{\Omega}SFormer: Dual-Modal {\Omega}-like Super-Resolution Transformer Network for Cross-scale and High-accuracy Terraced Field Vectorization Extraction

Ce Zhang; Chang Li; Yongjun Zhang; Yu Wang

REVIEW 5 major objections 4 minor 2 references

{\Omega}SFormer: Dual-Modal {\Omega}-like Super-Resolution Transformer Network for Cross-scale and High-accuracy Terraced Field Vectorization Extraction

T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read ΩSFormer fuses high-resolution imagery with low-resolution DEM to extract terraced-field vectors at 0.976 mIoU.

desk verdict Useful new dataset, but the paper's central cross-scale claim contradicts its own Table 4; fix that before believing the 0.976 number. read the letter →

arxiv 2411.17088 v1 pith:XVBOJLUR submitted 2024-11-26 cs.CV

Chang Li , Yu Wang , Ce Zhang , Yongjun Zhang This is my paper · ORCID

classification cs.CV

keywords terracedfieldextractionvectorizationdual-modalfusionsuper-resolutiontransformersemanticsegmentationdigitalelevationmodelsoilandwaterconservation

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 claims that terraced fields—stepped farmlands built for soil and water conservation—can be extracted from remote sensing data as ready-to-use vector polygons with near-perfect agreement, reaching an mIoU of 0.976 and an overall accuracy of 0.981 on its same-scale 2 m dual-modal test set. To do this, it proposes ΩSFormer, a dual-branch Transformer whose Ω-like shape keeps high-resolution features alive through the encoder and fuses spectral imagery with digital elevation data at multiple scales. The paper also introduces DMRVD, a 22 441 km² dataset of 19 330 RGB images, 19 330 DEMs, and 371 150 hand-vectorized terraced-field polygons across nine counties in four Chinese provinces, which it says is the first dataset built specifically for deep-learning-based terraced-field vectorization extraction. If the reported results hold, automated monitoring of soil erosion, carbon sequestration, and agricultural land use would no longer need to rely on slow manual delineation.

What carries the argument

The carrying mechanism is the Ω-like network shape formed by the super-resolution Transformer connection module (SRTCM): at each encoder stage the original high-resolution features are fused with progressively downsampled features, making the network wide at the top and narrow at the bottom, with the two modality branches arranged symmetrically in a dual-modal, dual-branch (D2MB) input scheme. Multi-head self-attention on non-overlapping windows preserves local edge detail while 3×3 depthwise convolutions enable cross-window exchange. The spatial topological semantic relationship optimization (STSRO) then augments each pixel's representation by comparing it with whole object regions, and the vectorization extraction module (VEM) reparameterizes a contour vibration neural network to iteratively evolve segmentation boundaries into smooth vector polygons.

What would settle it

Retrain on labels created without any DEM or 3D terrain model, then compare dual-modal and RGB-only mIoU; if the 0.165-point margin shrinks to near zero, the claimed terrain-modal contribution is mostly label leakage rather than learned generalization.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a dual-modal, dual-branch super-resolution Transformer can make terraced-field extraction nearly indistinguishable from human annotation. The network fuses high-resolution RGB imagery with lower-resolution DEM data by continuously re-injecting original high-resolution features into each downsampling stage, so that fine edge information is never lost and the DEM is effectively super-resolved by the imagery. A spatial-topological refinement stage then sharpens boundary pixels using object-region relationships, and a contour-vibration network iteratively deforms the segmentation boundary into smooth vector outlines. Against U-Net, U-Net++, DeepLabV3+, HRNet, W-Net, W-Net++, HRFormer, SegFormer, and SAM, the paper reports the best mIoU of 0.976 and OA of 0.981, with mIoU improvements of 0.165, 0.297, and 0.128 over the best single-modal imagery, single-modal DEM, and dual-modal baselines. The abstract presents these as cross-scale dual-modal results; the paper's Table 4 reports the same 0.976/0.981 values for the same-scale 2 m dual-modal configuration.

Load-bearing premise

The load-bearing premise is that the DEM is an independent input, when in fact the ground-truth labels were created by looking at imagery on a 3D terrain surface derived from that same DEM, so the DEM's value may be baked into the labels rather than learned by the network.

