pith. sign in

arxiv: 2604.16147 · v1 · submitted 2026-04-17 · 💻 cs.CV · cs.AI

SWNet: A Cross-Spectral Network for Camouflaged Weed Detection

Henry O. Velesaca , Luigi Miranda , Angel D. Sappa This is my paper

Pith reviewed 2026-05-10 08:04 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords camouflaged weed detectioncross-spectral networknear-infraredweed segmentationbimodal fusionpyramid vision transformeredge-aware refinementagricultural computer vision
0
0 comments X

The pith

A cross-spectral network fuses visible and near-infrared images to detect weeds that blend with crops in dense fields.

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

The paper introduces SWNet to address plant camouflage where invasive weeds mimic the appearance of crops in visible light. It builds a network that adds near-infrared data to exploit how chlorophyll reflects differently in that spectrum, allowing separation of targets that look identical to the eye. The architecture combines a Pyramid Vision Transformer v2 backbone for long-range context, a gated fusion module for dynamic spectral integration, and an edge-aware refinement step for cleaner boundaries. Experiments on the Weeds-Banana dataset show the method outperforms ten prior approaches in segmentation accuracy. The work establishes that cross-spectral fusion plus boundary guidance is required for reliable performance inside thick agricultural canopies.

Core claim

SWNet is a bimodal end-to-end cross-spectral network that uses a Pyramid Vision Transformer v2 backbone to capture long-range dependencies, a Bimodal Gated Fusion Module to dynamically integrate Visible and Near-Infrared information, and an Edge-Aware Refinement module to produce sharper object boundaries. By leveraging the physiological differences in chlorophyll reflectance captured in the NIR spectrum, the proposed architecture effectively discriminates targets that are otherwise indistinguishable in the visible range and outperforms ten state-of-the-art methods on the Weeds-Banana dataset.

What carries the argument

The Bimodal Gated Fusion Module that dynamically integrates Visible and Near-Infrared information to exploit chlorophyll reflectance differences for discriminating camouflaged weeds.

Load-bearing premise

Physiological differences in chlorophyll reflectance captured in the NIR spectrum are sufficient to discriminate targets indistinguishable in the visible range under dense agricultural canopy conditions.

What would settle it

If a visible-spectrum-only network matches or exceeds SWNet accuracy on the Weeds-Banana dataset, the claimed necessity of NIR integration would not hold.

Figures

Figures reproduced from arXiv: 2604.16147 by Angel D. Sappa, Henry O. Velesaca, Luigi Miranda.

Figure 1
Figure 1. Figure 1: The overall architecture of the proposed SWNet. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example of RGB, NIR and mask images of the Weeds-Banana [ [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results on COD techniques that have achieved first or second place in at least one of the metrics in Table [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of three distinct configurations: a version [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

This paper presents SWNet, a bimodal end-to-end cross-spectral network specifically engineered for the detection of camouflaged weeds in dense agricultural environments. Plant camouflage, characterized by homochromatic blending where invasive species mimic the phenotypic traits of primary crops, poses a significant challenge for traditional computer vision systems. To overcome these limitations, SWNet utilizes a Pyramid Vision Transformer v2 backbone to capture long-range dependencies and a Bimodal Gated Fusion Module to dynamically integrate Visible and Near-Infrared information. By leveraging the physiological differences in chlorophyll reflectance captured in the NIR spectrum, the proposed architecture effectively discriminates targets that are otherwise indistinguishable in the visible range. Furthermore, an Edge-Aware Refinement module is employed to produce sharper object boundaries and reduce structural ambiguity. Experimental results on the Weeds-Banana dataset indicate that SWNet outperforms ten state-of-the-art methods. The study demonstrates that the integration of cross-spectral data and boundary-guided refinement is essential for high segmentation accuracy in complex crop canopies. The code is available on GitHub: https://cod-espol.github.io/SWNet/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces SWNet, a bimodal end-to-end cross-spectral network for camouflaged weed detection in dense agricultural environments. It uses a Pyramid Vision Transformer v2 (PVTv2) backbone to capture long-range dependencies, a Bimodal Gated Fusion Module to integrate visible and near-infrared (NIR) channels by exploiting chlorophyll reflectance differences, and an Edge-Aware Refinement module for sharper boundaries. The central claim is that SWNet outperforms ten state-of-the-art methods on the Weeds-Banana dataset.

