{"title":"DATerra: Boundary and structure aware domain adaptation for cross-regional terraced parcel extraction","authors":"Zhiwei Zhang, Xingyuan Zeng, Xinkai Kong, Juepeng Zheng, Haoyuan Liang, Guowei Zhou, Bohan Shi, Lingyuan Zhao, Jianxi Huang, Haohuan Fu","doi":"10.1016/j.jag.2026.105579","DOIUrl":null,"url":null,"abstract":"<div><div>Irregular agricultural parcels, particularly terraced fields, exhibit complex geometries and fine-scale boundaries that are highly sensitive to cross-regional variations in imaging conditions and landscape patterns. Although unsupervised domain adaptation (UDA) has shown promise for remote sensing segmentation under domain shifts, most existing methods primarily focus on semantic alignment while failing to adequately preserve boundary geometry and parcel-level structural consistency, often leading to boundary adhesion and distorted parcel delineation in target domains. To address these limitations, we propose DATerra, a boundary and structure aware domain adaptation framework for cross-regional terraced parcel extraction. The framework couples two architecture-level boundary-modeling modules with two training-level structure-preserving strategies: (i) Boundary-Guided Attention Fusion (BGAF), which injects geometry-aware cues into multi-scale feature learning; (ii) Boundary-aware Uncertainty-guided Gating (BUDG), which selectively refines ambiguous boundary regions while preserving reliable parcel interiors; and (iii) Parcel Structural Consistency (PSC), together with Structure-Aware ClassMix (SA-ClassMix), which promotes stable parcel predictions under partial target-domain observations and preserves parcel–boundary structural units during cross-domain mixing. Extensive experiments on the GTPBD-China benchmark, covering four geographic regions and twelve cross-domain transfer scenarios, demonstrate that DATerra substantially improves boundary delineation and object-level parcel integrity. On average, DATerra improves OIS and ODS by 14.00 and 10.76 percentage points, respectively, while reducing GUC and GTC by 20.14 and 9.98 points compared with the runner-up methods. Pixel-level metrics are also consistently improved, with an average gain of 1.23%. These results highlight the importance of explicitly modeling boundary geometry and structural consistency for robust cross-regional terraced parcel extraction.</div></div>","PeriodicalId":73423,"journal":{"name":"International journal of applied earth observation and geoinformation : ITC journal","volume":"153 ","pages":"Article 105579"},"PeriodicalIF":8.2000,"publicationDate":"2026-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"International journal of applied earth observation and geoinformation : ITC journal","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1569843226004954","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2026/9/15 0:00:00","PubModel":"Epub","JCR":"Q1","JCRName":"REMOTE SENSING","Score":null,"Total":0}
引用次数: 0
Abstract
Irregular agricultural parcels, particularly terraced fields, exhibit complex geometries and fine-scale boundaries that are highly sensitive to cross-regional variations in imaging conditions and landscape patterns. Although unsupervised domain adaptation (UDA) has shown promise for remote sensing segmentation under domain shifts, most existing methods primarily focus on semantic alignment while failing to adequately preserve boundary geometry and parcel-level structural consistency, often leading to boundary adhesion and distorted parcel delineation in target domains. To address these limitations, we propose DATerra, a boundary and structure aware domain adaptation framework for cross-regional terraced parcel extraction. The framework couples two architecture-level boundary-modeling modules with two training-level structure-preserving strategies: (i) Boundary-Guided Attention Fusion (BGAF), which injects geometry-aware cues into multi-scale feature learning; (ii) Boundary-aware Uncertainty-guided Gating (BUDG), which selectively refines ambiguous boundary regions while preserving reliable parcel interiors; and (iii) Parcel Structural Consistency (PSC), together with Structure-Aware ClassMix (SA-ClassMix), which promotes stable parcel predictions under partial target-domain observations and preserves parcel–boundary structural units during cross-domain mixing. Extensive experiments on the GTPBD-China benchmark, covering four geographic regions and twelve cross-domain transfer scenarios, demonstrate that DATerra substantially improves boundary delineation and object-level parcel integrity. On average, DATerra improves OIS and ODS by 14.00 and 10.76 percentage points, respectively, while reducing GUC and GTC by 20.14 and 9.98 points compared with the runner-up methods. Pixel-level metrics are also consistently improved, with an average gain of 1.23%. These results highlight the importance of explicitly modeling boundary geometry and structural consistency for robust cross-regional terraced parcel extraction.
期刊介绍:
The International Journal of Applied Earth Observation and Geoinformation publishes original papers that utilize earth observation data for natural resource and environmental inventory and management. These data primarily originate from remote sensing platforms, including satellites and aircraft, supplemented by surface and subsurface measurements. Addressing natural resources such as forests, agricultural land, soils, and water, as well as environmental concerns like biodiversity, land degradation, and hazards, the journal explores conceptual and data-driven approaches. It covers geoinformation themes like capturing, databasing, visualization, interpretation, data quality, and spatial uncertainty.