Editorial extensions

If this is right

If correct, terraced-field mapping for soil-erosion monitoring can be automated across large, heterogeneous regions at near-human accuracy.
The vector output eliminates the pixelation and storage bloat of raster masks, giving infinitely scalable, directly analyzable field boundaries.
High-resolution imagery can act as a super-resolution prior for coarse DEMs, recovering terrain detail that the DEM alone cannot provide.
Separate dual-branch encoding is a better fusion strategy than stacking modalities into extra channels, because each branch can correct its own alignment errors before fusion.
DMRVD gives the research community a reusable open benchmark for terraced-field vectorization, with 371 150 labeled polygons across diverse terrain.

Reading between the lines

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

Editorial inference: since the labels were created by visual interpretation of imagery overlaid on a 3D terrain model built from the DEM, the DEM's measured contribution may include label leakage; an unbiased test needs labels produced independently of elevation data.
Editorial inference: the reported cross-scale advantage should be read with care, because the same paper's Table 4 shows the 2 m/12.5 m cross-scale configuration at mIoU 0.883, below the same-scale 2 m result of 0.976; a genuine super-resolution fusion mechanism would be expected to narrow that gap.
Editorial inference: a natural extension is to test ΩSFormer on terraced-field types and geographies outside the nine Chinese counties, for example Mediterranean or Andean terraces, to see whether the near-perfect accuracy reflects dataset-specific regularities or general terrain understanding.

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.

Desk Editor's Note

read the letter

Colleague, you should know two things about this one. The DMRVD dataset is genuinely new and open: 19,330 image-DEM pairs, 371k vector polygons, nine counties across China. That is a real resource for terraced field mapping and worth having. The ΩSFormer network, on the other hand, is an engineering stack of known modules—HRFormer-style high-res transformers, OCNet-like object context, CVNet for contour vectorization—and the paper's central claim about cross-scale fusion isn't supported by its own numbers.

Here's the problem. The abstract says cross-scale dual-modal datasets achieved mIoU 0.976 and OA 0.981, and that this beats single-modal and dual-modal baselines by 0.165, 0.297, 0.128. Table 4 tells a different story: the cross-scale experiment (2 m RGB, 12.5 m DEM, 2 m label) gives 0.883 mIoU. The 0.976 belongs to the same-scale 2m/2m/2m condition. So the headline result is not the cross-scale result; the cross-scale result is substantially worse. The claimed improvements are tied to the same-scale number, so they don't evidence cross-scale capability either. That's an internal contradiction in the paper's main selling point, not an interpretation difference.

There's also a label-leakage worry. Labels were made by annotating over a 3D model that displays the DEM (Section 2.2), so DEM information is baked into the ground truth. That makes it hard to interpret the apparent gain from adding DEM. Add no error bars, a circular dataset-quality argument (Section 4.4.1 validates the dataset by the authors' own model's accuracy), and no quantitative vectorization metrics—the vectorization module is shown only qualitatively.

The net is: the dataset is worth something, the network is a plausible engineering contribution, but the paper as written overclaims. The good news is that the problems are correctable in revision: rewrite the abstract to match Table 4, run the cross-scale comparisons with an unbiased label protocol or at least disclose the entanglement, add uncertainty and vector boundary metrics, and reframe the claim from 'cross-scale super-resolution works' to 'same-scale works and cross-scale is a harder open problem.'

I'd send it to peer review—the dataset justifies referee time—but with the expectation of major revision. The reader's REJECT is fair for the current version; I'd soften it to 'major revision required' because the underlying resource is real and the central conflict can be fixed without new collection.

Referee Report

5 major / 4 minor

Summary. The paper proposes ΩSFormer, a dual-modal Transformer network for terraced-field vectorization extraction (TFVE) from high-resolution RGB imagery and DEM data, together with a new dual-modal raster-vector dataset (DMRVD) covering nine counties in China. The authors claim that ΩSFormer is the first model to address cross-modal and cross-scale super-resolution fusion for TFVE, and report very high accuracy: mIoU 0.976 and OA 0.981, with improvements of 0.165, 0.297, and 0.128 over the best single-modal imagery, single-modal DEM, and dual-modal baselines, respectively. The manuscript includes comparisons with U-Net, U-Net++, DeepLabV3+, HRNet, HRFormer, SegFormer, W-Net, W-Net++, and SAM, an ablation of STSRO, and cross-scale experiments with three resolution combinations.

