Application of Generative Adversarial Network (GAN) for Synthetic Training Data Creation to improve performance of ANN Classifier for extracting Built-Up pixels from Landsat Satellite Imagery

Amritendu Mukherjee; Dipanwita Sinha Mukherjee; Parthasarathy Ramachandran

REVIEW 5 major objections 4 minor 18 references

Application of Generative Adversarial Network (GAN) for Synthetic Training Data Creation to improve performance of ANN Classifier for extracting Built-Up pixels from Landsat Satellite Imagery

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Adding GAN-generated synthetic pixels raises ANN built-up classification accuracy from 93.31% to 99.83% on a held-out Landsat 7 test set.

desk verdict A clear, honest demonstration of GAN-based augmentation that fails to isolate the GAN's contribution due to a missing duplication control. read the letter →

arxiv 2501.19283 v1 pith:RKBKSIXY submitted 2025-01-31 cs.CV cs.LG

Amritendu Mukherjee , Dipanwita Sinha Mukherjee , Parthasarathy Ramachandran This is my paper · ORCID

classification cs.CVcs.LG

keywords GenerativeAdversarialNetworksynthetictrainingdatabuilt-upclassificationLandsat7ANNclassifierKolmogorov-SmirnovtestBallDivergencekappacoefficient

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

The paper tries to show that a simple GAN can generate synthetic built-up pixels from just 100 real examples, and that adding these synthetic pixels to the training set of an ANN classifier improves the classifier's ability to separate built-up from non-built-up pixels in Landsat 7 imagery. The claim matters because pixel-level classification of low-resolution multispectral imagery is typically limited by the scarcity of pure single-class training pixels. If the approach works, it offers a low-cost way to boost classifier performance without additional manual labeling.

What carries the argument

The central object is a four-layer GAN whose generator maps 100-dimensional uniform noise to 6 Landsat spectral bands (B1-B6) and whose discriminator scores real versus synthetic pixels; distributional fidelity of generated sets is then verified with per-band Kolmogorov-Smirnov tests and a Ball Divergence test for the joint distribution. These verified synthetic sets are fed to a single-hidden-layer ANN classifier whose accuracy on the test set is monitored as more generated pixels are added.

What would settle it

Take the same 100 original built-up pixels, make three sets of 100 random copies with replacement, add them to the training set in the same increments, and train the same ANN classifier; if accuracy and kappa reach 0.9983 and 0.9958, then the GAN's generative modeling is not responsible for the improvement.

Watch

Extended reading notes

Core claim

On a held-out test set of 2000 built-up and 5000 non-built-up pixels, the ANN classifier's overall accuracy rose from 0.9331 to 0.9983 and its kappa coefficient from 0.8277 to 0.9958 as three sets of 100 GAN-generated built-up pixels were incrementally added to the original 100 built-up training pixels. The generated pixels were shown to match the original marginal band distributions (by per-band Kolmogorov-Smirnov tests) and joint six-band distribution (by Ball Divergence test) at p-values above 0.05. The authors interpret this steady improvement as evidence that the GAN provides a faithful generative augmentation for small training sets.

Load-bearing premise

The claimed gain from GAN-generated pixels is not compared against simply duplicating original pixels, so the improvement may reflect more training examples rather than the GAN's learned distribution.

Editorial extensions

If this is right

Adding generated built-up pixels in three steps of 100 raises overall accuracy from 0.9331 to 0.9983 and kappa from 0.8277 to 0.9958 on a held-out set of 7000 pixels.
The generated pixel sets clear per-band Kolmogorov-Smirnov and joint Ball Divergence tests against the original distribution, supporting the claim that they are statistically similar to real built-up pixels.
The improvement pattern suggests that the GAN's synthetic examples are informative enough to act as a data-augmentation strategy when pure-class training pixels are scarce.
The same pipeline (GAN generation followed by distribution checks and incremental ANN training) can be transferred to other low-resolution multispectral classification tasks.

Reading between the lines

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

A control condition that adds an equal number of duplicated original pixels would separate the benefit of larger sample size from the benefit of GAN-generated samples.
The experiment uses a very small ANN (one hidden layer, two hidden units); whether the same gains appear for larger or deeper classifiers remains open.
The methodology could be re-run with a VAE or a diffusion model on the same 100-pixel set to test whether the accuracy ceiling is GAN-specific.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: the paper gives a reproducible recipe for using a small GAN to inflate a 100-pixel built-up training set and reports a large accuracy jump in a Landsat ANN classifier. The distribution checks are a nice touch. But the headline gain is not attributable to the GAN because there is no control that adds duplicated original pixels, and the ANN hyperparameters drift across configurations.

