PaperlayerView exact revision on arXiv

Source-linked AI summary

Topologically Consistent Agricultural Parcel Vectorization with Semantic-Guided Diffusion and Topology-Aware Polygonization

Weiqin Jiao, Xiaolong Zuo, Claudio Persello

arXiv:2609.07520v1cs.CV

TL;DR

Parcel maps require regular, low-redundancy polygons with reusable shared boundaries, yet coverage or object-level scores alone do not ensure a coherent map. The paper combines semantic-guided edge–vertex diffusion with partition-first polygonization from a shared graph. Across two datasets it reports effectiveness on geometric, object-level, and topological criteria, including zero measured intrusion and improved shared edge recall on AI4SmallFarms.

  • Problem

    Practical parcel maps need regular geometry and reusable shared boundaries, but coverage or object-level scores alone do not ensure a coherent map.

    Source: §1 · p. 2, §5 · p. 35

  • Method

    Joint edge–vertex latent diffusion with supervised multi-cue conditioning generates parcel primitives, followed by partition-first polygonization from a shared boundary graph.

    Source: §1 · p. 6

  • Results

    Experiments on two datasets demonstrate effectiveness across geometric, object-level, and topological criteria; on AI4SmallFarms, shared edge recall rises from 38.16% for B2PNet to 43.39%, with zero measured intrusion.

    Source: §4.4 · p. 23, §6 · p. 37

  • Takeaways & Limitations

    The results support treating agricultural parcel vectorization as map-level reconstruction, with reliable primitives and topology-aware reconstruction both important to coherent, low-redundancy maps.

    Source: §6 · p. 37

    In the two-stage design, primitive errors can propagate into reconstruction, including corner omission or incorrect parcel faces.

    Source: §5 · p. 37

Abstract

from arXiv · show

Agricultural parcel polygons play a fundamental role in geospatial applications such as precision agriculture, land administration, and crop monitoring. Beyond regular polygon geometry and low vertex redundancy, practical parcel maps should avoid topological conflicts and preserve common boundaries between adjacent fields. Yet this requirement remains largely unresolved: segmentation-based methods mainly produce parcel masks or raster boundary cues and rely on heuristic raster-to-vector conversion, instance- and contour-based methods reconstruct parcels independently, and recent vector-oriented methods improve polygon regularity but do not explicitly recover adjacent parcels from a shared topological structure. To address this gap, we propose a semantic-guided diffusion framework for topologically consistent agricultural parcel vectorization. It couples joint edge--vertex latent diffusion with supervised multi-cue conditioning to generate geometrically regularised parcel-boundary and vertex primitives while suppressing false-positive responses. A topology-aware parcel polygon reconstruction method then converts these primitives into regular polygons by reconstructing parcel faces from a common planar graph, enabling adjacent predicted parcels to reuse shared boundaries and avoid mutual interior intrusion. Extensive experiments on the AI4SmallFarms and iFLYTEK datasets evaluate parcel vectorization in terms of pixel-level coverage, geometric fidelity, object-level correctness, and topological consistency. The results show strong and competitive performance, with zero measured intrusion ratio and the highest shared-edge recall, demonstrating the potential of the proposed framework for accurate, regular, and topologically consistent agricultural parcel vectorization.

1. Introduction

Agricultural parcel maps must combine regular, low-redundancy polygons with valid shared boundaries, but existing raster and independently reconstructed vector methods do not ensure this topology. The paper couples semantic-guided edge–vertex diffusion with partition-first polygon reconstruction from a shared graph, reporting strong performance particularly in geometric fidelity and topological consistency.

