GTPBD-MM: A Global Terraced Parcel and Boundary Dataset with Multi-Modality

Zhiwei Zhang; Xingyuan Zeng; Xinkai Kong; Kunquan Zhang; Haoyuan Liang; Bohan Shi; Juepeng Zheng; Jianxi Huang

cs.CVApr 14, 2026

GTPBD-MM: A Global Terraced Parcel and Boundary Dataset with Multi-Modality

Authors: Zhiwei Zhang, Xingyuan Zeng, Xinkai Kong, Kunquan Zhang, Haoyuan Liang, Bohan Shi, Juepeng Zheng, Jianxi Huang,

Organizations: Sun Yat-sen University · Northeastern University · National Supercomputing Center in Shenzhen · China Agricultural University ·

Abstract

Agricultural parcel extraction plays an important role in remote sensing-based agricultural monitoring, supporting parcel surveying, precision management, and ecological assessment. However, existing public benchmarks mainly focus on regular and relatively flat farmland scenes. In contrast, terraced parcels in mountainous regions exhibit stepped terrain, pronounced elevation variation, irregular boundaries, and strong cross-regional heterogeneity, making parcel extraction a more challenging problem that jointly requires visual recognition, semantic discrimination, and terrain-aware geometric understanding. Although recent studies have advanced visual parcel benchmarks and image-text farmland understanding, a unified benchmark for complex terraced parcel extraction under aligned image-text-DEM settings remains absent. To fill this gap, we present GTPBD-MM, the first multimodal benchmark for global terraced parcel extraction. Built upon GTPBD, GTPBD-MM integrates high-resolution optical imagery, structured text descriptions, and DEM data, and supports systematic evaluation under Image-only, Image+Text, and Image+Text+DEM settings. We further propose Elevation and Text guided Terraced parcel network (ETTerra), a multimodal baseline for terraced parcel delineation. Extensive experiments demonstrate that textual semantics and terrain geometry provide complementary cues beyond visual appearance alone, yielding more accurate, coherent, and structurally consistent delineation results in complex terraced scenes.

Figures & tables

Figure 1. Overview of terraced parcel extraction challenges and multimodal motivation. The top row illustrates the complementary information provided by image, text, and DEM modalities. The bottom row compares three input settings: the Image-only model (HBGNet ( Zhao et al., 2025 ) ) suffers from semantic confusion (red box) and boundary ambiguity (yellow box); the Image+Text model (LISA ( Lai et al., 2024 ) ) mitigates semantic confusion but still fails to resolve the boundary ambiguity; by jointly modeling image, text, and DEM, our method effectively addresses both issues and yields more complete and structurally consistent parcel delineation results.

Dataset	Research Focus	Pub.	Input	Resolution	Annotation	Coverage	Terraced Scenes
GFSAD30 ( Thenkabail et al., 2021 )	Cropland Mapping	USGS’21	Image	30 m	Mask	Global	No
GTM ( Li et al., 2025 )	Cropland Mapping	JAG’25	Image	10 m	Mask	Global	Yes
AI4Boundaries ( d’Andrimont et al., 2023 )	Parcel Delineation	ESSD’23	Image	10 m, 1 m	Boundary, Vector	Europe	No
FHAPD ( Zhao et al., 2025 )	Parcel Delineation	ISPRS’25	Image	1–2 m	Mask	China	No
FTW ( Kerner et al., 2025 )	Parcel Delineation	AAAI’25	Image	10 m	Mask, Parcel	Global	No
GTPBD ( Zhang et al., 2025 )	Parcel Delineation	NeurIPS’25	Image	0.5–0.7 m	Boundary, Mask, Parcel	Global	Yes

Table 1. Comparison of representative datasets for agricultural parcel analysis in terms of research focus, input modality, spatial resolution, annotation type, geographic coverage, and whether terraced scenes are included.

Figure 2. Overview of GTPBD-MM. Top: unified multimodal design with aligned modalities, hierarchical annotations, and evaluation tasks. Bottom: spatial sampling distribution, with global coverage and a zoom-in view of China.

Figure 3. Dataset statistics of GTPBD-MM. (a) Regional- and Country-level area distribution. (b) Word cloud of text descriptions.

				Pixel-level					Edge-level		Object-level
Model	Method Family	Input	Pub.	Rec. $\uparrow$	F1 $\uparrow$	OA $\uparrow$	mIoU $\uparrow$	mAcc $\uparrow$	OIS $\uparrow$	ODS $\uparrow$	GOC $\downarrow$	GUC $\downarrow$	GTC $\downarrow$
U-Net ( Ronneberger et al., 2015 )	Gen. Sem. Seg.	I	MICCAI’15	57.15	65.42	72.67	55.87	71.32	22.97	15.57	44.51	39.47	47.11
PSPNet ( Zhao et al., 2017 )	Gen. Sem. Seg.	I	CVPR’17	63.12	69.39	72.83	58.29	72.97	18.93	12.71	20.00	46.47	40.63
Deeplabv3 ( Chen et al., 2017a )	Gen. Sem. Seg.	I	CVPR’18	81.07	77.77	79.03	65.25	79.21	28.29	20.04	21.76	40.85	38.42
SegFormer ( Xie et al., 2021 )	Gen. Sem. Seg.	I	NeurIPS’21	69.07	71.66	79.14	63.75	77.21	22.97	15.70	26.10	41.45	41.10
Mask2Former ( Cheng et al., 2022 )	Gen. Sem. Seg.	I	CVPR’22	64.20	69.69	74.73	58.93	73.82	36.00	26.33	33.92	38.03	41.98

Table 2. Comprehensive benchmark results on GTPBD-MM . Gen. Sem. Seg. , Parcel Delin. , VL Seg. , and MM Parcel Delin. denote General Semantic Segmentation , Parcel Delineation , Reasoning Segmentation , and Multimodal Parcel Delineation , respectively.

