Interpretable multi-label learning enables complete-site screening of remediation technologies from historical engineering cases
Authors: Guo, Tianfeng; Xu, Jiayang; Zhu, Lei; Fang, Tingting; Zhang, Ruoxuan; Wang, Yong; Hu, Dapeng; Lyu, Tong; Wen, Guoguang; Zhou, Jianan; Kang, Yulin
Source journal: Environmental research (2026-11-17)
DOI: 10.1016/j.envres.2026.126223
Abstract
Selecting remediation technologies for contaminated sites is a systems-level decision involving contaminant mixtures, risk intensity, spatial extent, medium properties and management constraints. Current data-driven studies focus on contaminant mapping, risk identification or single-process performance, lacking whole-site multi-technology screening. Here we examined whether pre-selection information from historical engineering projects could retrieve documented remediation functions. We curated 168 soil and 105 groundwater remediation projects from traceable Chinese engineering records as complete-site samples, using pre-selection information as predictors and historically implemented technologies including seven soil and four groundwater functional classes as multilabel outcomes. For soil, seven statistical and machine-learning approaches were evaluated under identical multilabel-stratified fivefold validation with inner-fold hyperparameter tuning and benchmarked against a feature-free historical-frequency comparator. Out-of-fold SHAP attribution, incremental information-set analysis and contamination-complexity-dependent Top-k recovery were subsequently used. All seven achieved higher Macro-AP than the historical-frequency comparator, with Random Forest achieving 0.594 versus 0.276; mean historical-technology coverage increased from 39.8% at Top-1 to 61.8% at Top-2 and 76.4% at Top-3, while exploratory groundwater validation yielded a Macro-AP of 0.668 versus 0.419. SHAP attribution showed that medium/hydrogeology and contaminant composition contributed 29.3% and 25.8%, respectively, while the other three information groups together contributed 44.9%. Historical technology portfolios broadened systematically with contamination complexity, increasing from 1.29 functional classes at single-contaminant sites to 2.34 at organic-inorganic co-contaminated sites, while Top-3 completely recovered the historical technology set for 58.9% of soil sites. Candidate width therefore emerges as a site-dependent screening variable, while ranked candidates still require technology-specific evaluation of feasibility, process compatibility, residual risk and implementation constraints before scheme development. By converting fragmented historical remediation practice into auditable complete-site evidence, this framework establishes a reproducible interface between data-driven inference and engineering judgment for transparent decision support in complex contaminated-site management.
Related topics
Related open access articles
- Prevalent Remediation Technologies for Organic Contaminants
- INDENTIFICATION OF HYDROGEOCHEMICAL PROCESS OF GROUNDWATER IN AND AROUND KANIGIRI, PODILI, ONGOLE AREAS OF PRAKASAM DISTRICT, ANDHRA PRADESH, INDIA.
- Removing of Hardness Salts from Groundwater by Thermogenic Synthesis Zeolite
- Assessment of Groundwater Quality Parameters in Northern Region of Zanzibar Island
- Removing of Hardness Salts from Groundwater by Thermogenic Synthesis Zeolite
- Phytoremediation of contaminated soil Lead and Cadmium by Brassica júncea (L.) Czern plant
Abstract and bibliographic data from PubMed (U.S. National Library of Medicine). Copyright of the abstract belongs to the original publisher or authors.