What's new: not much, honestly. GAN-based data augmentation is standard. The application to six-band Landsat pixel vectors with a tiny training set is routine. But the paper is transparent about the architecture and the evaluation, and it ships a concrete protocol that someone else could run. The KS and Ball Divergence tests at least show the generated samples are statistically similar to the originals, which is more than many papers in this area do.

Soft spots: the missing duplication control is the load-bearing one. With 100 built-up and 400 non-built-up pixels, adding any built-up examples rebalances the classes and strengthens the minority gradient signal. The monotonic improvement in Table 3 could just be more data. The weight decay changes from 0.4 to 0.3 to 0.1 across rows, a second confound. No error bars or repeated runs, and a single study site. The distribution tests have low power at n=100; failing to reject similarity is not proof the generated points add information beyond the original spread. I also agree with the reader that the paper ignores a large existing literature on GANs for satellite data, which weakens the positioning.

Who should read it: remote sensing practitioners who want a simple augmentation recipe and are aware that the evidence here is suggestive, not definitive. The paper is a decent demonstration, not a new method.

Recommendation: it deserves a serious referee, because the question is meaningful and the fix is straightforward—add a duplication baseline, fix the weight decay, and ideally run a second site. On the current evidence I would not accept the claim as proven, but I would not desk-reject it.

Referee Report

5 major / 4 minor

Summary. The paper proposes a GAN-based data augmentation method for ANN classification of built-up pixels from Landsat 7 imagery. From 100 original built-up pixels, the authors train a simple GAN and generate three sets of 100 synthetic built-up pixels each; they then train an ANN classifier on the original 100 built-up and 400 non-built-up pixels, and incrementally add the generated built-up sets. On a held-out test set (2000 built-up and 5000 non-built-up pixels), they report that accuracy improves from 0.9331 to 0.9983 and kappa from 0.8277 to 0.9958. Distribution similarity between generated and original pixels is assessed via Kolmogorov–Smirnov tests per band and a Ball Divergence test for the joint distribution.

Significance. If the causal attribution were established, this would be a practical and low-cost way to improve classification in small-sample remote sensing settings, and the paper provides a clear experimental template with honest reporting of all numerical results and appropriate distributional checks. The main barrier is that the experimental design does not isolate the effect of the generative model: the observed gains are equally consistent with a pure class-rebalancing or sample-size effect. The manuscript is therefore not yet ready for publication as it stands, but the core question is well-posed and the required additional experiments are straightforward.

major comments (5)

[Section 4, Table 3] The central attribution of the improvement to GAN-generated data is not established because the design lacks a duplication control. The training set starts with 100 built-up and 400 non-built-up pixels (a 4:1 imbalance); adding 100, 200, or 300 additional built-up examples—whether generated or simply copies of the original 100—rebalances the class ratio and increases minority-class gradient signal. The monotonic accuracy and kappa increases in Table 3 are therefore fully consistent with a sample-size/class-rebalancing effect. The authors must include a control condition where the same numbers of duplicated original built-up pixels are added, and show that GAN-generated data outperforms that control.
[Section 3, Table 3] The ANN hyperparameters are not held constant across the compared configurations: weight decay lambda changes from 0.4 to 0.3 to 0.1 as generated sets are added, and the number of hidden units is re-selected by grid search for each configuration. This introduces a second confound: the accuracy gains could partly stem from better regularization or other hyperparameter choices rather than from the added data. The comparison should either fix all hyperparameters across all training configurations, or report the performance of each configuration with the same hyperparameter set.
[Section 4, Table 3] All reported accuracy measures come from a single run; there are no error bars, confidence intervals, or repeated cross-validation results. Given the stochastic nature of GAN training and ANN initialization, the observed monotonic trend might not be reproducible under different random seeds. The authors should repeat the entire pipeline (GAN training, data generation, ANN training, evaluation) multiple times and report mean and standard deviation (or confidence intervals) for accuracy and kappa.
[Section 4, Tables 1 and 2] The distribution tests show that the generated samples are statistically indistinguishable from the original 100 pixels, but they do not demonstrate that the generated samples add any information beyond the original distribution. With n=100 per set, the KS and Ball Divergence tests have limited power, and the high p-values are consistent with the GAN simply reproducing the original distribution. If the generated data merely copy the original 100 pixels, one would expect the same improvement as from duplication; the experiment needs a test that specifically measures whether the generated data expand the learned decision boundary relative to simple duplication.
[Section 3, GAN architecture] The GAN training procedure is not described in sufficient detail for reproducibility: there is no learning rate, optimizer, number of epochs, batch size, or convergence criterion, and the GAN loss (Equation 1) has a typo that obscures the exact objective. Without these details, an independent researcher cannot reproduce the generated sets, and the reader cannot assess whether the GAN was trained to convergence or whether the final generated pixels are sensitive to training settings.