Significance. If the empirical claims hold with proper validation, the work targets a practical challenge in precision agriculture where homochromatic camouflage hinders robotic weed detection. The cross-spectral fusion strategy and public code release are strengths that could support follow-on research in multimodal agricultural vision. However, the absence of supporting metrics and analyses limits immediate impact assessment.

major comments (3)
  1. [Abstract] Abstract: The assertion that 'Experimental results on the Weeds-Banana dataset indicate that SWNet outperforms ten state-of-the-art methods' is made without any reported quantitative metrics (e.g., mIoU, F1, or pixel accuracy), comparison tables, ablation results, or statistical tests, preventing evaluation of the central empirical claim.
  2. [Introduction and Experiments] Motivation and Experiments: No per-class VIS vs. NIR separability statistics, qualitative examples showing camouflage failure in VIS but success in NIR, or ablation isolating the NIR channel and Bimodal Gated Fusion Module contribution are provided; this leaves open whether gains derive from cross-spectral discrimination or from the PVTv2 backbone and edge refinement alone, undermining the core physiological motivation.
  3. [Experiments] Experiments: No dataset statistics (image count, train/test split, class balance in Weeds-Banana), training hyperparameters, or implementation details are reported, rendering the outperformance claim non-reproducible and unevaluable.
minor comments (2)
  1. [Abstract] The GitHub link contains a possible typo ('cod-espol' instead of a standard organization name); verify and correct for accessibility.
  2. [Method] Notation for the Bimodal Gated Fusion Module and Edge-Aware Refinement could be formalized with equations or pseudocode to improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments that identify key areas where additional detail will strengthen the presentation of our results and motivation. We have revised the manuscript accordingly to address each point.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The assertion that 'Experimental results on the Weeds-Banana dataset indicate that SWNet outperforms ten state-of-the-art methods' is made without any reported quantitative metrics (e.g., mIoU, F1, or pixel accuracy), comparison tables, ablation results, or statistical tests, preventing evaluation of the central empirical claim.

    Authors: We agree that the abstract should include quantitative support. In the revised manuscript we have added the key metrics (mIoU and F1-score) to the abstract and expanded the experiments section with full comparison tables, ablation results, and statistical tests so that the outperformance claim can be directly evaluated. revision: yes

  2. Referee: [Introduction and Experiments] Motivation and Experiments: No per-class VIS vs. NIR separability statistics, qualitative examples showing camouflage failure in VIS but success in NIR, or ablation isolating the NIR channel and Bimodal Gated Fusion Module contribution are provided; this leaves open whether gains derive from cross-spectral discrimination or from the PVTv2 backbone and edge refinement alone, undermining the core physiological motivation.

    Authors: We acknowledge the value of explicit evidence for the physiological motivation. The revision now contains per-class VIS/NIR separability statistics, qualitative examples that illustrate camouflage failure in the visible spectrum and its resolution in NIR, and an ablation study that isolates the contribution of the NIR channel together with the Bimodal Gated Fusion Module. These additions demonstrate that the performance gains are attributable to cross-spectral fusion rather than the backbone or edge refinement alone. revision: yes

  3. Referee: [Experiments] Experiments: No dataset statistics (image count, train/test split, class balance in Weeds-Banana), training hyperparameters, or implementation details are reported, rendering the outperformance claim non-reproducible and unevaluable.

    Authors: We thank the referee for highlighting the reproducibility gap. The revised manuscript includes a new subsection with complete Weeds-Banana dataset statistics (image counts, train/test splits, class balance), the full set of training hyperparameters, and implementation details (optimizer, learning-rate schedule, hardware, and code repository). revision: yes

Circularity Check

0 steps flagged

No significant circularity in empirical architecture proposal and evaluation

full rationale

The manuscript proposes the SWNet architecture (PVTv2 backbone + bimodal gated fusion + edge refinement) and reports empirical outperformance on the Weeds-Banana dataset against ten external baselines. No derivation chain, equations, or first-principles predictions exist that reduce to fitted inputs by construction. The physiological NIR reflectance motivation is stated as domain knowledge rather than derived within the paper, and no self-citation load-bearing steps or self-definitional loops are present. The central claim remains an independent empirical comparison on held-out data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard deep-learning assumptions that transformers capture long-range dependencies and that NIR reflectance differences are discriminative for chlorophyll-based camouflage; no explicit free parameters, axioms, or invented physical entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5494 in / 1038 out tokens · 34020 ms · 2026-05-10T08:04:54.516158+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Camouflaged object detection via context- aware cross-level fusion.IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6981–6993, 2022

    Geng Chen, Si-Jie Liu, Yu-Jia Sun, Ge-Peng Ji, Ya-Feng Wu, and Tao Zhou. Camouflaged object detection via context- aware cross-level fusion.IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6981–6993, 2022. 3, 6

    work page 2022

  2. [2]