Significance. If the central claims were correct, the paper would contribute a useful dataset and architecture for terraced-field extraction, an application relevant to soil and water conservation. The release of DMRVD and code is a strength, and the comparison against multiple baselines is a positive feature. However, the paper's headline claim of cross-scale dual-modal performance is internally contradicted by its own Table 4, and the label-production procedure entangles the DEM input with the ground truth. These issues undermine the main contribution as stated, so the current manuscript cannot be accepted as a reliable demonstration of cross-scale super-resolution TFVE.

major comments (5)

[Abstract vs. Table 4] The abstract states that 'For cross-scale dual-modal datasets, ΩSFormer achieved the best performance with mIOU and OA values of 0.976 and 0.981 respectively,' but Table 4 defines the cross-scale configuration as RGB 2m, DEM 12.5m, and label 2m, and reports mIoU 0.883 for that row. The 0.976 value belongs to the same-scale 2m/2m/2m configuration. This is an internal contradiction in the paper's central claim about cross-scale performance.
[Section 4.3 and Table 4] The paper's own cross-scale experiment produces the worst accuracy among the three configurations: 0.883 mIoU, below both same-scale 2m (0.976) and same-scale 12.5m (0.921), and the text explicitly says that 'the semantic segmentation results of cross-scale are the worst.' The improvements of 0.165, 0.297, and 0.128 reported in the abstract are computed from Table 2's 0.976 result, which is the same-scale high-resolution result, not the cross-scale result. No baseline comparisons under the cross-scale condition are provided, so the claim that ΩSFormer is best under cross-scale input is unsupported.
[Section 2.2, label production] The ground-truth labels were produced by overlaying high-resolution imagery on a 3D model and visually interpreting boundaries, where the 3D model is constructed from the DEM that is later used as the second input modality. Because the DEM is entangled with label construction, the reported dual-modal gains over single-modal imagery may reflect label leakage rather than learned terrain understanding. The manuscript does not analyze or control for this dependency.
[Section 4.4.1] The quality of DMRVD is validated by the accuracy that the authors' own model achieves on it ('The quality of DMRVD can be evaluated based on the degree of accuracy achieved in the classification process'), which is circular. An independent assessment of label quality—for example, inter-annotator agreement or comparison with field survey data—is needed to substantiate the dataset contribution.
[Tables 2–4] All accuracy values are point estimates from a single training run with no error bars, repeated trials, or statistical significance tests. Consequently, the reported improvements (e.g., 0.128 over the best dual-modal baseline, 0.035 from STSRO in Table 3) cannot be distinguished from run-to-run variation, and the superiority claims are not statistically grounded.

minor comments (4)

[References and Table 2] SegFormer is cited as (Yuan et al. 2021a) in Section 3 and in the comparison network list, but the Introduction correctly attributes SegFormer to Xie et al. (2021); the citation key should be corrected to avoid ambiguity with HRFormer.
[Section 3.5, Eq. (10)] Equation (10) and the surrounding text contain garbled subscripts and superscripts (e.g., the indices in the discrete contour vibration equation) that make the formula difficult to read; the notation should be cleaned up.
[Table 2, SAM row] The comparison with SAM lacks any description of how SAM was prompted or adapted for binary terraced-field segmentation, making the comparison impossible to reproduce.
[General notation] The metric is spelled inconsistently as both 'mIOU' and 'mIoU'; please standardize the notation throughout the manuscript.

Circularity Check

3 steps flagged · score 6.0 of 10

Dual-modal gains are partly built into the labels via DEM-assisted annotation, and the abstract's 0.976 mIoU 'cross-scale' result is the same-scale 2m/2m/2m number relabeled; dataset quality is also validated by the accuracy achieved on itself.

self definitional [Section 2.2, 'Label production'; Fig. 2 caption]
"Due to the distinctive dual-modal structure of this study, remotely sensed imagery is overlaid onto the 3D model space for annotation, forming the basis for visual interpretation. ... Fig. 2. DMRVD display. The first row is remotely sensed imagery, the second row is DEM, the third row is 3D display of DEM, and the fourth row is binarized label."