I

T

, and

D

denote image, text, and DEM inputs, respectively. Pub. indicates the publication venue/year. Best and second-best completed results are highlighted in bold and underline , respectively.

Figure 4. Overview of our proposed Elevation and Text Guided Terraced Parcel Network (ETTerra), which integrates cross-modal interaction and spatial terrain modulation for multimodal terraced parcel extraction.

Figure 5. Qualitative comparison of different methods on GTPBD-MM. Red boxes highlight typical semantic confusion in non-parcel regions without sufficient textual guidance, while yellow boxes show geometric recovery errors caused by the absence of DEM cues. Our ETTerra produces more complete, coherent, and structurally consistent parcel delineation results.

Figure 6. Edge-level error analysis of different methods on GTPBD-MM. We visualize correct edges, false positive edges, and false negative edges together with OIS and mAcc.

Figure 7. Object-level under-segmentation error analysis of different methods on GTPBD-MM. We visualize GUC-based error distribution together with GUC and mAcc.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8. More representative cases of GTPBD-MM from different regions. From top to bottom, the samples are collected from Chongqing, Guangxi, Zhejiang, and Guangdong in China, followed by Vietnam and Indonesia.

Figure 9. Visual comparison between representative agricultural parcel datasets and GTPBD-MM. Existing datasets mainly focus on regular or relatively flat farmland scenes and provide limited modality or annotation forms, whereas GTPBD-MM presents more complex terraced parcels together with aligned image, text, DEM, and fine-grained annotations.

Category	Setting	Value
Input	Image size	$512\times 512$
Input	DEM size	$512\times 512$
Input	Text max length	77
Backbone	Text encoder	CLIP
Backbone	Segmentation backbone	SAM
Optimizer	Type	AdamW

Appendix

Table 3. Hyperparameter settings of ETTerra on GTPBD-MM.

Figure 10. More boundary visualization results on representative regions. For each case, we compare the predictions of Ours, LISA, and HBGNet with the ground truth. The bottom row in each case shows edge-level error visualization, where red, blue, and green denote false positive edges, false negative edges, and correct edges, respectively.

Figure 11. More object-level error visualization results on representative regions. For each case, we compare the predictions of Ours, LISA, and HBGNet with the ground truth. The error maps are colored according to GUC values, where darker blue indicates larger under-segmentation errors.

Multi-Angular Remote Sensing

See all 8 papers

Delineate Anything v2: A Global Foundation Model for Field Delineation
Jul 21, 2026Mykola Lavreniuk, Nataliia Kussul, Andrii Shelestov +4Remote Sensing Image Segmentation Agricultural Remote Sensing
European Space Agency, Frascati, Italy · Space Research Institute NASU-SSAU, Kyiv, Ukraine · University of Maryland, College Park, Maryland, USA +1
Hedgementation = Hedgerow Segmentation: A Remote Sensing Benchmark
Jun 22, 2026Nathan Senyard, Salem Hamdani, Astrid Zhang +4Remote Sensing Image Segmentation Agricultural Remote Sensing
UBC CS · INSAT Tunisia · Vector Institute +1
TRNet: Topography-Guided Frequency Rectification and Structure-Aware Decoding for Multimodal Paddy Rice Segmentation
Aug 4, 2026Kaiwen Xiao, Chunlong Fu, Liping Zheng +1Agricultural Image Analysis Remote Sensing Image Segmentation
School of Computer Science, Sichuan University Jinjiang College, Meishan 620860, China

Jul 21, 2026J/K move · Enter open · S save

Delineate Anything v2: A Global Foundation Model for Field Delineation

Mykola Lavreniuk, Nataliia Kussul, Andrii Shelestov, Yevhenii Salii+3

European Space Agency, Frascati, Italy · Space Research Institute NASU-SSAU, Kyiv, Ukraine · University of Maryland, College Park, Maryland, USA+1

Remote Sensing Image Segmentation Agricultural Remote Sensing Geospatial Foundation Models Remote Sensing Benchmarks

Open paper AskPDF arXiv

Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities, they frequently fail in geospatial domains due to topological complexity, cropland texturing patterns, and a lack of physical scale awareness. In this work, we introduce Delineate Anything v2, a globally scalable foundation model designed specifically for wide-area field boundary mapping. We construct FBIS-73M, a 73-million-instance multi-resolution dataset spanning 61 countries. To address the pervasive issue of multi-field administrative parcel merging, we introduce a resolution-specific data curation pipeline that leverages topological image-space adaptation to homogenize merged parcels and strengthen weak physical boundaries. Furthermore, we establish a novel, manually curated evaluation benchmark covering 100 countries to assess independent zero-shot generalization. Our results show that Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework, by 0.284 [email protected] (+103.3% relative gain), while maintaining execution speeds suitable for rapid national- and global-scale deployment, as demonstrated by nationwide mapping of Ukraine (603,000 km^2) in 5.4 hours on a consumer-grade workstation. Code, pre-trained weights, the FBIS-73M dataset, and ready-to-use national-scale vector boundary products are publicly available at https://lavreniuk.github.io/Delineate-Anything/.

Abstract

Multi-Angular Remote Sensing

Explore similar work

Delineate Anything v2: A Global Foundation Model for Field Delineation

Hedgementation = Hedgerow Segmentation: A Remote Sensing Benchmark

TRNet: Topography-Guided Frequency Rectification and Structure-Aware Decoding for Multimodal Paddy Rice Segmentation