    Boundary-guided network for camou- flaged object detection.Knowledge-based systems, 248: 108901, 2022

    Tianyou Chen, Jin Xiao, Xiaoguang Hu, Guofeng Zhang, and Shaojie Wang. Boundary-guided network for camou- flaged object detection.Knowledge-based systems, 248: 108901, 2022. 3, 6, 7

    work page 2022

  3. [3]

    A tutorial on the cross-entropy method.Annals of operations research, 134:19–67, 2005

    Pieter-Tjerk De Boer, Dirk P Kroese, Shie Mannor, and Reuven Y Rubinstein. A tutorial on the cross-entropy method.Annals of operations research, 134:19–67, 2005. 4

    work page 2005

  4. [4]

    Camouflaged object de- tection

    Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object de- tection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777–2787,

    work page

  5. [5]

    Concealed object detection.IEEE transactions on pat- tern analysis and machine intelligence, 44(10):6024–6042,

    Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection.IEEE transactions on pat- tern analysis and machine intelligence, 44(10):6024–6042,

    work page

  6. [6]

    Cross-modal feature fusion for field weed mapping using rgb and near-infrared imagery.Agriculture, 14(12):2331, 2024

    Xijian Fan, Chunlei Ge, Xubing Yang, and Weice Wang. Cross-modal feature fusion for field weed mapping using rgb and near-infrared imagery.Agriculture, 14(12):2331, 2024. 1

    work page 2024

  7. [7]

    Res2net: A new multi-scale backbone architecture.IEEE transactions on pattern analysis and machine intelligence, 43(2):652–662,

    Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture.IEEE transactions on pattern analysis and machine intelligence, 43(2):652–662,

    work page

  8. [8]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InConf. on Computer Vision and Pattern Recognition, pages 770–778,

    work page

  9. [9]

    High-resolution it- erative feedback network for camouflaged object detection

    Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution it- erative feedback network for camouflaged object detection. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 881–889, 2023. 3, 6, 7

    work page 2023

  10. [10]

    Deep gradient learn- ing for efficient camouflaged object detection.Machine In- telligence Research, 20(1):92–108, 2023

    Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learn- ing for efficient camouflaged object detection.Machine In- telligence Research, 20(1):92–108, 2023. 3, 6

    work page 2023

  11. [11]

    Magnet: A camouflaged object detec- tion network simulating the observation effect of a magnifier

    Xinhao Jiang, Wei Cai, Zhili Zhang, Bo Jiang, Zhiyong Yang, and Xin Wang. Magnet: A camouflaged object detec- tion network simulating the observation effect of a magnifier. Entropy, 24(12):1804, 2022. 2

    work page 2022

  12. [12]

    Scale-aware modulation meet transformer

    Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lian- wen Jin. Scale-aware modulation meet transformer. InInt. Conf. on Computer Vision, pages 6015–6026, 2023. 3

    work page 2023

  13. [13]

    Modeling aleatoric uncertainty for camouflaged object detection

    Jiawei Liu, Jing Zhang, and Nick Barnes. Modeling aleatoric uncertainty for camouflaged object detection. InProceed- ings of the IEEE/CVF winter conference on applications of computer vision, pages 1445–1454, 2022. 3, 6

    work page 2022

  14. [14]

    Deep- roadmapper: Extracting road topology from aerial images

    Gell ´ert M ´attyus, Wenjie Luo, and Raquel Urtasun. Deep- roadmapper: Extracting road topology from aerial images. InInt. Conf. on Computer Vision, pages 3438–3446, 2017. 4

    work page 2017

  15. [15]

    Weed detection using deep learning: A systematic literature review.Sensors, 23(7): 3670, 2023

    Nafeesa Yousuf Murad, Tariq Mahmood, Abdur Rahim Mo- hammad Forkan, Ahsan Morshed, Prem Prakash Jayaraman, and Muhammad Shoaib Siddiqui. Weed detection using deep learning: A systematic literature review.Sensors, 23(7): 3670, 2023. 2

    work page 2023

  16. [16]

    Plant camouflage: ecology, evolution, and implications.Trends in ecology & evolution, 33(8):608–618, 2018

    Yang Niu, Hang Sun, and Martin Stevens. Plant camouflage: ecology, evolution, and implications.Trends in ecology & evolution, 33(8):608–618, 2018. 1

    work page 2018

  17. [17]

