of the training set for the unsupervised HTS-VAE is underspecified. The paper repeatedly refers to learning 'normal patterns,' but it never states whether the VAE is trained on all time windows, only on windows centered on the known static label, or only on windows that are themselves classified as stable by some prior step. If anomalous windows are included in training, the reconstruction baseline b and the anomaly score S are biased, and the self-training loop in Section 2.7 can propagate errors. Section 4.1 gives only optimizer and architecture hyperparameters; the data split used for Stage-1 training, the number of windows per sample, and the precise definition of the baseline b (the 'N sliding windows from the training set') are not reported. This information is necessary to reproduce the method and to determine whether the anomaly detector is genuinely unsupervised or implicitly supervised by the static labels.
  4. [Section 2.7 and Tables 3-4] The main empirical claim relies on a self-training loop whose failure modes are not isolated. The relabeling classifier is trained on time steps that the same anomaly detector labels as stable, and then it is used to relabel the same detector's anomalies; the final OA/Kappa are therefore not an independent test of the anomaly detector's correctness. The authors do evaluate F1-A against independent manual labels, which is good, but they do not report how classification accuracy changes as a function of anomaly-detection precision/recall, or compare against a version trained on oracle-stable time steps. I recommend adding such an analysis; otherwise the reported improvement over baselines could partly reflect error compensation between the two stages rather than the proposed representation learning.
minor comments (5)
  1. [Section 6 and Abstract] The conclusion reports 'average OA of 95.08%, 95.01%, and 91.64% in anomaly detection, dynamic sample generation, and dynamic mapping,' but the first two quantities are F1 scores (F1-A and F1-S), not OA; the terminology should be corrected.
  2. [Section 5.2, Table 4] The text says 'the rule-based interpolation method performs worst,' but Table 4 does not include this baseline; either add it to the table or remove the claim.
  3. [Table 1, NC row] In the Newcastle row, TasGen's OA (91.11 +/- 0.61) is lower than TSSCD's (92.17 +/- 0.39), so the unqualified 'outperforming' language in the Abstract should be qualified; the body text does say 'almost regions,' which is accurate but should be reflected in the abstract.
  4. [Section 2.6 and Algorithm 1] The scalar anomaly score S in Eq. (13) is not defined per band and time point, but the baseline b and the attribution score use S0_{c,t}; the paper should specify how S is decomposed into a C x T score map.
  5. [Algorithm 1, line 6] The notation 'x_n union x_a \ (c,t)' is ambiguous; it should be written as (x_n union x_a) \ {(c,t)} to make clear that only the selected dimension is removed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: dynamic-sample relabeling is externally validated against per-timestamp manual labels; self-citations are contextual, not load-bearing.

full rationale

The central derivation chain is empirically self-contained. Stage 1 learns an unsupervised reconstruction-based anomaly score from the HTS-VAE; Stage 2 trains a classifier on time steps assumed stable and relabels anomalous steps, and both anomaly detection and relabeling are then scored against independently labeled per-timestamp manual labels. The final mapping accuracy is also compared with static samples and with external baselines (CCDC, LandTrendr, TSSCD) using held-out reference labels. No fitted parameter is renamed as a prediction: the static label y0 is an input, but the relabeled dynamic labels at other timestamps are not defined to equal y0, and they are verified against manual labels at every time point. The self-citations are not load-bearing. The citation to Gong et al. (2024) for 'stable classification theory' appears only in the conclusion and is immediately followed by 'which is also proven in our experiments,' so the evidence for the claim is the paper's own independent evaluation rather than a self-citation chain. Other self-citations (e.g., Yuan et al. 2025) are background context on wetland dynamics. There is no imported uniqueness theorem and no ansatz smuggled in via citation. The most serious weakness in the derivation is the Section 2.4 assertion that sharing Deconv1D parameters makes the two delta distributions equal, i.e., q(d1:T | et) = p(z1:T | es), which is not generally true and makes the stated ELBO cancelation invalid; Eq. (11) also uses undefined symbols, and Algorithm 1's stopping condition appears reversed relative to Eq. (13). These are formal correctness and reproducibility concerns that undermine the stated variational bound, but they do not reduce the empirical mapping result to its own inputs by construction. Under the circularity-specific rubric, the appropriate finding is no significant circularity, with a small score adjustment only for the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters and domain assumptions. No new physical entities are introduced. The most fragile elements are the anomaly-to-change correspondence and the unsupported ELBO cancellation.