1. Introduction

  • Parcel maps need accurate, regular, low-redundancy polygons that avoid overlap and represent each common boundary once for neighboring fields.These maps support field-level information, crop monitoring, yield estimation, land administration, and agricultural planning.Source: §1 · p. 2
  • Raster methods rely on heuristic vector conversion, while contour- and vector-oriented methods generally reconstruct neighboring parcels without a common topological structure.Such approaches may produce shifted or duplicated boundaries, overlaps, or inconsistent simplification.Source: §1 · p. 4
  • Partition-first reconstruction creates a full-coverage region partition, then uses a shared boundary graph and vertex-guided simplification to recover compact parcel polygons.Adjacent parcels can reuse graph edges, reducing redundant vertices while preserving structural anchors.Source: §1 · pp. 5–6
  • The framework is evaluated on AI4SmallFarms and iFLYTEK across coverage, geometric fidelity, object-level correctness, and topological consistency.The topological evaluation distinguishes intrusion from reference-based shared-boundary recovery using intrusion ratio and shared edge recall.Source: §1 · pp. 5–6
  • Experiments show strong performance across evaluated metrics, with particularly clear gains in geometric fidelity, intrusion avoidance, and shared-edge recovery.The evaluation covers pixel-level coverage, geometric fidelity, object-level correctness, and topological consistency.Source: §1 · pp. 5–6
  • The framework couples joint edge–vertex latent diffusion with supervised multi-cue conditioning to generate regular primitives while mitigating false-positive responses.The conditioning integrates image context, boundary-sensitive cues, and parcel-region cues.Source: §1 · pp. 5–6

2. Related Work

Prior work progresses from raster delineation to instance and vector representations, but these approaches do not ensure shared-boundary reuse among neighboring parcels. The paper draws on topology-aware reconstruction and couples it with semantic-guided primitive generation for adjacent same-class parcels.

2.1. Raster-based agricultural parcel delineation

  • Raster-based methods improve boundary-sensitive parcel delineation, but polygons still require heuristic conversion and independently simplified contours may duplicate or shift common boundaries.Raster outputs can merge adjacent same-class parcels; contour simplification also trades geometric fidelity against vertex efficiency.Source: §2.1 · pp. 7–8

2.2. Parcel extraction beyond semantic segmentation

  • Instance and contour methods produce parcel instances or vertex sequences, while EEDNet and B2PNet target compact polygons; none reconstructs neighbors from a common boundary graph.Independent instances can retain overlaps or gaps, while errors in repaired edges or learned connectivity can propagate to inconsistent boundaries.Source: §2.2 · pp. 8–9

2.3. Polygonal object vectorization and topology-aware reconstruction

  • Polygonal object and building methods generally use instance-wise outlines or segmentation-based polygonization, which does not ensure shared-boundary reuse for adjacent agricultural parcels.ACPV-Net demonstrates partition-to-graph reconstruction for multi-class regions, but agricultural parcels need parcel-wise partitioning within a single semantic class.Source: §2.3 · pp. 9–10
  • The proposed design combines semantic-guided joint-latent diffusion, a full-coverage region partition, and shared-graph polygon reconstruction with vertex-guided simplification.Reliable boundary primitives help separate same-class parcel regions before topology-aware reconstruction.Source: §2.3 · pp. 10–11

3. Methodology

The method couples semantic-guided generation of edge, vertex, and parcel-mask primitives with partition-first polygon reconstruction on a shared boundary graph. For large tiles, it fuses patch-level evidence before reconstructing polygons once at tile level.

3.1. Overview

  • The framework predicts edge, vertex, and parcel-mask primitives, then reconstructs polygons from a shared-boundary graph rather than simplifying parcel outlines independently.This lets adjacent parcels reuse recovered graph edges.Source: §3.1 · p. 11

3.2. Primitive generation network

  • Joint latent diffusion generates edge masks and vertex heatmaps, modeling their spatial dependency so vertex responses remain consistent with recovered boundaries.The parcel mask is predicted discriminatively and retained to filter non-parcel faces.Source: §3.2 · p. 12, §3.2.3 · p. 15

3.2.1. Shared semantic feature extraction

  • VMamba and UPerNet produce a shared multi-scale semantic feature map that supplies context to task-specific branches and the diffusion condition.The backbone is fine-tuned jointly with the rest of the network.Source: §3.2.1 · pp. 12–13

3.2.2. Supervised multi-cue conditioning

  • Supervised edge and parcel branches form a multi-cue condition from edge-aware, parcel-aware, and shared features to guide diffusion toward parcel-valid structures.Parcel-mask supervision supplies region semantics that help suppress false edge and vertex responses in non-parcel regions.Source: §3.2.2 · p. 14

3.2.3. Joint edge-vertex latent diffusion

  • Edge and vertex targets are encoded as aligned latents and concatenated into a six-channel denoising state, allowing the denoiser to model boundary continuity and vertex localization jointly.The denoising objective is combined with edge and parcel-mask segmentation losses.Source: §3.2.3 · pp. 14–15

