Research map / Vision-language and multimodal / Human and embodied modeling
DCV2I: A Practical Approach for Supporting Geographers’ Visual Interpretation in Dune Segmentation with Deep Vision Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents DCV2I, an ArcGIS-integrated tool wrapping an Attention U-Net with ASPP and VGG16 encoder for automatic dune segmentation in Mu Us Sandy Land imagery, evaluated on a new dune dataset and deployed with geographers.
Problem
Manual visual interpretation of dune imagery by geographers is time-consuming and labor-intensive, and existing GOBIA tools and general segmentation models (e.g., SAM) perform poorly on dune segmentation.
Approach
Embed an ASPP module into the Attention U-Net (VGG16 encoder) with focal loss for dune segmentation, and wrap it in an ArcGIS plug-in so geographers get editable automatic segmentations without interactive prompting or leaving their GIS workflow.
Key results
- Proposed model reaches 96.72% accuracy, 72.96% F1, 80.67% mPA and 75.09% MIoU, versus 92.97% accuracy and 72.91% MIoU for U-Net on the dune test set.
- Two professional geographers used DCV2I after a 5-minute tutorial, with over 95% of segments used as-is, and task time dropped from a couple of days to under an hour.
- Ablation shows adding both ASPP and attention to U-Net gives the best MIoU (75.09 vs 72.91 baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace U-Net with DCV2I | U-Net · Evaluation Area | accuracy: 92.97 -> 96.72 | Helped |
| Replace U-Net with DCV2I | U-Net · Evaluation Area | F1 Score: 67.86 -> 72.96 | Helped |
| Replace U-Net with DCV2I | U-Net · Evaluation Area | mPA: 77.94 -> 80.67 | Helped |
| Replace U-Net with DCV2I | U-Net · Evaluation Area | MIoU: 72.91 -> 75.09 | Helped |
| Add Attention mechanism only (ablation) | U-Net · Evaluation Area | accuracy: 92.97 -> 95.41 | Helped |
| Add Attention mechanism only (ablation) | U-Net · Evaluation Area | F1 Score: 67.86 -> 70.99 | Helped |
| Add Attention mechanism only (ablation) | U-Net · Evaluation Area | mPA: 77.94 -> 80.21 | Helped |
| Add Attention mechanism only (ablation) | U-Net · Evaluation Area | MIoU: 72.91 -> 74.62 | Helped |
| Add ASPP module only (ablation) | U-Net · Evaluation Area | accuracy: 92.97 -> 94.78 | Helped |
| Add ASPP module only (ablation) | U-Net · Evaluation Area | F1 Score: 67.86 -> 69.92 | Helped |
| Add ASPP module only (ablation) | U-Net · Evaluation Area | mPA: 77.94 -> 79.65 | Helped |
| Add ASPP module only (ablation) | U-Net · Evaluation Area | MIoU: 72.91 -> 74.89 | Helped |
Surprising findings
- Geographers could not get satisfying results with SAM after several dozen interactions, suggesting interactive generalist models are premature for this domain.
Limitations
- Inference waiting time of 10-15 minutes per image on personal computers is still long.
- Evaluated on a single study area (Mu Us Sandy Land) with a small dataset of 1,537 images.
- Did not include native GOBIA tools in quantitative comparison because they gave little meaningful output.
Experimental setup
Datasets: VOC (pre-training for VGG16) Mu Us Sandy Land Pleiades satellite dune dataset
Benchmarks: Mu Us Sandy Land dune dataset
Metrics: accuracy F1 score mPA MIoU
Baselines: FCN U-Net SegNet DeepLabV3+
Builds on: U-Net Attention U-Net DeepLabV3+ Focal Loss
Closest papers by meaning
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.83)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.83)
- LithoBench: Benchmarking Large Multimodal Models for Remote-Sensing Lithology Interpretation (NeurIPS 2026 · similarity 0.83)
- UV-SAM: Adapting Segment Anything Model for Urban Village Identification (AAAI 2024 · similarity 0.83)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.82)
- DynamicVL: Benchmarking Multimodal Large Language Models for Dynamic City Understanding (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.