Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Change Monitoring: A Hyperbolic Representative Learning Framework and a Dataset for Long-term Fine-grained Tree Change Detection
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 UAVTC, a UAV-captured long-term high-resolution tree change dataset, and a Hyperbolic Siamese Network (HSN) for tree change detection that also transfers to cross-domain face anti-spoofing.
Problem
Fine-grained long-term tree change detection is hard because existing datasets use low-resolution imagery and lack continuous fine-grained annotations; tree changes mix hierarchical physiological changes with environmental noise.
Approach
A Siamese network whose change embeddings are mapped via exponential map into the Poincaré ball, with hyperbolic binary logistic regression and feature clipping, to model hierarchical tree changes and suppress extrinsic noise.
Key results
- HSN with ResNet101 backbone reaches 86.83% accuracy and 0.7372 F1 on UAVTC, versus 82.09% / 0.5947 for the ResNet101 Euclidean baseline.
- Hierarchical learning adds about 20% F1 over the Siamese baseline (0.5947 to 0.7164 w/ HL in ablation), and feature clipping raises F1 to 0.7372.
- On cross-domain face anti-spoofing, HSN achieves the best HTER in the OCI→M protocol (11.98%) and the best AUC in most protocols.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HSN instead of the ResNet101 baseline | UAVTC tree change detection · UAVTC test | F1: 0.5947 -> 0.7372 | Helped |
| use HSN instead of SA-FAS | cross-domain face anti-spoofing · Oulu-NPU + CASIA + Idiap → MSU-MFSD | HTER: 13.17 -> 11.98 | Helped |
| use HSN instead of SA-FAS | cross-domain face anti-spoofing · Oulu-NPU + MSU-MFSD + Idiap → CASIA | HTER: 24.48 -> 17.46 | Helped |
| use HSN instead of SA-FAS | cross-domain face anti-spoofing · Oulu-NPU + CASIA + MSU-MFSD → Idiap | HTER: 19.79 -> 18.75 | Helped |
| use HSN instead of SA-FAS | cross-domain face anti-spoofing · Idiap + CASIA + MSU-MFSD → Oulu-NPU | HTER: 14.23 -> 13.45 | Helped |
| add hierarchical learning to the Siamese network (ablation) | Siamese network for UAVTC · UAVTC test | F1: 0.5947 -> 0.7164 | Helped |
| add feature clipping to the Siamese network with hierarchical learning (ablation) | Siamese network for UAVTC · UAVTC test | F1: 0.7164 -> 0.7372 | Helped |
| vary the Poincaré-ball curvature c from 0.1 to 1.0 (ablation) | HSN for UAVTC · UAVTC test | accuracy: 80.71; 80.80; 79.69; 79.74; 80.27 | Mixed |
| vary the Poincaré-ball curvature c from 0.1 to 1.0 (ablation) | HSN for UAVTC · UAVTC test | accuracy: 81.43; 81.86; 80.08; 80.02; 81.00 | Mixed |
| vary the Poincaré-ball embedding dimension (ablation) | HSN for UAVTC · UAVTC test | accuracy | Mixed |
Surprising findings
- Hyperbolic operations add only marginal compute overhead (0.4437 s vs 0.4358 s per iteration).
Limitations
- Only 68 trees from a single 110m x 140m plot in one mixed urban forest, limiting generalizability.
- Hyperbolic operations incur some computational overhead.
- Few comparable fine-grained change detection baselines exist.
Experimental setup
Datasets: UAVTC ImageNet Oulu-NPU CASIA Idiap Replay Attack MSU-MFSD
Benchmarks: UAVTC Oulu-NPU/CASIA/Idiap/MSU-MFSD cross-domain FAS protocols
Metrics: accuracy F1-score HTER AUC
Baselines: ResNet101 VGG16 InceptionV3 MobileNet Tree-blossom recognition SSDG-R SSAN-R PatchNet SA-FAS
Builds on: Hyperbolic Neural Networks Poincaré embeddings Siamese network
Closest papers by meaning
- Bringing SAM to new heights: leveraging elevation data for tree crown segmentation from drone imagery (NeurIPS 2025 · similarity 0.84)
- BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.83)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.83)
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.82)
- BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (NeurIPS 2024 · similarity 0.82)
- BioCLIP: A Vision Foundation Model for the Tree of Life (CVPR 2024 · similarity 0.81)
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.