2026-10-06T21:36:52Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33942475 2026-09-19T13:57:01Z category_26998 category_27331 item_type_3 month_year_09_2026

AncientTreeDB zhibo wang (25075945) Other environmental sciences not elsewhere classified Environmental history ancient and notable trees Chinese local gazetteers historical ecology cultural heritage tree survival modeling loss-cause attribution species distribution conservation policy CHANTT v1.0 (Chinese Historical Ancient and Notable Tree Traits) is a machine-learning-ready, multi-table relational dataset of ancient and notable trees (名木古树) recorded in Chinese historical local gazetteers and related documentary sources, covering the period from antiquity to 2024 CE. MOTIVATION. Modern tree-census inventories rarely extend beyond a few decades, which limits the study of long-term tree survival and of the ecological, cultural, military, religious and institutional forces that shape the fate of ancient trees. Chinese local gazetteers (方志), general gazetteers, temple gazetteers, stelae, biji notebooks and official histories preserve several millennia of scattered records on individual trees. This dataset converts that unstructured documentary heritage into a structured, linkable, FAIR-compliant research resource. CONSTRUCTION PIPELINE. (1) Source collection: 142 historical titles. (2) Layout analysis: vertical-text line segmentation, restoration of double-column small characters, removal of seals and illustrations. (3) Layout-aware OCR with rare- and variant-character dictionary enhancement, emitting character-level confidence. (4) LLM-based structured extraction of tree name, species, locality, planting dynasty, legend and loss cause, with per-field confidence. (5) Spatiotemporal alignment: historical administrative units mapped to present-day units; reign years converted to CE years. (6) Semantic de-duplication (SemHash) plus cross-gazetteer same-name merging; overall redundancy removal rate 21.8%. (7) Three-tier verification (GOLD/SILVER/BRONZE) with dual-annotator cross-checking. (8) Automated quality screening and compliance review: coordinates generalised to county level with random perturbation, no personal data, CC BY 4.0 licence. CONTENTS — ten linked CSV tables (UTF-8, ~6.2 MB uncompressed). T1_tree_master.csv — 8,638 tree records x 50 fields (5,297 extant, 3,338 lost, 3 replaced); 61 species; 32 provincial-level units; 17 habitat types; 32-class loss-cause taxonomy nested in 6 families; fields cover species identity (Chinese name, Latin name, family, genus), georeference, habitat tenure, planting dynasty and estimated year, estimated age and age class, protection grade, current status, loss cause and year, legend attributes, cultural value score, mention counts, OCR and extraction confidence, five exposure indices (urban, drought, flood, earthquake, war) and verification tier. T2_gazetteer_mentions.csv — 24,939 mention records linked to T1 by tree_id. T3_lost_trees.csv — 3,338 loss records with documented/inferred status and loss period. T4_legend_custom.csv — 3,616 legend, custom and belief records with protection-effect score. T5_hazard_events.csv — 33 historically documented hazard and warfare events (Huaxian 1556, Tancheng 1668, Haiyuan 1920, Tangshan 1976, Wenchuan 2008 earthquakes; Dingwu Famine; 1931 and 1998 floods; Huayuankou dike breach; Taiping and Anglo-French war damage). T6_climate_geo.csv — 153 city-level aggregations of thermal and precipitation means, survival rate, species richness and five risk indices. T7_policy_events.csv — 38 policy, legal and institutional events, 1952 State Council forest-protection directive to the 2025 Regulations on the Protection of Ancient and Notable Trees (State Council Order No. 800), including negative historical episodes. T8_annotation_quality.csv — 5,200 sampled records of dual-annotator accuracy, agreement, threshold pass rate and field completeness. T9_source_books.csv — 142 source titles with author, dynasty of compilation/printing and document category. T10_ocr_quality.csv — 3,200 sampled pages of layout type, character error rate, layout-analysis accuracy, rare-character recall and reading-order accuracy. The deposit also contains code/ (knowledge base, dataset construction, model training and figure-generation scripts) and results/ (metrics.json, cluster assignments and profiles, feature importance, dimensionality-reduction projections), plus README.md, DATACITE.xml, LICENSE.txt, FILE_INDEX and MD5 checksums. TECHNICAL VALIDATION. A multilayer perceptron (ANT-Net, hidden layers 192-96-48, ReLU, Adam, L2 = 3e-3, early stopping) predicts survival status with ROC-AUC 0.906 (accuracy 0.844, F1 0.876), exceeding logistic regression (0.899), random forest (0.904) and histogram gradient boosting. Spatial and temporal hold-out generalisation averages ~0.89 ROC-AUC across six macro-regions and four dynastic periods. Ablation identifies habitat tenure as the most informative feature family: removing it reduces ROC-AUC from 0.906 to 0.753. Loss-cause attribution reaches 0.56 accuracy over six cause families and 0.65 top-3 accuracy over fine-grained classes. Kaplan-Meier survival analysis, permutation importance and K-means/Ward/DBSCAN clustering outputs are provided in results/. USAGE NOTES. For clue discovery, query T1 by province, habitat type and species jointly and rank by n_source_mentions. For ecological change reconstruction, couple T1 with T5 and T6 to build regional tree-loss intensity series. For modelling, use results/metrics.json as the baseline and either filter verification_tier != 'BRONZE' or weight samples by extract_conf. LIMITATIONS. Gazetteer records are systematically biased toward old, famous and temple-associated trees; BRONZE-tier values are inferred extraction outputs, not individually verified, and must not be used as administrative determinations for individual trees; tree ages are estimates rather than dendrochronological measurements; gazetteers from the humid south survive in greater numbers, producing a documentary survival bias. REUSE AND LICENCE. CC BY 4.0. Findable through a figshare DOI and DataCite metadata; accessible by anonymous HTTPS download; interoperable through CSV/UTF-8 open formats, documented field names, foreign-key linkage on tree_id and Latin names aligned to the Flora of China; reusable through an explicit licence, full provenance (book → juan → page → record), versioning and documentation of methods, limitations and usage. RELATED MATERIAL. Companion Data Descriptor submitted to Scientific Data (Zhibo Wang, Jiangsu Ocean University). Cite as: Wang, Z. CHANTT v1.0: A Multi-Table Dataset of Chinese Ancient and Notable Trees Extracted from Historical Local Gazetteers by AI-Based Text Recognition. figshare, 2026. DOI: 10.6084/m9.figshare.XXXXXXX. Licence: CC BY 4.0. 2026-09-19T13:57:01Z Dataset Dataset 10.6084/m9.figshare.33942475.v2 https://figshare.com/articles/dataset/CHANTT_v1_0/33942475 CC BY 4.0