minor comments (4)

[Table 1] Table 1 lists "B1 (BLUE)" twice and omits "B6 (SWIR2)"; the last row should presumably be B6. Additionally, "B4 (INR)" should be "B4 (NIR)", and the columns for the three generated sets appear to be misaligned with the band rows.
[Equation (1)] Equation (1) has a missing closing parenthesis in the generator term: it should read log(1 - D(G(z))), not log(1 - D(G(z)). The authors should correct this typo for clarity.
[Section 2, paragraph 2] The sentence beginning "True and False Color Composite images ... have been provided in Figure 1 Google Earth Engine 3(GEE) for the same time period has been used" is grammatically broken and should be rewritten to clearly state that Google Earth Engine was used to create the training and testing sets.
[General] The paper does not include a data or code availability statement. Given the manual verification of training pixels and the small experimental scale, providing at least the list of pixel coordinates or the trained model outputs would greatly improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the GAN-augmented ANN result is an empirical measurement, not a prediction forced by definitions or self-citation.

full rationale

The paper reports an experiment: a GAN is trained on 100 original built-up pixels, three 100-pixel synthetic sets are generated, distribution tests (KS and Ball Divergence) are run, and an ANN is trained with incremental additions of the synthetic built-up pixels and tested on a held-out set (Section 4, Table 3). No step in this chain is circular in the sense required by the review rules. The GAN is not fitted to the test set; the ANN is evaluated on 2000 built-up and 5000 non-built-up test pixels that are separate from the 100+400 training pixels. The accuracy and kappa improvements (0.9331 to 0.9983 and 0.8277 to 0.9958) are observed outcomes, not quantities that reduce to the input by construction. There are no load-bearing self-citations: references [1]-[14] are standard methodological sources and none are by the authors. The main weakness is experimental, not circular: the design lacks a control that adds an equal number of duplicated original built-up pixels, so the reported gains may reflect class rebalancing or additional minority-class samples rather than a generative-model advantage; the varying weight-decay parameter (lambda from 0.4 to 0.1 across Table 3 rows) is an additional confound. These are correctness and interpretation concerns, not circularity. Since the core claim is not definitionally tied to its inputs, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce new physical or conceptual entities. The free parameters are all architectural and training choices. The axioms are standard assumptions about GAN convergence and statistical tests. The most important unstated assumption is that the manual GEE labels are correct and that the single study site is representative.

free parameters (6)

GAN hidden layer sizes = G: 100-100-100-6, D: 6-100-100-1
Chosen by hand after experimentation (Section 3). These architectural choices affect the generated data distribution and are not derived from theory.
GAN activation functions = Sigmoid, ReLU, ReLU
Chosen by hand (Section 3). The activation sequence is a design decision that influences training dynamics.
Latent noise dimension = 100
Chosen by hand (Section 3). The dimensionality of z is arbitrary.
ANN hidden units = 2
Selected by grid search with 10-fold CV (Section 3). The optimal value of 2 is a free parameter tuned on the training data.
ANN weight decay lambda = 0.4, 0.3, 0.1, 0.1
Varies per training configuration from grid search (Section 3, Table 3). These are tuned hyperparameters.
Number of generated pixels per set = 100 per set, 3 sets
Arbitrarily chosen (Section 3). The number of synthetic samples added is a free decision that could affect the degree of improvement.

