Forward-Facing Near-Infrared Adds Little to Colour for Farm-Machinery Traversability: A Site-Disjoint Evaluation of Sensor-Dependent Spatial Leakage

Sungwoo Kang

cs.CVSep 7, 2026

Forward-Facing Near-Infrared Adds Little to Colour for Farm-Machinery Traversability: A Site-Disjoint Evaluation of Sensor-Dependent Spatial Leakage

Authors: Sungwoo Kang

Organizations: Department of Electrical and Computer Engineering, Korea University, Seoul, Republic of Korea

Abstract

Near-infrared (NIR) imaging does not consistently outperform standard color cameras for daytime agricultural traversability once spatial data leakage is eliminated. Prior benchmarks suggesting an NIR advantage used sequence-level splits that permitted spatially autocorrelated imagery into test sets, artificially inflating NIR performance, especially on difficult paddy-boundary segmentation. When evaluated across strictly held-out recording sites using the AI Hub autonomous driving corpus, none of the four tested configurations (color, NIR, a luminance control, or their fusion) reliably surpasses standard color across any of the five traversability classes. Furthermore, apparent performance gains observed at a single held-out site completely fail to replicate when the held-out site is rotated, demonstrating that the observed benefits were site-specific artifacts rather than generalizable improvements. Correcting for spatial leakage erases the apparent lead of NIR entirely rather than uniformly degrading performance across all sensor types. Consequently, adding a forward-facing NIR camera to a daylight agricultural vehicle sensor suite remains unproven, and evaluations compromised by location leakage risk distorting sensor rankings and misleading procurement decisions.

Figures & tables

	Site-blocked IoU				Leakage inflation
Class	Colour	Lum.	NIR	Fusion	Colour	Lum.	NIR	Fusion
Paddy edge	$0.059_{\pm 0.043}$	$0.011_{\pm 0.017}$	$0.015_{\pm 0.016}$	$0.035_{\pm 0.016}$	$-0.008$	$+0.027$	$+0.067$	$+0.007$
Paddy, untilled	$0.067_{\pm 0.016}$	$0.065_{\pm 0.013}$	$0.100_{\pm 0.018}$	$0.066_{\pm 0.005}$	$+0.051$	$+0.034$	$+0.092$	$+0.060$
Paddy, tilled	$0.507_{\pm 0.025}$	$0.388_{\pm 0.037}$	$0.418_{\pm 0.030}$	$0.553_{\pm 0.030}$	$+0.080$	$+0.090$	$+0.104$	$+0.044$
Field furrow	$0.344_{\pm 0.013}$	$0.253_{\pm 0.029}$	$0.203_{\pm 0.014}$	$0.239_{\pm 0.037}$	$+0.033$	$+0.060$	$+0.047$	$+0.094$
Field levee	$0.538_{\pm 0.037}$	$0.475_{\pm 0.022}$	$0.379_{\pm 0.013}$	$0.539_{\pm 0.021}$	$+0.032$	$+0.044$	$+0.129$	$+0.044$

Table 1: Segmentation accuracy under the site-disjoint protocol, beside the leakage inflation the sequence-level split admits. Accuracies are the intersection-over-union on the intersection ground truth, as a mean over five seeds with its standard deviation, for the four input configurations. The inflation is not the difference of the two blocks. It is the per-frame paired difference between the sequence-level-split and site-blocked models over the frames both protocols scored (Section 3 ), whereas the two blocks are scored on different frames. Near-infrared carries the largest inflation on 5 of the 6 classes shown. The exception is the field furrow, where fusion is larger. The sequence-level split’s own accuracies, the in-plot-tree control that the site-blocked split leaves with nine validation frames, and the non-target standing-rice class are in Table 1 .