free parameters (5)
  • anomaly detection baseline score b = computed as average reconstruction score over training windows
    Defined in Section 2.6 as b = (1/N) sum of S0 over training windows; determines which time steps are flagged and relabeled.
  • top 10% attribution significance threshold = 10%
    Section 2.7: 'significant if ... falls within the top 10% of all attribution scores'; this hand-set threshold filters which anomalies drive relabeling.
  • sliding window length = 30
    Section 4.1: sliding window length set to 30; controls temporal context.
  • Monte Carlo sample count = 40
    Section 4.1: 'The number of samples is set to 40 for Monte Carlo integration'.
  • Gibbs iteration count = 10
    Section 4.1: 'the imputation time is set to 10 for Gibbs iteration'.
assumptions (3)
  • domain assumption Land surface type changes manifest as statistically significant anomalies in the temporal-spectral reflectance sequence.
    Section 2.1; the entire anomaly-based relabeling pipeline depends on this mapping.
  • domain assumption The reconstruction probability of the HTS-VAE is a valid proxy for normality of a time point.
    Section 2.5; anomaly score S is negative log-likelihood; the thresholding assumes this score separates true changes from natural variation.
  • ad hoc to paper The delta-distribution cancellation q(d1:T | et) = p(z1:T | es) is valid when decoder and encoder Deconv1D layers share parameters.
    Section 2.4, Eq. (11) and surrounding text; the cancellation is asserted without proof and appears to equate distributions with different conditioning variables.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic mapping from static labels: remote sensing dynamic sample generation with temporal-spectral embedding." pith.science (2026). https://pith.science/paper/5AWF6FXZ

@misc{pith2026250602574,
  author       = {Pith},
  title        = {Pith review of: Dynamic mapping from static labels: remote sensing dynamic sample generation with temporal-spectral embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5AWF6FXZ}},
  note         = {Machine review of arXiv:2506.02574}
}
read the original abstract

Accurate remote sensing geographic mapping requires timely and representative samples. However, rapid land surface changes often render static samples obsolete within months, making manual sample updates labor-intensive and unsustainable. To address this challenge, we propose TasGen, a two-stage Temporal spectral-aware Automatic Sample Generation method for generating dynamic training samples from single-date static labels without human intervention. Land surface dynamics often manifest as anomalies in temporal-spectral sequences. %These anomalies are multivariate yet unified: temporal, spectral, or joint anomalies stem from different mechanisms and cannot be naively coupled, as this may obscure the nature of changes. Yet, any land surface state corresponds to a coherent temporal-spectral signature, which would be lost if the two dimensions are modeled separately. To effectively capture these dynamics, TasGen first disentangles temporal and spectral features to isolate their individual contributions, and then couples them to model their synergistic interactions. In the first stage, we introduce a hierarchical temporal-spectral variational autoencoder (HTS-VAE) with a dual-dimension embedding to learn low-dimensional latent patterns of normal samples by first disentangling and then jointly embedding temporal and spectral information. This temporal-spectral embedding enables robust anomaly detection by identifying deviations from learned joint patterns. In the second stage, a classifier trained on stable samples relabels change points across time to generate dynamic samples. To not only detect but also explain surface dynamics, we further propose an anomaly interpretation method based on Gibbs sampling, which attributes changes to specific spectral-temporal dimensions.

Figures

Figures reproduced from arXiv: 2506.02574 by the authors.

