Characterizing Image-Derived Bridge–Environment Color Appearance in a Linear Cultural Heritage Corridor Using Semantic Segmentation

The Beijing–Hangzhou Grand Canal

Bibliographic Data

ID24218786
AuthorsYaru Liu (0000-0003-3178-6352, School of Art and Design, Nanjing Forestry University, Nanjing 210037, China), Xingye Yang (0009-0004-8736-8592, School of Art and Design, Nanjing Forestry University, Nanjing 210037, China), Ruixia Wang (0009-0001-1046-4787, School of Art and Design, Nanjing Forestry University, Nanjing 210037, China)
Year2026
Volume9
Issue10
Pages387
Publication date2026-09-24
Open AccessYes
TypeARTICLE
VenueHeritage (JOURNAL)
Journal identifiersISSN: 2571-9408 • E-ISSN: 2571-9408
PublisherMDPI AG (PUBLISHER • IT)
DOI10.3390/heritage9100387
OpenAlexW7214164892
LanguageEN
References cited35

Bridges are recurrent visual elements within linear cultural heritage corridors, but their recorded color appearance and within-scene relationships with water and vegetation remain insufficiently characterized at corridor scale. This study develops a semantic-segmentation-based framework for analyzing image-derived bridge–environment color appearance in publicly accessible photographs of the Beijing–Hangzhou Grand Canal. From 410 images, 393 unique bridges were identified; after duplicate control, image-quality screening, and mask validation, 356 unique bridges represented by one high- or medium-quality image each were retained for the primary analysis. Mask2Former was used to extract bridge, water, and vegetation regions. Bridge masks were manually corrected, and environmental masks from a stratified 48-image subset were expert-reviewed for validation and sensitivity analysis. Bridge-body appearance was quantified using Saturation, Value, saturation-thresholded circular Hue, saturation-weighted Hue Concentration Index, Colorfulness, Color Richness, dominant-color clustering, and CIELAB/CIEDE2000 metrics. All inferential analyses were conducted at the unique-bridge level. Construction period and primary structural material were not consistently documented across the public-image dataset and were therefore not used as classification variables; the analysis concerns recorded bridge appearance rather than material- or period-specific heritage color. Bridge bodies displayed a muted palette dominated by low-saturation gray-white, gray-blue, gray-brown, and gray-green tones. Sectional differences were most consistently expressed in circular Hue, Value, and Colorfulness; Saturation was weak and non-robust, while Hue Concentration Index and pixel-controlled color-diversity measures showed no significant sectional differences. Median within-image CIEDE2000 distance was lower for bridge–water than for bridge–vegetation comparisons (15.58 versus 20.00). Bridges were generally less saturated but slightly brighter than their combined environmental backgrounds. Bridge and environmental indicators showed weak-to-moderate positive co-variation, interpreted only as within-image association because all semantic regions shared the same acquisition and processing conditions. The principal findings remained stable across image-quality subsets, Saturation and valid-region thresholds, equal-pixel sampling, alternative diversity measures, and expert-reviewed mask substitution. The framework offers a reproducible tool for preliminary heritage-image documentation and corridor-scale pattern identification, but it does not provide calibrated measurements of material color or evidence of causal environmental effects.

canal bridges · CIEDE2000 color difference · Grand Canal · Mask2Former · semantic segmentation · Color perception and design · Color Science and Applications · Cultural Heritage Management and Preservation

Tools

Open DOIOpen Access