Figure 1: Paddy-edge accuracy on a fixed set of frames from one held-out super-site, scored by two models that differ only in whether they saw the super-site during training. Bars give the pooled intersection-over-union on the intersection ground truth for colour, the luminance control, and near-infrared. Near-infrared gains most from having seen the site and colour least, with the luminance control between them. The frames are identical and neither model trained on them. The gap is recognition of the site rather than the site being harder to segment.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	Site-blocked				Sequence-level split
Class	Colour	Lum.	NIR	Fusion	Colour	Lum.	NIR	Fusion
Paddy edge	$0.059_{\pm 0.043}$	$0.011_{\pm 0.017}$	$0.015_{\pm 0.016}$	$0.035_{\pm 0.016}$	$0.009_{\pm 0.005}$	$0.025_{\pm 0.014}$	$0.339_{\pm 0.025}$	$0.055_{\pm 0.055}$
Paddy, untilled	$0.067_{\pm 0.016}$	$0.065_{\pm 0.013}$	$0.100_{\pm 0.018}$	$0.066_{\pm 0.005}$	$0.235_{\pm 0.047}$	$0.131_{\pm 0.023}$	$0.239_{\pm 0.029}$	$0.237_{\pm 0.033}$
Paddy, tilled	$0.507_{\pm 0.025}$	$0.388_{\pm 0.037}$	$0.418_{\pm 0.030}$	$0.553_{\pm 0.030}$	$0.525_{\pm 0.044}$	$0.491_{\pm 0.032}$	$0.747_{\pm 0.040}$	$0.536_{\pm 0.061}$
Field furrow	$0.344_{\pm 0.013}$	$0.253_{\pm 0.029}$	$0.203_{\pm 0.014}$	$0.239_{\pm 0.037}$	$0.366_{\pm 0.016}$	$0.310_{\pm 0.022}$	$0.311_{\pm 0.017}$	$0.332_{\pm 0.014}$
Field levee	$0.538_{\pm 0.037}$	$0.475_{\pm 0.022}$	$0.379_{\pm 0.013}$	$0.539_{\pm 0.021}$	$0.661_{\pm 0.014}$	$0.615_{\pm 0.022}$	$0.607_{\pm 0.016}$	$0.658_{\pm 0.006}$

Appendix

Table 1: Per-class segmentation accuracy under the site-blocked protocol and the sequence-level split, for every annotated class the runs scored. Each cell is the intersection-over-union on the intersection ground truth, as a mean over five seeds with its standard deviation. The site-blocked block repeats Table 1 . The sequence-level-split block is what the leakage inflation there is measured against. In-plot trees are not evaluable under the site-blocked split, which retains nine validation frames of that class. Standing rice is neither a traversability target nor a control and is reported here only for completeness.

Trained on rice paddy alone
	Evaluated on rice paddy		Evaluated on dry field
Class	Colour	NIR	Colour	NIR
Paddy edge	$0.207_{\pm 0.096}$	$0.079_{\pm 0.022}$	—	—
Paddy, untilled	$0.000_{\pm 0.000}$	$0.000_{\pm 0.000}$	—	—
Paddy, tilled	$0.685_{\pm 0.031}$	$0.585_{\pm 0.022}$	$0.870_{\pm 0.031}$	$0.649_{\pm 0.055}$
Field furrow	$0.115_{\pm 0.039}$	$0.000_{\pm 0.000}$	$0.024_{\pm 0.016}$	$0.000_{\pm 0.000}$

Appendix

Table 1: Intersection-ground-truth accuracy of the configurations trained on a single driving environment, evaluated on both environments of the same held-out super-site, as a mean over five seeds with its standard deviation. The block a row belongs to names the training environment, so the column matching it is in-domain and the other is transfer. Trained on one environment and evaluated on the other, colour leads near-infrared on every target cell either configuration scores above zero on. Near-infrared leads on the intersection ground truth on the field furrow and field levee in domain, and the ordering reverses on the own-label ground truth in every one, so no cell here is read as a configuration advantage. These runs carry no criterion of their own and the accuracies are read as descriptive. A dash marks a class with fewer than

50

held-out frames in that environment.

Model	Colour	Lum.	NIR	Fusion	95% CI, NIR
MiT-B0 (primary)	$-0.008$	$+0.027$	$+0.067$	$+0.007$	$[+0.001,\,+0.135]$
MiT-B2	$+0.042$	$+0.041$	$+0.100$	$+0.046$	$[+0.034,\,+0.175]$
UPerNet/ConvNeXt-V2-T	$+0.038$	$+0.029$	$+0.069$	$+0.032$	$[+0.001,\,+0.157]$

Appendix

Table 1: Encoder size and architecture family. Both panels are read on the frames the site-blocked split holds out, so the model varies and the held-out site does not. What they test is whether an ordering depends on the encoder or the decoder, not whether it holds at another site. (a) The paddy-edge leakage inflation, the per-frame paired difference between the sequence-level-split and site-blocked models over the

322

frames both protocols scored, for the four input configurations. Near-infrared carries the largest inflation on all three models. The interval is on the near-infrared cell, the one cell of the panel a protocol criterion is declared on. The others are descriptive. (b) The near-infrared accuracy margin over the capacity-matched luminance control on the tilled paddy surface under the site-blocked protocol. The margin spans zero on all three models, and has the opposite sign on UPerNet/ConvNeXt-V2-T. Intervals are sequence-clustered 95% bootstrap intervals.