Figure 1
Figure 1. The workflow of each dynamic mapping paradigm. (a) is the workflow of our proposed method; (b) is the workflow of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The land surface change is a form of spatiotemporal anomaly. This set of figures and the NDVI variation show how the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of TasGen. (a) is the variational net, and (b) is the generative net, and (c) is the sample [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The sample data structure. This produces a compressed temporal embedding et that captures temporal-spectral dynamics such as phe￾nological cycles or disturbance trends. The embedding is then decoded via Deconv1D D layers to reconstruct a smoothed version of the input: …
Figure 5
Figure 5. Figure 5: The study areas and the classification systems in our study. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The results of two tasks in six regions (in %), including anomaly detection, dynamic sample relabeling and dynamic [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The visualization results of anomaly detection. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The results of Gibbs sampling-based anomaly attribution. The attribution score of B1 band is the largest. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: The visualization results of dynamic mapping performance of TasGen in SZ and PL. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: The visualization results of dynamic mapping, with the comparison with TSSCD and CCDC. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: The visualization results of change detection performance of TSSCD and CCDC, with the same areas of (b)-(d) as the [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 33 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    , author Vinyals, O

    author Bengio, S. , author Vinyals, O. , author Jaitly, N. , & author Shazeer, N. ( year 2015 ). title Scheduled sampling for sequence prediction with recurrent neural networks . journal Advances in neural information processing systems \/ , volume 28 \/

    work page 2015

  3. [3]

    , & author George, E

    author Casella, G. , & author George, E. I. ( year 1992 ). title Explaining the gibbs sampler . journal The American Statistician \/ , volume 46 \/ , pages 167--174

    work page 1992

  4. [4]

    , author Wang, J

    author Chen, S. , author Wang, J. , & author Gong, P. ( year 2023 ). title Robot: A spatiotemporal fusion model toward seamless data cube for global remote sensing applications . journal Remote Sensing of Environment \/ , volume 294 \/ , pages 113616

    work page 2023

  5. [5]

    , author Wang, J

    author Chen, S. , author Wang, J. , author Liu, Q. , author Liang, X. , author Liu, R. , author Qin, P. , author Yuan, J. , author Wei, J. , author Yuan, S. , author Huang, H. et al. ( year 2024 ). title Global 30-m seamless data cube (2000--2022) of land surface reflectance generated from landsat-5, 7, 8, 9 and modis terra constellations . journal Earth ...

    work page 2024

  6. [6]

    , author Gulcehre, C

    author Chung, J. , author Gulcehre, C. , author Cho, K. , & author Bengio, Y. ( year 2014 ). title Empirical evaluation of gated recurrent neural networks on sequence modeling . journal arXiv preprint arXiv:1412.3555 \/ ,

    arXiv 2014

  7. [7]

    author Davidson, N. C. ( year 2014 ). title How much wetland has the world lost? long-term and recent trends in global wetland area . journal Marine and Freshwater Research \/ , volume 65 \/ , pages 934--941

    work page 2014

  8. [8]

    , author S nderby, S

    author Fraccaro, M. , author S nderby, S. K. , author Paquet, U. , & author Winther, O. ( year 2016 ). title Sequential neural models with stochastic layers . journal Advances in neural information processing systems \/ , volume 29 \/

    work page 2016

Show all 43 references
  1. [9]

    , & author O’Neill, B

    author Gao, J. , & author O’Neill, B. C. ( year 2020 ). title Mapping global urban land for the 21st century with data-driven simulations and shared socioeconomic pathways . journal Nature communications \/ , volume 11 \/ , pages 2302

    2020

  2. [10]

    ( year 1989 )

    author Geweke, J. ( year 1989 ). title Bayesian inference in econometric models using monte carlo integration . journal Econometrica: Journal of the Econometric Society \/ , (pp. pages 1317--1339 )

    1989

  3. [11]

    ( year 2012 )

    author Gong, P. ( year 2012 ). title Remote sensing of environmental change over china: A review . journal Chinese Science Bulletin \/ , volume 57 \/ , pages 2793--2801

    2012

  4. [12]

    , author Chen, B

    author Gong, P. , author Chen, B. , author Li, X. , author Liu, H. , author Wang, J. , author Bai, Y. , author Chen, J. , author Chen, X. , author Fang, L. , author Feng, S. et al. ( year 2020 ). title Mapping essential urban land use categories in china (euluc-china): Prelimi...