assumptions (4)

domain assumption The GAN converges to the true data distribution and generates samples representative of the original pixel distribution.
The paper relies on successful GAN training without analyzing mode collapse or convergence (Section 3). The distribution tests only compare moments of the generated and original sets, not full representativeness.
standard math Universal approximation theorem justifies a single hidden layer ANN with 2 nodes.
Invoked in Section 3 to argue that one hidden layer suffices for any bounded continuous function. This is a standard theorem, but it does not guarantee that 2 nodes are sufficient for the specific LULC classification task.
domain assumption KS and Ball Divergence tests are valid for the sample sizes and distributions involved.
Section 4 uses these nonparametric tests on 100-sample sets. Their p-values are used as evidence of distributional equivalence, but the tests may have limited power with small samples.
domain assumption Manually verified Google Earth Engine pixels are accurate ground truth.
The training and test labels come from manual verification on GEE (Section 2). Any label noise would propagate into the reported accuracy and kappa.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Application of Generative Adversarial Network (GAN) for Synthetic Training Data Creation to improve performance of ANN Classifier for extracting Built-Up pixels from Landsat Satellite Imagery." pith.science (2026). https://pith.science/paper/RKBKSIXY

@misc{pith2026250119283,
  author       = {Pith},
  title        = {Pith review of: Application of Generative Adversarial Network (GAN) for Synthetic Training Data Creation to improve performance of ANN Classifier for extracting Built-Up pixels from Landsat Satellite Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKBKSIXY}},
  note         = {Machine review of arXiv:2501.19283}
}

abstract

Training a neural network for pixel based classification task using low resolution Landsat images is difficult as the size of the training data is usually small due to less number of available pixels that represent a single class without any mixing with other classes. Due to this scarcity of training data, neural network may not be able to attain expected level of accuracy. This limitation could be overcome using a generative network that aims to generate synthetic data having the same distribution as the sample data with which it is trained. In this work, we have proposed a methodology for improving the performance of ANN classifier to identify built-up pixels in the Landsat$7$ image with the help of developing a simple GAN architecture that could generate synthetic training pixels when trained using original set of sample built-up pixels. To ensure that the marginal and joint distributions of all the bands corresponding to the generated and original set of pixels are indistinguishable, non-parametric Kolmogorov Smirnov Test and Ball Divergence based Equality of Distributions Test have been performed respectively. It has been observed that the overall accuracy and kappa coefficient of the ANN model for built-up classification have continuously improved from $0.9331$ to $0.9983$ and $0.8277$ to $0.9958$ respectively, with the inclusion of generated sets of built-up pixels to the original one.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

[1]

The fundamental problems that these models try to solve are, (i) to estimate the underlying distribution of the observed data and (ii) to generate sample data from the same

INTRODUCTION In recent times, there has been a lot of research activities in the area of Generative Models[1, 2] as it demonstrates great potentials to improve existing state of the technology in vari- ous areas like Language Models, Image & Video Generation, Code Generation, Speech Processing etc. The fundamental problems that these models try to solve a...

work page
[2]

DATA & STUDY AREA In order to investigate the classification performance of the proposed method, Landsat7 Enhanced Thematic Mapper Plus (ETM+) image corresponding to January 2017 is sourced from U.S.G.S Earth Explorer 2. The area of the study site is approximately 12100 km2(1◦ Latitude×1◦ Longitude) and contains Jaipur which is the capital and the largest...

work page Pith review arXiv 2017
[3]

METHODOLOGY First, we’ve developed a simple GAN architecture to generate synthetic Built-Up pixels using actual Built-Up pixels from the original training set as described in Section 2. In the GAN framework, for the purpose of learning the distribution pg, a differentiable function G(z; θg) represented by a neural net- work is crafted to map the input noi...

work page 2025
[4]

RESULTS & DISCUSSIONS As discussed previously in Section 3, p-values associated with the Kolmogorov–Smirnov (KS) test that has been performed to confirm that the individual bands corresponding to each set of generated Built-Up pixels and the same bands from the original set of Built-Up pixels represent same distribution, have been reported in Table 1. As ...

work page 2025
[5]

Deep gen- erative models: Survey,

Achraf Oussidi and Azeddine Elhassouny, “Deep gen- erative models: Survey,” in 2018 International confer- ence on intelligent systems and computer vision (ISCV) . IEEE, 2018, pp. 1–8

work page 2018
[6]

An introduction to deep generative modeling,

Lars Ruthotto and Eldad Haber, “An introduction to deep generative modeling,” GAMM-Mitteilungen, vol. 44, no. 2, pp. e202100008, 2021

work page 2021
[7]

Deep gaus- sian mixture models,

Cinzia Viroli and Geoffrey J McLachlan, “Deep gaus- sian mixture models,” Statistics and Computing, vol. 29, pp. 43–51, 2019

work page 2019
[8]

Auto-encoding variational bayes,

Diederik P Kingma, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

arXiv 2013

Show all 18 references

[9]

