Decoding spatial genes of living heritage in traditional villages: TV-RSI-413 and MCPNet

Traditional villages encode settlement-scale “spatial genes”. We present TV-RSI-413, a dataset of 2478 on-site images from 413 Jiangxi villages (ground DSLR, low-altitude oblique UAV, and open street-view), each pixel-labelled into 22 classes. A three-stage audit yields high inter-rater agreement (Cohen’s κ = 0.92). Building on this resource, we propose MCPNet, a multiscale context-perceptual segmentation network designed to preserve cadastral edges and network continuity. Under the official in-domain protocol (20 spatial-gene classes), MCPNet achieves 34.7% mean accuracy and 24.3% mean IoU, outperforming DeepLabV3+ by 5.1 points in mean accuracy and 2.6 points in mean IoU, with higher boundary and connectivity scores. In a practice-facing setting that includes background and occluders, overall accuracy reaches 85.3% and mean IoU 70.7%. We release fixed splits, evaluation scripts, and code for FAIR reuse. The 22 classes further aggregate into functional archetypes linked to heritage values, enabling auditable priorities for conservation, ecological zoning, micro-restoration, parcelisation, and risk-aware governance.