The ground-truth vectors were created by overlaying imagery on a 3D model space whose geometry is the DEM (Fig. 2 shows the DEM's 3D display directly above the labels). That same DEM is then used as the second input modality of ΩSFormer. Consequently, the label already encodes topographic information from the DEM, so the claimed dual-modal gains (mIoU +0.165 over RGB-only, +0.297 over DEM-only, +0.128 over dual-modal baselines in Table 2) are partly guaranteed by the annotation procedure rather than by independent terrain understanding. DEM and labels are entangled by construction.
renaming known result [Abstract vs. Table 4 (Section 4.3)]
"Abstract: 'For cross-scale dual-modal datasets, ΩSFormer achieved the best performance with mIOU and OA values of 0.976 and 0.981 respectively.' Table 4: '2m 12.5m 2m 0.883 0.905' (cross-scale) and '2m 2m 2m 0.976 0.981' (same-scale)."

The 0.976/0.981 numbers in Table 4 belong to the same-scale high-resolution condition (RGB 2m, DEM 2m, label 2m), not to the cross-scale condition (RGB 2m, DEM 12.5m, label 2m), which is reported as 0.883/0.905. The abstract's headline cross-scale result is therefore the same-scale result renamed as 'cross-scale'; the central cross-scale super-resolution claim is not supported by the experiment that actually varies resolution.

1 more flagged steps

other [Section 4.4.1, 'The quality of DMRVD']
"Commonly, the quality of a dataset can be evaluated based on the degree of accuracy achieved in the classification process ... Table 2 shows the semantic segmentation accuracy, indicating that DMRVD performs better overall compared to SOTA networks. For DMRVD, ΩSFormer achieves the highest accuracy ... This suggests that DMRVD is more applicable and of superior quality."

The dataset's quality is validated by the high accuracy that the authors' own models achieve on the same dataset. This closes a self-referential loop: the model is judged good because it scores high on the dataset, and the dataset is judged good because the model scores high on it. No external benchmark, independent label check, or held-out cross-region validation is used to break the loop.

full rationale

The paper contains no fitted-parameter-then-predicted loop and no load-bearing self-citation chain; the network equations are standard attention/convolution blocks adapted from cited external work. The main circularity risk is in the data/label construction and in the reporting of the headline number. First, Section 2.2 states that annotation was performed by overlaying imagery on a 3D model space whose geometry is the DEM, and Fig. 2 confirms that the 3D display is generated from the DEM. Since the same DEM is the second input modality, the dual-modal accuracy advantage is partly an artifact of the label-generation process rather than independent evidence of cross-modal fusion. Second, the abstract's 0.976 mIoU 'cross-scale' claim is not the cross-scale result in Table 4 (0.883); it is the same-scale 2m/2m/2m result, so the central claim is a re-labeling of a different experimental condition rather than a derivation from cross-scale inputs. Third, Section 4.4.1 validates dataset quality by the accuracy obtained on the dataset itself, a self-referential evaluation. These issues are concrete and quotable, but they are partial: the architecture, ablations, and comparisons are independently described, and not every reported improvement is forced by definition. Overall circularity is therefore moderate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the accuracy of hand-drawn labels, the validity of resampling a 12.5m DEM to 2m for the main experiment, and the absence of spatial leakage in the 8:1:1 split. No new physical or theoretical entities are introduced. Three free hyperparameters or design choices are unstated or unreported.

free parameters (3)

Loss weights gamma and theta (Eq. 18) = not reported
Final loss combines L1, L2, L3 with weights gamma and theta; values are never given, and no sensitivity analysis is shown.
Dropout rate = 0.6 to 0.8
The paper states a range, not the exact value used in the reported runs.
Screening ratio of terraced to non-terraced tiles = 1:1
Dataset screened to a 1:1 ratio, a design choice that affects mIoU and is not representative of natural terrain prevalence.