	Site-identification accuracy
Configuration	Frame brightness	Supervising annotation
Near-infrared	$0.440$	$0.646$
Luminance control	$0.534$	$0.640$
Majority baseline	$0.472$	$0.472$

Appendix

Table 1: The three mechanisms that would explain the concentration of leakage in near-infrared without any appeal to the spectral band, all measured from the frames without training anything. (a) Super-site identified by one nearest-centroid classifier, scored by holding out whole driving sequences over the

15\,790

frames of

94

sequences, from the mean and standard deviation of the frame and from the class composition of the annotation the configuration is supervised by. The luminance control’s row reads the colour annotation, which supervises colour and fusion too. Near-infrared does not exceed the majority baseline on brightness and luma does, so near-infrared brightness carries no site signature colour brightness lacks. Both annotations exceed that baseline by a wide margin and differ from each other by

+0.006

(95% CI

[-0.009,\,+0.021]

), so the annotation supervising near-infrared marks the site no more strongly than the one supervising the other three. (b) Median agreement between the two paddy-edge annotations, as scored and after each frame’s measured camera offset is removed, against the same agreement at the control’s offset, and with the share of annotated edge pixels the intersection keeps. The two cameras are a median

7

px apart vertically at the working resolution, against

1

px for the same estimator on frames paired across driving sequences, which is what it reports from no alignment at all. The control’s offset is read on the corpus sample, where alignment is claimed to recover something. Alignment lifts the corpus sample and leaves both comparison sets below even its unaligned agreement. The corpus sample is stratified over the whole corpus and is not what any comparison is scored on. The other two rows are the frame sets a comparison is scored on, and they are eroded to within a point of each other, which is what lets an ordering that agrees across both be read as more than an annotation artifact.

Agricultural Image Analysis

Agricultural image analysis uses computer vision and multimodal learning to monitor crops, livestock, produce quality, and field conditions.

See all 94 papers

TEA-AgriVLN: Traversability Estimation Alarm for Agricultural Vision-and-Language Navigation
Jul 30, 2026Xiaobei Zhao, Xingqi Lyu, Xin Chen +1Embodied Navigation Vision-Language Navigation
China Agricultural University · China Agricultural University-Sichuan Advanced Agricultural & Industrial Institute
Enabling 24-hour Agricultural Robotics: Unsupervised Day-to-Night Cross-Modal Image Translation for Nighttime Visual Navigation
Jul 13, 2026Robel Mamo, Rajitha de Silva, Grzegorz Cielniak +1Image-to-Image Translation Robot Navigation
LaSER Lab, Kennesaw State University, Marietta, GA 30060, USA. · Lincoln Institute for Agri-Food Technology, University of Lincoln, Lincoln LN2 2LG, UK.
HSI-Road Relabeled: Surface-Aware Road-Scene Segmentation
Sep 10, 2026Imad Ali Shah, Imran Mehmood, Enda Ward +3Remote Sensing Image Segmentation Semantic Segmentation
School of Engineering and Ryan Institute, University of Galway, Ireland · Valeo Vision Systems, Tuam, Ireland

Jul 30, 2026J/K move · Enter open · S save

TEA-AgriVLN: Traversability Estimation Alarm for Agricultural Vision-and-Language Navigation

Xiaobei Zhao, Xingqi Lyu, Xin Chen, Xiang Li

China Agricultural University · China Agricultural University-Sichuan Advanced Agricultural & Industrial Institute

Embodied Navigation Vision-Language Navigation Agricultural Robotics VLMs for Robotics Robot Traversability Estimation

Open paper AskPDF arXiv

Vision-and-Language Navigation in Continuous Environments (VLN-CE) requires an agent to follow a natural language instruction, predicting a sequence of low-level actions to navigate a robot from a starting point to a target location. The A2A benchmark and the AgriVLN method pioneeringly extended VLN-CE from indoor scenes to agricultural scenes, while we observed a challenging distinction: In indoor scenes, whether a zone is traversable tends to be clear to classify, such as wood floors are traversable but concrete walls are not. In agricultural scenes, however, this issue tends to be ambiguous, such as an unripe cornfield might be traversable for a robotic dog but might be non-traversable for a human. To address this issue, we propose the TEA module, which estimates the traversability of the camera image, then alarm the decision-maker for rethinking when the predicted action does not align with the traversability map. We integrate it into the AgriVLN backbone to build our TEA-AgriVLN method. When evaluated on A2A, it improves Success Rate (SR) from 0.47 to 0.54 and Navigation Error (NE) from 2.91 m to 2.70 m, showing the state-of-the-art performance in the agricultural VLN-CE domain. We further implement the ablation studies and the case study, discussing the effectiveness and limitations of TEA on different ground categories and scene classes. Code: https://github.com/AlexTraveling/TEA-AgriVLN.

Abstract

Agricultural Image Analysis

Explore similar work

TEA-AgriVLN: Traversability Estimation Alarm for Agricultural Vision-and-Language Navigation

Enabling 24-hour Agricultural Robotics: Unsupervised Day-to-Night Cross-Modal Image Translation for Nighttime Visual Navigation

HSI-Road Relabeled: Surface-Aware Road-Scene Segmentation