Research map / Vision-language and multimodal / Human and embodied modeling
Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era
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.
Introduces Archaeoscape, an open-access, expert-annotated airborne laser scanning (ALS) archaeology dataset covering 888 km² in Cambodia, and benchmarks modern segmentation models on it.
Problem
Detecting subtle Angkorian archaeological features under dense jungle canopy in ALS data is hard, and deep learning progress is blocked by a lack of expert-annotated, open-access ALS archaeology datasets.
Approach
A large, expert-annotated ALS archaeology dataset (orthophotos plus LiDAR-derived nDTM with 5-class polygon annotations) released under credentialized open access with misuse safeguards, benchmarked across CNN, ViT and hierarchical ViT segmentation models with elevation channel added to RGB.
Key results
- Best model PVTv2-b1 (224 input, ImageNet1K pretraining) reaches 52.1 avg IoU and 88.7 OA; U-Net at 512 input reaches 52.8 avg IoU and 89.6 OA.
- Models trained on RGB only perform below 30% mIoU, versus about 50% with elevation data, showing elevation is the most informative channel under canopy.
- Foundation models (DINOv2, CLIP, ScaleMAE, DOFA) perform poorly, with ViT-L ScaleMAE at 30.4 avg IoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase U-Net input size from 224 to 512 | U-Net, ImageNet1K · Archaeoscape test | mIoU: 50.5 -> 52.8 | Helped |
| Increase PVTv2 input size from 224 to 512 | PVTv2, ImageNet1K · Archaeoscape test | mIoU: 52.1 -> 52.2 | Helped |
| Use RGB plus elevation rather than RGB only (ablation) | U-Net, ImageNet1K · Archaeoscape test | mIoU: 34.2 -> 50.5 | Helped |
| Use RGB plus elevation rather than RGB only (ablation) | ViT-S, DINOv2 · Archaeoscape test | mIoU: 29.0 -> 41.9 | Helped |
| Use RGB plus elevation rather than RGB only (ablation) | PVTv2-b1, ImageNet1K · Archaeoscape test | mIoU: 33.9 -> 52.1 | Helped |
| Remove RGB channels and use elevation only (ablation) | U-Net, ImageNet1K · Archaeoscape test | temple IoU: 33.3 -> 28.8 | Hurt |
| Remove RGB channels and use elevation only (ablation) | PVTv2-b1, ImageNet1K · Archaeoscape test | temple IoU: 32.3 -> 27.8 | Hurt |
| Remove RGB channels and use elevation only (ablation) | U-Net, ImageNet1K · Archaeoscape test | hydro IoU: 32.7 -> 37.8 | Helped |
| Replace proposed initialization with fully random initialization (ablation) | PVTv2-b1, ImageNet1K · Archaeoscape test | mIoU: 52.1 -> 46.0 | Hurt |
| Randomly initialize the first layer while retaining other pretrained weights (ablation) | PVTv2-b1, ImageNet1K · Archaeoscape test | mIoU: 52.1 -> 44.4 | Hurt |
| Use LoRA rank 32 instead of proposed initialization and fine-tuning (ablation) | PVTv2-b1, ImageNet1K · Archaeoscape test | mIoU: 52.1 -> 46.1 | Hurt |
| Fine-tune DINOv2 with LoRA instead of full fine-tuning (ablation) | DINOv2 small · Archaeoscape test | mIoU | Hurt |
Surprising findings
- CNN U-Net outperforms most ViT and foundation-model variants on this task.
- Foundation models pretrained on natural or satellite imagery adapt poorly to the elevation channel.
Limitations
- Domain shift between two acquisition campaigns with different sensors and weather conditions.
- Annotation ambiguity for faint, gradually sloping features and possible unannotated features in background.
- Focuses only on Khmer civilization in Cambodia, so conclusions may not generalize to other cultures.
Experimental setup
Datasets: Archaeoscape
Benchmarks: Archaeoscape
Metrics: IoU mean IoU overall accuracy
Baselines: U-Net DeepLabv3 ViT-S ViT-B ViT-L HybViT DOFA SWIN-S SWIN-B PCPVT-S PVTv2-b1 ScaleMAE
Builds on: U-Net DeepLabv3 Swin Transformer PVTv2 DINOv2 LoRA
Closest papers by meaning
- LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images (CVPR 2024 · similarity 0.84)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.83)
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.82)
- VaseVQA-3D: Benchmarking 3D VLMs on Ancient Greek Pottery (ICLR 2026 · similarity 0.81)
- Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration (CVPR 2025 · similarity 0.81)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.80)
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.