    Zoom in and out: A mixed-scale triplet network for camouflaged object detection

    Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. InConf. on Com- puter Vision and Pattern Recognition, pages 2160–2170,

    work page

  18. [18]

    deepnir: Datasets for generating synthetic nir im- ages and improved fruit detection system using deep learning techniques.Sensors, 22(13):4721, 2022

    Inkyu Sa, Jong Yoon Lim, Ho Seok Ahn, and Bruce Mac- Donald. deepnir: Datasets for generating synthetic nir im- ages and improved fruit detection system using deep learning techniques.Sensors, 22(13):4721, 2022. 1, 2

    work page 2022

  19. [19]

    Edge-aware mirror network for camouflaged object detection

    Dongyue Sun, Shiyao Jiang, and Lin Qi. Edge-aware mirror network for camouflaged object detection. In2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2465–2470. IEEE, 2023. 3, 6

    work page 2023

  20. [20]

    Context-aware cross-level fusion network for camouflaged object detection

    Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. InIJCAI, pages 1025–1031, 2021. 4

    work page 2021

  21. [21]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR,

    work page

  22. [22]

    Unveiling the hidden: Early detection of in- vasive vegetation in crops with uav multispectral imaging

    Henry O Velesaca, Andrea Mero, Hector Villegas, and An- gel D Sappa. Unveiling the hidden: Early detection of in- vasive vegetation in crops with uav multispectral imaging. Smart Agricultural Technology, page 101875, 2026. 5, 6, 8

    work page 2026

  23. [23]

    Cross-modal oriented object detection of uav aerial images based on im- age feature.IEEE Transactions on Geoscience and Remote Sensing, 62:1–21, 2024

    Huiying Wang, Chunping Wang, Qiang Fu, Dongdong Zhang, Renke Kou, Ying Yu, and Jian Song. Cross-modal oriented object detection of uav aerial images based on im- age feature.IEEE Transactions on Geoscience and Remote Sensing, 62:1–21, 2024. 2

    work page 2024

  24. [24]

    Assisted refinement network based on channel information interaction for camouflaged object de- tection

    Kuan Wang, Xiuhong Li, Yulong Bai, Songlin Li, Mengge Lu, and Zhenhong Jia. Assisted refinement network based on channel information interaction for camouflaged object de- tection. InInt. Conf. on Multimedia Retrieval, pages 2058– 2062, 2025. 3, 6, 7

    work page 2058

  25. [25]

    Efficient camouflaged object detection network based on channel reconstruction and hy- brid attention

    Kuan Wang, Xiuhong Li, Songlin Li, Yulong Bai, Boyuan Li, Mengge Lu, and Zhenhong Jia. Efficient camouflaged object detection network based on channel reconstruction and hy- brid attention. InInt. Conf. on Multimedia Retrieval, pages 2063–2067, 2025. 3, 6, 7

    work page 2063

  26. [26]

    Assisted refinement network based on channel information interaction for camouflaged and salient object detection.arXiv preprint arXiv:2512.11369, 2025

    Kuan Wang, Yanjun Qin, Mengge Lu, Liejun Wang, and Xi- aoming Tao. Assisted refinement network based on channel information interaction for camouflaged and salient object detection.arXiv preprint arXiv:2512.11369, 2025. 3, 6, 7

    work page arXiv 2025

  27. [27]

    Pvt v2: Improved baselines with pyramid vision transformer

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415–424, 2022. 2, 3, 4

    work page 2022

  28. [28]

    F3net: fusion, feedback and focus for salient object detection

    Jun Wei, Shuhui Wang, and Qingming Huang. F3net: fusion, feedback and focus for salient object detection. InProceed- ings of the AAAI conference on artificial intelligence, pages 12321–12328, 2020. 4

    work page 2020

  29. [29]

    Color indices for weed identi- fication under various soil, residue, and lighting conditions

    David M Woebbecke, George E Meyer, Kenneth V on Bar- gen, and David A Mortensen. Color indices for weed identi- fication under various soil, residue, and lighting conditions. Transactions of the ASAE, 38(1):259–269, 1995. 2

    work page 1995

  30. [30]

    Plantcamo: Plant camouflage detection,

    Jinyu Yang, Qingwei Wang, Feng Zheng, Peng Chen, Ale ˇs Leonardis, and Deng-Ping Fan. Plantcamo: Plant camou- flage detection.arXiv preprint arXiv:2410.17598, 2024. 1

    work page arXiv 2024

  31. [31]

    Detecting camouflaged object in fre- quency domain

    Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in fre- quency domain. InConf. on Computer Vision and Pattern Recognition, pages 4504–4513, 2022. 2

    work page 2022