acceptodds
Under review as a conference paper at ICLR 2027

Improving Land Cover Classification Under Limited Data Using Advanced Augmentation and Transfer Learning

Abstract

Land cover classification from satellite imagery under limited labeled data remains a critical challenge in operational remote sensing, where annotation costs are prohibitive. This paper presents a systematic label-efficiency study of advanced data augmentation combined with transfer learning for remote sensing (RS) scene classification, evaluated across four public benchmarks: EuroSAT (10 classes, 27,000 images), UC Merced (21 classes, 2,100 images), NWPU-RESISC45 (45 classes, 31,500 images), and AID (28 classes,  8,400 images), totaling  69,000 images across 104 land cover classes. We combine an ImageNet-pretrained ResNet-50 backbone with three augmentation strategies—Mixup, CutMix, and Random Erasing—evaluated at six training data ratios from 1% to 50% with three independent random seeds per experiment. Transfer learning provides the most significant improvement, yielding gains of 10–28% over from-scratch baselines. Augmentation benefit is strongly data-regime dependent: mixing augmentations yield +21–31% gains under extreme label scarcity (1% data) but converge toward baseline at sufficient data (≥20%). The optimal augmentation strategy correlates significantly with dataset-specific inter-class overlap, quantified by the Intra/Inter-Class Distance Ratio (point-biserial r=0.261, p=0.0075, n=104 classes). EuroSAT CutMix at 1% training matches scratch-trained models using 20× more labeled data. Mechanistic evidence is provided via t-SNE visualization and Grad-CAM attention analysis. A practical three-step augmentation selection framework is proposed and validated.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
data augmentation, land cover classification, label efficiency, remote sensing, transfer learning, Mixup, CutMix, limited labeled data

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_29PLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This manuscript investigates the effectiveness of data augmentation and ImageNet-pretrained weights for remote sensing scene classification. Although the experiments demonstrate that these techniques can improve classification performance, the manuscript offers little, if any, original contribution. The findings largely reiterate well-established knowledge in both the machine learning and remote sensing communities. The work falls substantially short of the novelty and scientific contribution expected of a research article.

Critical strengths

The manuscript presents a series of straightforward experiments illustrating the effectiveness of data augmentation and pretrained weights for remote sensing scene classification. Based on experimental results from four datasets, the authors further propose a framework for selecting data augmentation strategies. However, the generalizability of this framework has not been established through evaluation on larger or more diverse datasets.

Critical weaknesses
  1. Lack of novelty and scientific contribution. The manuscript reads more like a course assignment report than an original research article. Its primary findings, data augmentation and ImageNet pretraining benefit remote sensing scene classification, have been widely recognized and extensively investigated in the existing literature. Reconfirming these established observations through additional experiments does not, by itself, constitute a meaningful scientific contribution.
  2. Limited generalizability of the proposed augmentation selection framework. The proposed framework appears to be derived specifically from the four datasets considered in this study. Its applicability beyond these datasets remains unclear, as no experiments are conducted on additional, larger-scale, or more recent benchmarks. Consequently, there is insufficient evidence to support its general applicability or practical value.
  3. Potentially fabricated references. Some references in the manuscript appear questionable and may not correspond to verifiable publications. Such as [5],[14],[18].
Questions for authors

The manuscript lacks a substantive original contribution, primarily reproduces findings that are already well established, and provides insufficient evidence for the generalizability of its proposed framework. The concerns regarding potentially unverifiable references further undermine confidence in the scholarly rigor of the work. These issues are fundamental rather than matters that can be adequately addressed through a routine rebuttal.

Reviewer_cjdqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is a tech report written by an LLM with several hallucinated references and author lists. Furthermore, it's not well polished or ready for review, see the entire page 6 is just 4 plots with a bunch of empty space and the paper contains many bulleted lists.