    2020

  5. [13]

    , author Liu, H

    author Gong, P. , author Liu, H. , author Zhang, M. , author Li, C. , author Wang, J. , author Huang, H. , author Clinton, N. , author Ji, L. , author Li, W. , author Bai, Y. et al. ( year 2019 ). title Stable classification with limited sample: Transferring a 30-m resolution ...

    2019

  6. [14]

    , author Wang, J

    author Gong, P. , author Wang, J. , & author Huang, H. ( year 2024 ). title Stable classification with limited samples in global land cover mapping: Theory and experiments . journal Sci. Bull \/ , volume 69 \/ , pages 1862--1865

    2024

  7. [15]

    , author Yan, J

    author He, H. , author Yan, J. , author Liang, D. , author Sun, Z. , author Li, J. , & author Wang, L. ( year 2024 ). title Time-series land cover change detection using deep learning-based temporal semantic segmentation . journal Remote Sensing of Environment \/ , volume 305 ...

    2024

  8. [16]

    , author Wang, J

    author Huang, H. , author Wang, J. , author Liu, C. , author Liang, L. , author Li, C. , & author Gong, P. ( year 2020 ). title The migration of training samples towards dynamic global land cover mapping . journal ISPRS Journal of Photogrammetry and Remote Sensing \/ , volume ...

    2020

  9. [17]

    author Kennedy, R. E. , author Yang, Z. , & author Cohen, W. B. ( year 2010 ). title Detecting trends in forest disturbance and recovery using yearly landsat time series: 1. landtrendr—temporal segmentation algorithms . journal Remote Sensing of Environment \/ , volume 114 \/ ...

    2010

  10. [18]

    author Kingma, D. P. , & author Dhariwal, P. ( year 2018 ). title Glow: Generative flow with invertible 1x1 convolutions . journal Advances in neural information processing systems \/ , volume 31 \/

    2018

  11. [19]

    , author Gong, P

    author Li, C. , author Gong, P. , author Wang, J. , author Zhu, Z. , author Biging, G. S. , author Yuan, C. , author Hu, T. , author Zhang, H. , author Wang, Q. , author Li, X. et al. ( year 2017 ). title The first all-season sample set for mapping global land cover with lands...

    2017

  12. [20]

    , author Zhao, Y

    author Li, Z. , author Zhao, Y. , author Han, J. , author Su, Y. , author Jiao, R. , author Wen, X. , & author Pei, D. ( year 2021 ). title Multivariate time series anomaly detection and interpretation using hierarchical inter-metric and temporal embedding . In booktitle Proce...

    2021

  13. [21]

    , & author Yu, S

    author Lin, Q. , & author Yu, S. ( year 2018 ). title Losses of natural coastal wetlands by land conversion and ecological degradation in the urbanizing chinese coast . journal Scientific reports \/ , volume 8 \/ , pages 15046

    2018

  14. [22]

    author Mariano, D. A. , author dos Santos, C. A. , author Wardlow, B. D. , author Anderson, M. C. , author Schiltmeyer, A. V. , author Tadesse, T. , & author Svoboda, M. D. ( year 2018 ). title Use of remote sensing indicators to assess effects of drought and human-induced lan...

    2018

  15. [23]

    author Nicholls, R. J. ( year 2004 ). title Coastal flooding and wetland loss in the 21st century: changes under the sres climate and socio-economic scenarios . journal Global environmental change \/ , volume 14 \/ , pages 69--86

    2004

  16. [24]

    , author Cottam, A

    author Pekel, J.-F. , author Cottam, A. , author Gorelick, N. , & author Belward, A. S. ( year 2016 ). title High-resolution mapping of global surface water and its long-term changes . journal Nature \/ , volume 540 \/ , pages 418--422

    2016

  17. [25]

    author Seto, K. C. , author G \"u neralp, B. , & author Hutyra, L. R. ( year 2012 ). title Global forecasts of urban expansion to 2030 and direct impacts on biodiversity and carbon pools . journal Proceedings of the National Academy of Sciences \/ , volume 109 \/ , pages 16083--16088

    2012

  18. [26]

    , author Doan, T

    author Soubry, I. , author Doan, T. , author Chu, T. , & author Guo, X. ( year 2021 ). title A systematic review on the integration of remote sensing and gis to forest and grassland ecosystem health attributes, indicators, and measures . journal Remote Sensing \/ , volume 13 \...

