Fusing conveyor-belt images with existing in-line weight records for flue-cured tobacco leaf grading: a strict date-held-out evaluation
- 1School of Computer and Electrical Engineering, Hunan University of Arts and Science
- 2Key Laboratory of Hunan Province for Control Technology of Distributed Electric Propulsion Air Vehicle
Abstract
Deep learning models for flue-cured tobacco leaf grading are often evaluated using randomly split datasets, which can place images from the same acquisition day in both training and test sets. This may overestimate performance because batch, lighting, and moisture conditions are shared across the splits. This study examined whether existing in-line weight records can improve image-based grading under a stricter, date-grouped evaluation protocol. We used 25,325 paired conveyor-belt images and weight records collected on 34 acquisition dates during the 2025 season. Acquisition dates were kept separate across the training, validation, and test sets. We compared image-only, weight-only, and four image–weight fusion models. A ResNet-50 classifier achieved 86.2% top-1 accuracy with random splitting but only 43.4% when whole acquisition dates were held out, showing that random splits substantially overstate real-world performance. Weight alone performed poorly and varied strongly across batches. Combining weight with images produced small gains of about 1–2 percentage points; among the four fusion architectures of the main comparison, only late fusion yielded a statistically significant improvement over the image-only model, and two deeper fusion baselines (FiLM feature conditioning and cross-modal attention) added in a follow-up experiment did not exceed it either. Scalar gating for adaptively weighting the two modalities did not reliably identify image-model failures. These results indicate that grading studies should report how data are split, disclose date-level groupings, and distinguish genuine generalization from mere random-seed stability. Overall, in-line weight records provide limited but useful complementary information without requiring additional hardware. Date-grouped evaluation is essential for assessing whether tobacco grading models can generalize across production batches.
Related
- Effects of Sequentially Removing Low-Abundance Taxon Records from Community Matrices on Aquatic Insect Community Evaluation in Mountain Streams.
- Adaptations of the pregnant murine heart to pre-existing hypertension
- Urban Vietnam at the microbiome transition: Co-existing enterotypes reflect the global industrialization patterns
- FAnet: A lightweight end-to-end framework for Fructus Aurantii identification and quality grading
- Breeder age outweighs egg grading method on hatchability and chick quality in broiler hatcheries.
BibTeX
@article{xiong2026fusing,
title = {{Fusing conveyor-belt images with existing in-line weight records for flue-cured tobacco leaf grading: a strict date-held-out evaluation}},
author = {Qi Xiong and Wuke Li and Wenxue Tan},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1970460},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1970460/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Fusing conveyor-belt images with existing in-line weight records for flue-cured tobacco leaf grading: a strict date-held-out evaluation
AU - Qi Xiong
AU - Wuke Li
AU - Wenxue Tan
PY - 2026
DA - 2026/10/26
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1970460
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1970460/full
AB - Deep learning models for flue-cured tobacco leaf grading are often evaluated using randomly split datasets, which can place images from the same acquisition day in both training and test sets. This may overestimate performance because batch, lighting, and moisture conditions are shared across the splits. This study examined whether existing in-line weight records can improve image-based grading under a stricter, date-grouped evaluation protocol. We used 25,325 paired conveyor-belt images and weight records collected on 34 acquisition dates during the 2025 season. Acquisition dates were kept separate across the training, validation, and test sets. We compared image-only, weight-only, and four image–weight fusion models. A ResNet-50 classifier achieved 86.2% top-1 accuracy with random splitting but only 43.4% when whole acquisition dates were held out, showing that random splits substantially overstate real-world performance. Weight alone performed poorly and varied strongly across batches. Combining weight with images produced small gains of about 1–2 percentage points; among the four fusion architectures of the main comparison, only late fusion yielded a statistically significant improvement over the image-only model, and two deeper fusion baselines (FiLM feature conditioning and cross-modal attention) added in a follow-up experiment did not exceed it either. Scalar gating for adaptively weighting the two modalities did not reliably identify image-model failures. These results indicate that grading studies should report how data are split, disclose date-level groupings, and distinguish genuine generalization from mere random-seed stability. Overall, in-line weight records provide limited but useful complementary information without requiring additional hardware. Date-grouped evaluation is essential for assessing whether tobacco grading models can generalize across production batches.
SN - 1664-462X
ER -