assumptions (4)

domain assumption Ground truth labels are accurate and unbiased
Section 2.2 describes visual interpretation with eight elements; no field verification, inter-annotator agreement, or quality control metrics are reported.
domain assumption Resampling the 12.5m DEM to 2m produces a valid 2m DEM for training
Table 1 setting (b) resamples DEM to 2m; the 0.976 mIoU result comes from this setting, yet no validation shows that resampling preserves terrain structure relevant to terrace boundaries.
domain assumption Train/validation/test tiles are independent despite shared geography
The paper only states an 8:1:1 ratio; without spatial block holdout, terraces from the same county can appear in both training and test, inflating accuracy.
standard math Standard deep learning training assumptions
Optimizer, learning rate schedule, and backpropagation are standard; not a load-bearing assumption beyond typical practice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of {\Omega}SFormer: Dual-Modal {\Omega}-like Super-Resolution Transformer Network for Cross-scale and High-accuracy Terraced Field Vectorization Extraction." pith.science (2026). https://pith.science/paper/XVBOJLUR

@misc{pith2026241117088,
  author       = {Pith},
  title        = {Pith review of: \OmegaSFormer: Dual-Modal \Omega-like Super-Resolution Transformer Network for Cross-scale and High-accuracy Terraced Field Vectorization Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVBOJLUR}},
  note         = {Machine review of arXiv:2411.17088}
}

read the original abstract

Terraced field is a significant engineering practice for soil and water conservation (SWC). Terraced field extraction from remotely sensed imagery is the foundation for monitoring and evaluating SWC. This study is the first to propose a novel dual-modal {\Omega}-like super-resolution Transformer network for intelligent TFVE, offering the following advantages: (1) reducing edge segmentation error from conventional multi-scale downsampling encoder, through fusing original high-resolution features with downsampling features at each step of encoder and leveraging a multi-head attention mechanism; (2) improving the accuracy of TFVE by proposing a {\Omega}-like network structure, which fully integrates rich high-level features from both spectral and terrain data to form cross-scale super-resolution features; (3) validating an optimal fusion scheme for cross-modal and cross-scale (i.e., inconsistent spatial resolution between remotely sensed imagery and DEM) super-resolution feature extraction; (4) mitigating uncertainty between segmentation edge pixels by a coarse-to-fine and spatial topological semantic relationship optimization (STSRO) segmentation strategy; (5) leveraging contour vibration neural network to continuously optimize parameters and iteratively vectorize terraced fields from semantic segmentation results. Moreover, a DMRVD for deep-learning-based TFVE was created for the first time, which covers nine study areas in four provinces of China, with a total coverage area of 22441 square kilometers. To assess the performance of {\Omega}SFormer, classic and SOTA networks were compared. The mIOU of {\Omega}SFormer has improved by 0.165, 0.297 and 0.128 respectively, when compared with best accuracy single-modal remotely sensed imagery, single-modal DEM and dual-modal result.

Figures

Figures reproduced from arXiv: 2411.17088 by the authors.

**Figure 5.** Illustrating the STSRO spatial topological semantic relationship for the pixel marked with the yellow square. (a) Atrous spatial pyramid pooling: the multi-scale context is a set of sparsely sampled pixels marked with blue squares and green squares. The pixels exhibiting different colors are indicative of disparate dilation rates. These pixels are dispersed throughout both the object region and the background region… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Aakerberg, A., Johansen, A.S., Nasrollahi, K., & Moeslund, T.B. (2022). Semantic segmentation guided real-world super -resolution. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 449-458 Amundson, R., Berhe, A.A., Hopmans, J.W., Olson, C., Sztein, A.E., & Sparks, D.L. (2015). Soil and human security in the 21st centur...

work page 2022
[2018]

Proceedings: Lecture Notes in Computer Science (LNCS 11211), 833-851 Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A.L

15th European Conference. Proceedings: Lecture Notes in Computer Science (LNCS 11211), 833-851 Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A.L. (2018b). DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. Ieee Transactions on Pattern Analysis and Machine Intelligence, 40, 834...

work page 2018

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Aakerberg, A., Johansen, A.S., Nasrollahi, K., & Moeslund, T.B. (2022). Semantic segmentation guided real-world super -resolution. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 449-458 Amundson, R., Berhe, A.A., Hopmans, J.W., Olson, C., Sztein, A.E., & Sparks, D.L. (2015). Soil and human security in the 21st centur...

work page 2022

[2] [2018]

Proceedings: Lecture Notes in Computer Science (LNCS 11211), 833-851 Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A.L

15th European Conference. Proceedings: Lecture Notes in Computer Science (LNCS 11211), 833-851 Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A.L. (2018b). DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. Ieee Transactions on Pattern Analysis and Machine Intelligence, 40, 834...

work page 2018