Generative adversar- ial networks,

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversar- ial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

2020
[10]

Denoising diffusion probabilistic models,

Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion probabilistic models,” Advances in neural in- formation processing systems , vol. 33, pp. 6840–6851, 2020

2020
[11]

Understanding diffusion models: A uni- fied perspective,

Calvin Luo, “Understanding diffusion models: A uni- fied perspective,” arXiv preprint arXiv:2208.11970 , 2022

2022 arXiv
[12]

Rose, Jacob J

Amy N. Rose, Jacob J. McKee, Marie L. Urban, and Eddie A. Bright, chapter LandScan 2017, LandScan. Oak Ridge National Laboratory, Oak Ridge, TN, 2018

2017
[13]

The kolmogorov-smirnov test for goodness of fit,

Frank J Massey Jr, “The kolmogorov-smirnov test for goodness of fit,” Journal of the American statistical As- sociation, vol. 46, no. 253, pp. 68–78, 1951

1951
[14]

Ball divergence: Nonparametric two sample test,

Wenliang Pan, Yuan Tian, Xueqin Wang, and Heping Zhang, “Ball divergence: Nonparametric two sample test,” Annals of statistics, vol. 46, no. 3, pp. 1109, 2018

2018
[15]

Multilayer feedforward networks are universal approx- imators,

Kurt Hornik, Maxwell Stinchcombe, and Halbert White, “Multilayer feedforward networks are universal approx- imators,” Neural networks, vol. 2, no. 5, pp. 359–366, 1989

1989
[16]

Deep learning,

Ian Goodfellow, “Deep learning,” 2016

2016
[17]

260–260, Springer US, Boston, MA, 2017

Kai Ming Ting, Confusion Matrix , pp. 260–260, Springer US, Boston, MA, 2017

2017
[18]

A coefficient of agreement for nominal scales,

Jacob Cohen, “A coefficient of agreement for nominal scales,” Educational and psychological measurement , vol. 20, no. 1, pp. 37–46, 1960

1960

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

The fundamental problems that these models try to solve are, (i) to estimate the underlying distribution of the observed data and (ii) to generate sample data from the same

INTRODUCTION In recent times, there has been a lot of research activities in the area of Generative Models[1, 2] as it demonstrates great potentials to improve existing state of the technology in vari- ous areas like Language Models, Image & Video Generation, Code Generation, Speech Processing etc. The fundamental problems that these models try to solve a...

work page

[2] [2]

DATA & STUDY AREA In order to investigate the classification performance of the proposed method, Landsat7 Enhanced Thematic Mapper Plus (ETM+) image corresponding to January 2017 is sourced from U.S.G.S Earth Explorer 2. The area of the study site is approximately 12100 km2(1◦ Latitude×1◦ Longitude) and contains Jaipur which is the capital and the largest...

work page Pith review arXiv 2017

[3] [3]

METHODOLOGY First, we’ve developed a simple GAN architecture to generate synthetic Built-Up pixels using actual Built-Up pixels from the original training set as described in Section 2. In the GAN framework, for the purpose of learning the distribution pg, a differentiable function G(z; θg) represented by a neural net- work is crafted to map the input noi...

work page 2025

[4] [4]

RESULTS & DISCUSSIONS As discussed previously in Section 3, p-values associated with the Kolmogorov–Smirnov (KS) test that has been performed to confirm that the individual bands corresponding to each set of generated Built-Up pixels and the same bands from the original set of Built-Up pixels represent same distribution, have been reported in Table 1. As ...

work page 2025

[5] [5]

Deep gen- erative models: Survey,

Achraf Oussidi and Azeddine Elhassouny, “Deep gen- erative models: Survey,” in 2018 International confer- ence on intelligent systems and computer vision (ISCV) . IEEE, 2018, pp. 1–8

work page 2018

[6] [6]

An introduction to deep generative modeling,

Lars Ruthotto and Eldad Haber, “An introduction to deep generative modeling,” GAMM-Mitteilungen, vol. 44, no. 2, pp. e202100008, 2021

work page 2021

[7] [7]

Deep gaus- sian mixture models,

Cinzia Viroli and Geoffrey J McLachlan, “Deep gaus- sian mixture models,” Statistics and Computing, vol. 29, pp. 43–51, 2019

work page 2019

[8] [8]

Auto-encoding variational bayes,

Diederik P Kingma, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

arXiv 2013