3.3. Topology-aware parcel polygonization

  • The polygonization pipeline partitions the image domain from the reconstructed edge mask, builds one graph from region transitions, and simplifies shared polylines using structural anchors and projected vertex peaks.It traces closed faces and retains those with sufficient support from the predicted parcel mask.Source: §3.3 · pp. 16–17, Figure 4 · p. 17

3.4. Tile-level vectorization from patch predictions

  • For large tiles, overlapping patch evidence is fused in tile coordinates before one tile-level partition and shared-graph polygonization, rather than merging patch-level polygons.Edge evidence is overlap-averaged; parcel masks and vertex heatmaps use non-overlapping valid crop regions.Source: §3.4 · pp. 17–18

4.1. Datasets

The evaluation uses AI4SmallFarms and iFLYTEK as complementary agricultural parcel datasets, with large test tiles processed through overlapping patches.

4.1. Datasets

  • AI4SmallFarms represents smallholder landscapes in Vietnam and Cambodia, while iFLYTEK represents cultivated-land mapping with non-uniform parcels and ambiguous boundaries.The authors describe the pair as supporting evaluation across different agricultural mapping conditions.Source: §4.1 · p. 18
  • AI4SmallFarms has 62 tiles and 439,001 polygons; experiments use aligned high-resolution Google Map RGB imagery and the original 43/9/10 train-validation-test tile split.Training and validation tiles yield 18,473 and 3,934 non-overlapping 512 × 512 patches, respectively.Source: §4.1 · p. 18
  • iFLYTEK experiments use RGB from 31 JiLin-1 tiles, with 16 training tiles and 15 test tiles; training tiles yield 2,486 patches and validation yields 629.Following the stated protocol, large test tiles are processed with overlapping 512 × 512 patches at 256-pixel stride.Source: §4.1 · pp. 18–19

4.2. Implementation details

Implementation uses dataset-specific training schedules and compares raster, contour/vectorization, and building-polygon baselines under stated tile-level processing procedures.

4.2. Implementation details

  • Training uses AdamW with a 6 × 10−5 base learning rate, batch size 8, and EMA decay 0.9999, for 240k AI4SmallFarms and 160k iFLYTEK iterations.Experiments use one NVIDIA L40S GPU.Source: §4.2 · p. 19
  • Baselines span raster parcel delineation, instance-wise contour extraction and vectorization, and building-oriented polygon extraction methods.The included methods are grouped by their output type and intended cross-domain generalization assessment.Source: §4.2 · p. 19
  • Baseline results are taken from B2PNet when available; otherwise, metrics are computed from its authors' predictions using specified tile-level conversion or stitching procedures.Raster outputs are stitched and converted by contour extraction plus Douglas–Peucker simplification; other polygon predictions use the stated fusion strategy, except B2PNet's image-level post-processing.Source: §4.2 · p. 19

4.3. Evaluation metrics

The evaluation measures parcel vectorization across pixel-level coverage, geometric fidelity, object-level correctness, and topological consistency. Complementary metrics assess mask and boundary overlap, polygon shape and complexity, matched-object detection, and neighboring-parcel relationships.

4.3. Evaluation metrics

  • Pixel-level coverage rasterizes polygon outputs and uses IoU for general mask overlap and Boundary IoU (BIoU) for boundary alignment.BIoU computes IoU between dilated boundary regions.Source: §4.3 · p. 20
  • Geometric fidelity uses CIoU, N-ratio, PoLiS, and Maximum Tangent Angle error (MTA) to assess shape overlap, vertex redundancy, boundary distance, and orientation consistency.CIoU penalizes vertex-count mismatch; lower PoLiS and MTA indicate better boundary fidelity and orientation consistency, respectively.Source: §4.3 · pp. 20–21
  • Object-level correctness uses one-to-one polygon matching at IoU thresholds of 0.50, 0.75, and 0.90 to assess precision, recall, and F1-score.Each predicted polygon can match at most one ground-truth polygon, and vice versa.Source: §4.3 · p. 21
  • Topological consistency uses Intrusion Ratio (IR) for predicted interior overlap and Shared Edge Recall (SER) for recovery of reference boundaries between adjacent parcels.Zero IR alone does not guarantee correct shared-boundary recovery, because parcels may have gaps or be merged.Source: §4.3 · pp. 21–22

4.4. Results on AI4SmallFarms

On AI4SmallFarms, the proposed method leads overall, improving geometric fidelity and shared-edge recovery over B2PNet while recording zero measured intrusion. It also achieves the best average recall and F1, though baseline families show different topology and geometry trade-offs.

