Research map / Detection and segmentation / Open-set and few-shot detection
FUSU: A Multi-temporal-source Land Use Change Segmentation Dataset for Fine-grained Urban Semantic Understanding
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 FUSU, a multi-temporal, multi-source land use change segmentation dataset (17 classes, ~30B annotated pixels, 847 km²) and a baseline FUSU-Net that jointly does segmentation and change detection using Sentinel time series plus bi-temporal high-resolution images.
Problem
Existing urban change segmentation datasets use coarse land cover classes, single-snapshot or sparse temporal annotations, and lack fine-grained land use semantics needed to study human activity in cities.
Approach
A fine-grained 17-class land use annotation over bi-temporal Google Earth images, expanded with monthly Sentinel-1/2 time series, plus a unified time-series architecture (U-TAE temporal branch fused with an HRNet bi-temporal branch) that performs segmentation and change detection end-to-end.
Key results
- FUSU-Net achieves 50.10 mIoU for intra-dataset semantic segmentation over 17 classes, above HRNet (46.88) and Deeplab-v3 (47.74).
- FUSU-Net reaches 79.80 IoU on binary change detection, outperforming DMINet (72.59) and other baselines.
- Ablation shows semantic change detection mIoU rising from 26.64 with no time-series images to 41.09 with all 25 Sentinel images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Sentinel time-series features to the HRNet-based segmentation model | FUSU-Net · FUSU intra-dataset semantic segmentation | mIoU: 47.74 -> 50.10 | Helped |
| use FUSU-Net instead of ICIFNet | FUSU semantic change detection · FUSU intra-dataset | mIoU: 36.17 -> 41.09 | Helped |
| use FUSU-Net instead of DMINet | FUSU binary change detection · FUSU intra-dataset | IoU: 72.59 -> 79.80 | Helped |
| use FUSU-Net instead of the strongest baseline | Semantic segmentation · Xi'an train, Jiaxing test | mIoU: 10.07 -> 11.65 | Helped |
| use FUSU-Net instead of K-net | Semantic segmentation · Jiaxing train, Xi'an test | mIoU: 10.59 -> 10.46 | Hurt |
| increase the number of time-series images from 0 to 25 (ablation) | FUSU-Net · FUSU semantic segmentation | mIoU: 46.72 -> 50.10 | Helped |
| increase the number of time-series images from 0 to 25 (ablation) | FUSU-Net · FUSU binary change detection | IoU: 65.51 -> 79.80 | Helped |
| increase the number of time-series images from 0 to 25 (ablation) | FUSU-Net · FUSU semantic change detection | mIoU: 26.64 -> 41.09 | Helped |
| add Sentinel-1 data (ablation) | FUSU-Net · FUSU semantic segmentation | mIoU: 51.17 -> 50.10 | Hurt |
Surprising findings
- Cross-city transfer collapses: training on one city and testing on the other drops mIoU to roughly 8-12%, versus 45-50% within-city.
- Removing Sentinel-1 SAR slightly improves segmentation mIoU (51.17 without vs 50.10 with), indicating simple concatenation is not ideal.
Limitations
- Covers only five urban districts in two Chinese cities, limiting geographic generality.
- Land use requires human activity understanding that imagery alone captures poorly, making high accuracy hard.
- Simple concatenation of Sentinel-1 and Sentinel-2 is suboptimal due to sensor gaps.
Experimental setup
Datasets: FUSU ImageNet
Benchmarks: FUSU
Metrics: IoU mIoU
Baselines: FCN PSPNet Fast-SCNN Deeplab-v3 HRNet K-net U-TAE BIT ChangeFormer ICIFNet DMINet SSCD-l
Builds on: U-TAE HRNet ASPP BIT
Closest papers by meaning
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.84)
- DynamicVL: Benchmarking Multimodal Large Language Models for Dynamic City Understanding (NeurIPS 2025 · similarity 0.83)
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.83)
- UV-SAM: Adapting Segment Anything Model for Urban Village Identification (AAAI 2024 · similarity 0.83)
- UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.82)
- UniChange: Unifying Change Detection with Multimodal Large Language Model (CVPR 2026 · 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.