    2021

  19. [27]

    , author Tarrio, K

    author Stanimirova, R. , author Tarrio, K. , author Turlej, K. , author McAvoy, K. , author Stonebrook, S. , author Hu, K.-T. , author Ar \'e valo, P. , author Bullock, E. L. , author Zhang, Y. , author Woodcock, C. E. et al. ( year 2023 ). title A global land cover training d...

    2023

  20. [28]

    , author Sutter, R

    author Wiens, J. , author Sutter, R. , author Anderson, M. , author Blanchard, J. , author Barnett, A. , author Aguilar-Amuchastegui, N. , author Avery, C. , & author Laine, S. ( year 2009 ). title Selecting and conserving lands for biodiversity: The role of remote sensing . j...

    2009

  21. [29]

    , author Fuchs, R

    author Winkler, K. , author Fuchs, R. , author Rounsevell, M. , & author Herold, M. ( year 2021 ). title Global land use changes are four times greater than previously estimated . journal Nature communications \/ , volume 12 \/ , pages 2501

    2021

  22. [30]

    author Wulder, M. A. , author Loveland, T. R. , author Roy, D. P. , author Crawford, C. J. , author Masek, J. G. , author Woodcock, C. E. , author Allen, R. G. , author Anderson, M. C. , author Belward, A. S. , author Cohen, W. B. et al. ( year 2019 ). title Current status of ...

    2019

  23. [31]

    , author Gong, P

    author Yang, J. , author Gong, P. , author Fu, R. , author Zhang, M. , author Chen, J. , author Liang, S. , author Xu, B. , author Shi, J. , & author Dickinson, R. ( year 2013 ). title The role of satellite remote sensing in climate change studies . journal Nature climate chan...

    2013

  24. [32]

    , & author Huang, X

    author Yang, J. , & author Huang, X. ( year 2021 ). title The 30\,m annual land cover dataset and its dynamics in china from 1990 to 2019 . journal Earth System Science Data \/ , volume 13 \/ , pages 3907--3925 . https://essd.copernicus.org/articles/13/3907/2021/. :10.5194/ess...

    2021 doi

  25. [33]

    , author Liang, X

    author Yuan, S. , author Liang, X. , author Lin, T. , author Chen, S. , author Liu, R. , author Wang, J. , author Zhang, H. , & author Gong, P. ( year 2025 ). title A comprehensive review of remote sensing in wetland classification and mapping . journal arXiv preprint arXiv:25...

    2025 arXiv

  26. [34]

    , author Lin, G

    author Yuan, S. , author Lin, G. , author Zhang, L. , author Dong, R. , author Zhang, J. , author Chen, S. , author Zheng, J. , author Wang, J. , & author Fu, H. ( year 2024 ). title Fusu: A multi-temporal-source land use change segmentation dataset for fine-grained urban sema...

    2024

  27. [35]

    , & author Weng, Q

    author Zhang, L. , & author Weng, Q. ( year 2016 ). title Annual dynamics of impervious surface in the pearl river delta, china, from 1988 to 2013, using time series landsat imagery . journal ISPRS Journal of Photogrammetry and Remote Sensing \/ , volume 113 \/ , pages 86--96

    2016

  28. [36]

    , & author Woodcock, C

    author Zhu, Z. , & author Woodcock, C. E. ( year 2014 ). title Continuous change detection and classification of land cover using all available landsat data . journal Remote sensing of Environment \/ , volume 144 \/ , pages 152--171

    2014

  29. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  30. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  31. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  34. [42]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  35. [43]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.