4.4. Results on AI4SmallFarms

  • 90.54% IoU, 86.65% BIoU, and 78.73% CIoU surpass B2PNet’s respective 88.32%, 81.95%, and 61.57%, while SER rises from 38.16% to 43.39%.The proposed method also records zero measured intrusion; the authors connect this result to reconstructing parcels from a shared boundary graph.Source: Table 1 · p. 22, §4.4 · p. 23
  • 59.59% average recall and 55.79% average F1 exceed B2PNet’s 57.43% and 54.39%, while E2EC has the highest average precision.The recall improvement is clearest at IoU thresholds 0.50 and 0.75; E2EC’s precision advantage accompanies lower recall and poorer geometric complexity.Source: §4.4 · p. 23, Table 2 · p. 23
  • Raster-based methods record zero measured intrusion but often merge adjacent parcels and recover true shared boundaries poorly.Their low shared-edge recall means zero intrusion alone does not establish correct shared-boundary topology.Source: §4.4 · pp. 23–24
  • Instance-wise contour methods recover shared edges better than raster methods but generally have higher intrusion; E2EC and E2EVAP exceed 1.0% intrusion and have poor CIoU and N-ratio.Their independently predicted contours can overlap or have non-coincident neighboring boundaries.Source: §4.4 · p. 24
  • PolyR-CNN and RoIPoly perform poorly across most metrics in this dense agricultural setting, where parcels are irregular and frequently share long boundaries.The paper attributes this to their RoI-based prediction being designed for relatively isolated building footprints.Source: §4.4 · p. 24
  • In the representative 58 cambodia tile, the proposed method shows regular polygons, fewer redundant vertices, and no visible mutual intrusion in the highlighted regions.Figure 5 compares aerial imagery and ground truth with SEANet, B2PNet, and the proposed method; red regions mark mutual intrusion.Source: §4.4 · pp. 24–25, Figure 5 · p. 25

4.5. Results on iFLYTEK

On iFLYTEK, the method maintains strong geometric fidelity and topological consistency, with zero measured intrusion and the highest shared-edge recall. It balances competitive coverage with the best precision and F1, while baseline outcomes differ across metrics.

4.5. Results on iFLYTEK

  • SEANet has the highest IoU, while the proposed method has the best BIoU and second-best IoU, indicating stronger boundary alignment with competitive coverage.Source: Table 3 · p. 26
  • 61.89% CIoU exceeds B2PNet’s 47.81%, while N-ratio changes from 1.58 to 0.91.The paper describes this as a better balance between shape fidelity and vertex redundancy.Source: §4.5 · p. 26
  • 21.96% shared-edge recall is the highest, up from B2PNet’s 19.58%, and the proposed method records zero measured intrusion.The results also report the best CIoU, N-ratio, and MTA for geometric fidelity.Source: Table 3 · p. 26, §4.5 · p. 27
  • At IoU 0.90, F1 improves from B2PNet’s 16.63% to 23.22%; the proposed method also has the highest average precision and F1.B2PNet has the highest average recall.Source: Table 4 · p. 26, §4.5 · p. 27
  • Across baselines, raster methods can have low intrusion through merging but extremely low shared-edge recall, while instance-wise methods recover more shared boundaries but retain intrusion or poor geometric complexity.Building-oriented methods remain unstable in this cultivated-land setting.Source: §4.5 · p. 27
  • On the CGJC 22 test tile, the proposed method maintains competitive coverage while producing more regular, vertex-efficient, and topologically consistent polygons over the whole tile.The compared methods show jagged or fragmented boundaries, missed regions, local intrusions, dense vertices, or patch-border artifacts.Source: §4.5 · pp. 27–28

4.6. Ablation study

The ablations isolate primitive generation, topology-aware polygonization, and tile-level reconstruction, showing how these components affect geometric, object-level, and topological performance.

