[Paper Review] Hi-UCD: A Large-scale Dataset for Urban Semantic Change Detection in Remote Sensing Imagery

Shiqi Tian, Ailong Ma+2 more|arXiv (Cornell University)|Nov 6, 2020

Remote-Sensing Image Classification22 references40 citations

TL;DR

Hi-UCD introduces a large-scale ultra-high-resolution aerial dataset for urban semantic change detection with three time phases and 9 land-cover classes, plus a benchmark of traditional and deep-learning methods. It enables refined, multi-temporal UCD research.

ABSTRACT

With the acceleration of the urban expansion, urban change detection (UCD), as a significant and effective approach, can provide the change information with respect to geospatial objects for dynamical urban analysis. However, existing datasets suffer from three bottlenecks: (1) lack of high spatial resolution images; (2) lack of semantic annotation; (3) lack of long-range multi-temporal images. In this paper, we propose a large scale benchmark dataset, termed Hi-UCD. This dataset uses aerial images with a spatial resolution of 0.1 m provided by the Estonia Land Board, including three-time phases, and semantically annotated with nine classes of land cover to obtain the direction of ground objects change. It can be used for detecting and analyzing refined urban changes. We benchmark our dataset using some classic methods in binary and multi-class change detection. Experimental results show that Hi-UCD is challenging yet useful. We hope the Hi-UCD can become a strong benchmark accelerating future research.

Motivation & Objective

• Address the lack of high-spatial-resolution data, semantic annotations, and long-range multi-temporal imagery for urban change detection (UCD).
• Create a large-scale, semantically annotated UCD dataset with ultra-high resolution and multi-temporal coverage.
• Provide a unified benchmark using traditional and deep-learning methods for binary and multi-class change detection on Hi-UCD.
• Enable analysis of refined urban changes through semantic labels and three-year temporal pairs.

Proposed method

• Assemble ultra-high-resolution aerial imagery at 0.1 m GSD of Tallinn, Estonia.
• Annotate 9 land-cover classes mapped to the Estonian Topographic Database and generate semantic change maps across years (2017-2018, 2018-2019, 2017-2019).
• Patch extraction into 1024x1024 tiles and filter patches with >200 change pixels to form the dataset.
• Provide three temporal pairs per area (two-year and three-year changes) and orthorectified images with seasonal consistency.
• Benchmark a suite of binary and multi-class change detection methods, including traditional (CVA, MAD, IRMAD) and deep-learning approaches (FC_Siam_diff, FC_Siam_conc, FC_EF, FC_EF_Res, Siam_Res_EF; Deeplab v3/v3+, PSPNet, FarSeg).
• Adopt standard training/evaluation settings (SGD, learning rate 0.01, polynomial decay, batch size 4, 10k or 20k iterations, data augmentation).

Experimental results

Research questions

• RQ1 Can ultra-high-resolution (0.1 m) aerial imagery with semantic annotations detect finer-grained urban changes than existing datasets?
• RQ2 How do traditional change-detection methods compare to deep-learning approaches on Hi-UCD for binary and multi-class tasks?
• RQ3 What are the performance and complexity trade-offs of state-of-the-art semantic segmentation models when applied to Hi-UCD’s multi-temporal data?

Key findings

• Hi-UCD provides 359 image pairs (2017-2018), 386 pairs (2018-2019), and 548 pairs (2017-2019) with semantic and change maps.
• Deep learning methods significantly outperform traditional change-detection methods on Hi-UCD across evaluated metrics.
• Among binary-change methods, Siamese and EF-based networks achieve the strongest performance, with higher OA and IoU than classic CVA/MAD/IRMAD baselines.
• For multi-class change detection, semantic segmentation models (e.g., Deeplab v3 family, PSPNet, FarSeg) provide competitive results, albeit with higher computational cost and sensitivity to classification accuracy.
• Hi-UCD’s ultra-high resolution (0.1 m) reveals fine spatial details and enables more precise boundary delineation of changes, while multi-temporal data (three phases) supports temporal analysis of urban dynamics.