[5] "A survey on virtual sample generation and its applications in small sample learning" I couldn't find this online, I suspect it's hallucinated [18] "A comprehensive survey on augmentation policy search for data augmentation" I couldn't find this online, I suspect it's hallucinated [14] the reference states the authors are "X. Wang, R. Tian" when they are actually Liya Wang and Alex Tien https://www.spiedigitallibrary.org/conference-proceedings-of-spie/12732/1273202/Remote-sensing-scene-classification-with-masked-image-modeling/10.1117/12.2680898.short [4] the reference states the authors are "D. Hong, B. Zhang, H. Li, Y. Li, J. Li, J. Yao, N. Yokoya, S. Li, J. Chanussot, P. Ghamisi" but the authors are "D. Hong, B. Zhang, X. Li, Y. Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jia, A. Plaza, P. Gamba, J. Atli Benediktsson, J. Chanussot"

Critical strengths

None

Critical weaknesses

The paper is a tech report written by an LLM with several hallucinated references and author lists. Furthermore, it's not well polished or ready for review, see the entire page 6 is just 4 plots with a bunch of empty space and the paper contains many bulleted lists.

[5] "A survey on virtual sample generation and its applications in small sample learning" I couldn't find this online, I suspect it's hallucinated [18] "A comprehensive survey on augmentation policy search for data augmentation" I couldn't find this online, I suspect it's hallucinated [14] the reference states the authors are "X. Wang, R. Tian" when they are actually Liya Wang and Alex Tien https://www.spiedigitallibrary.org/conference-proceedings-of-spie/12732/1273202/Remote-sensing-scene-classification-with-masked-image-modeling/10.1117/12.2680898.short [4] the reference states the authors are "D. Hong, B. Zhang, H. Li, Y. Li, J. Li, J. Yao, N. Yokoya, S. Li, J. Chanussot, P. Ghamisi" but the authors are "D. Hong, B. Zhang, X. Li, Y. Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jia, A. Plaza, P. Gamba, J. Atli Benediktsson, J. Chanussot"

Questions for authors

N/A

Reviewer_8JSgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents a label-efficiency study of data augmentation combined with ImageNet transfer learning for remote sensing scene classification. The paper reports that transfer learning gives the largest gain, that augmentation helps most in the low-data regime, and that the best augmentation correlates with an intra/inter-class distance ratio computed from pretrained features. Based on this ratio, a three-step rule for selecting the augmentation is proposed, and t-SNE and Grad-CAM visualizations are given as mechanistic evidence.

Critical strengths

The experiments cover four benchmarks and six training ratios with three seeds, which gives a consistent picture of the trends. However, this is not sufficient to support acceptance given the weaknesses below.

Critical weaknesses
  • The paper has no technical novelty. It applies existing augmentations and standard ImageNet transfer learning to standard scene classification benchmarks and reports their effectiveness. Transferring ImageNet-pretrained CNNs to remote sensing scene classification has been studied for about a decade, and the paper does not contribute anything beyond this. The related work section is also very thin and does not position the study against the extensive literature on label-efficient and few-shot remote sensing scene classification.
  • The main claims are not fully supported by the experiments. In Table 4, the reported augmentation gains at 1% data are compared against training from scratch, making it impossible to separate the effects of pretraining and augmentation. In addition, the proposed selection rule (Sec. 5.1) contradicts the best-performing strategies in Table 5 for EuroSAT and AID.
  • Some references are unreliable; for example, [14] (arXiv:2302.14256) is attributed to "X. Wang, R. Tian," whereas the actual authors are Liya Wang and Alex Tien.
Questions for authors
  • What is the accuracy of Pretrained+NoAug at 1%, 2%, 5%, and 10% training data?
  • At 1%, UC Merced provides only about 17 training images for 21 classes. How was stratified sampling performed, and how is this consistent with "25 samples/class" in the caption of Fig. 5?
  • How is the binary variable for the point-biserial correlation defined when three different best strategies are involved?
  • References [5] and [18] could not be located. Please verify all references.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.