4.6.1. Primitive generation

  • Semantic supervision produces the largest primitive-generation gains: CIoU rises from 71.67% to 76.20%, and vertex N-ratio moves from 1.42 to 1.21.Average precision and F1 rise from 46.33% and 47.58% to 55.03% and 54.62%, respectively.Source: §4.6.1 · p. 30
  • Joint edge–vertex diffusion improves geometric quality over vertex-only diffusion, moving N-ratio from 1.33 to 1.21 and MTA from 41.84 to 37.01.The joint-diffusion variant also achieves higher BIoU and CIoU and a more reference-like N-ratio.Source: §4.6.1 · p. 30
  • Multi-cue conditioning raises BIoU from 63.55% to 64.08% and CIoU from 76.20% to 77.59%, while improving N-ratio from 1.21 to 1.11.The fused shared, edge-aware, and parcel-aware cues also yield the highest average precision.Source: §4.6.1 · p. 30
  • The full primitive generator achieves the best BIoU and CIoU, most reference-like vertex efficiency, lowest orientation error, and highest object precision.Its edge and vertex primitives are visually cleaner than ablated variants, which show breaks, blurred responses, and redundant or false-positive vertices.Source: §4.6.1 · p. 31, Figure 7 · p. 31

4.6.2. Topology-aware polygonization

  • Douglas–Peucker simplification is tolerance-sensitive: ϵ = 0.5 and 1.0 leave N-ratios of 9.51 and 5.52, while ϵ = 2.0 improves vertex efficiency and CIoU.Even at ϵ = 2.0, mask-contour irregularities can retain redundant vertices; the full method offers a stronger geometric balance.Source: §4.6.2 · pp. 32–33, Figure 9 · p. 34
  • The full topology-aware polygonization has the strongest geometric balance, including best CIoU, lowest PoLiS, and zero measured intrusion among reconstruction alternatives.Independent contour simplification reaches 37.64% BIoU and 29.95% SER; skeleton-graph polygonization reaches 67.91% CIoU and 31.81% SER, versus the full method’s 77.59% CIoU and 38.61% SER.Source: §4.6.2 · pp. 32–33, Table 6 · p. 32

4.6.3. Tile-level vectorization

  • 86.65% BIoU, 59.59% average recall, 55.79% average F1, and 43.39% SER accompany 0.00% IR for tile-level reconstruction, versus B2PNet-style stitching.B2PNet-style stitching has 82.70% BIoU, 49.15% average recall, 50.29% average F1, 27.01% SER, and 1.46% IR.Source: Table 7 · p. 34, §4.6.3 · pp. 34–35
  • 86.65% BIoU, 59.59% average recall, 55.79% average F1, and 43.39% SER show better tile-level map quality than post hoc polygon stitching.The proposed strategy stitches intermediate primitives in tile coordinates and polygonizes once across the full tile, while B2PNet-style stitching merges or prunes patch-level polygons.Source: §4.6.3 · pp. 34–35

5. Discussion

The discussion frames parcel vectorization as map-level reconstruction: pixel and object scores alone do not establish topological coherence. The shared-graph pipeline supports boundary reuse, but its two-stage design remains vulnerable to primitive errors.

5. Discussion

  • Similar IoU or object-level scores can mask duplicated boundaries, overlaps, missing shared edges, or inconsistent adjacent polygons, motivating complementary geometric, object, and topology metrics.Intrusion ratio and shared edge recall explicitly measure whether neighboring parcels form a consistent vector map.Source: §5 · p. 35
  • Partition-first reconstruction organizes parcel faces in a common boundary graph, so neighbors reuse boundary chains and avoid mutual interior intrusion.The tile-level approach also aggregates intermediate primitives before polygonization rather than stitching independently generated patch polygons.Source: §5 · pp. 36–37
  • The two-stage design can propagate primitive errors: overlapping corner responses may omit corners, while weak, missing, or spurious boundaries may yield merged parcels, over-segmentation, or incorrect faces.The deterministic reconstruction does not explicitly model uncertainty in predicted primitives.Source: §5 · p. 37

6. Conclusion

The paper presents a diffusion-based primitive generator with shared-graph polygon reconstruction for agricultural parcel vectorization. Across two datasets, it reports effectiveness across geometric, object-level, and topological criteria.

6. Conclusion

  • The framework combines diffusion-generated geometric primitives with shared-graph polygon reconstruction for agricultural parcel vectorization.Source: §6 · p. 37
  • Reconstructing parcels from a common topological structure improves shared-boundary recovery and produces compact vector representations with low vertex redundancy.Source: §6 · p. 37
  • Experiments on AI4SmallFarms and iFLYTEK demonstrate effectiveness across geometric, object-level, and topological evaluation criteria.Source: §6 · p. 37

Related papers

Loading 2609.07520v1…