Organizations: Department of Geography, The University of Hong Kong, Hong Kong, China · Institute for Climate and Carbon Neutrality, the University of Hong Kong, Hong Kong, China · The University of Hong Kong Shenzhen Institute of Research and Innovation, Shenzhen, China · School of Information Engineering, Guangdong University of Technology, Guangzhou, China · Institute for Smart Cities, School of Architecture and Urban Planning, Shenzhen University, Shenzhen, China · College of Urban and Environmental Sciences, Academy of Frontier Interdisciplinary Research, Central China Normal University, Wuhan, 430079, China · Laboratory for Remote Sensing and Environmental Change (LRSEC), Department of Earth, Environmental and Geographical Sciences, University of North Carolina, Charlotte, NC 28223-0001, USA · Center of Excellence in Regional, Urban, and Built Environmental Analytics (RUBEA), Faculty of Architecture, Chulalongkorn University, Bangkok, 10330, Thailand
Building age is fundamental for urban infrastructure safety, operation, and planning, as it directly affects structural safety, and sustainable development. Existing methods mostly rely on detailed building information such as street view, function, or materials, which are harder to obtain than geospatial data like satellite imagery and Digital Surface Models (DSMs). However, existing methods based on those geospatial data mainly rely on optical imagery and local texture features, limiting their ability to capture structural information and distinguish buildings from different periods. Moreover, the inherent temporal ordering among building age categories is often overlooked. To address these limitations, we propose the first deep-learning framework for multi-building age estimation using widely available multi-source geospatial data, including remote sensing data and their derived products, as well as building footprint information that estimates age at both pixel and instance levels within a unified pipeline. Specifically, a large pretrained backbone with fine-tuning adapters extracts representative features, and raster-derived and property-derived fusion modules support pixel-wise and instance-wise predictions. Pixel-level estimation employs a pyramid decoder with Cosine-guided Mixture-of-Experts (CMoE) to fuse multi-source raster data across scales, capturing fine-grained building age differences, while instance-level prediction uses attention pooling and Property-Conditional Adaptive Instance Normalization (PCAdaIN) to extract instance-level representations and integrate property attributes. A Temporal Ordinal Constraint Loss (TOCL) enforces the sequential ordering of predicted ages. We also construct a building age benchmark for Hong Kong to enable comprehensive evaluation. The source code and datasets are available at https://github.com/littlebeen/RS-buildingAge.git.
Figures & tables
Figure 1: Comparison between our proposed framework and previous deep-learning-based building age estimation methods. Our framework not only provides multi-building instance-level estimation within a unified image-processing pipeline, but also enables pixel-level estimation. Pixel-level and instance-level supervision mutually reinforce each other, leading to improved building age estimation performance. (a) Single-building age estimation based on street-view data. (b) Multi-building age estimation using detailed building features and GNN. (c) Our multi-building age estimation framework based on widely available multi-source geospatial data .
Dataset
Acquired Place
Image Number
Building Number
Resolution (m)
Input Data Type
MapYourCity [ 5 ]
European cities
30,849
30,849
0.5
Street View, Satellite
NLBA [ 14 ]
Netherlands
5,000
Unknown
0.25
Satellite, Footprint
Hong Kong
Hong Kong
4,194
29,958
0.25
Satellite, DSM, LU, Footprint
Table 1: Comparison of existing open-source building age estimation datasets based on widely available multi-source geospatial data.
Figure 2: The proposed framework consists of two complementary fusion branches: a raster-derived data fusion module for pixel-level estimation and a property-derived data fusion module for instance-level estimation.
Figure 3: The structure of CMoE, which contains three modules: a modality mask gate, a cosine similarity adaptation module, and an MoE module.
Figure 4: The structures of Attention Pooling and PCAdaIN. Attention Pooling extracts instance-level features from the whole-image feature representations, while PCAdaIN fuses building property data with the corresponding raster features.
Data type
Acquisition time
Resolution (m)
True Digital Orthophoto
2025
0.25
DSM
2021
0.5
LU
1990, 2000, 2010, 2020
10
Building footprint
2025
Vector data
Building age
2026
Vector data
Table 2: Data source of the Hong Kong dataset
Figure 5: The distribution of building pixels across different construction periods in the labeled dataset.
Figure 6: The distribution of buildings across different construction periods in the labeled dataset.
Figure 7: (a) Average building height across different construction periods. (b) The spatial-temporal changes of built-up land from 1990 to 2020.
Type
Method
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
Unet [ 27 ]
37.48
35.63
21.83
35.80
32.50
15.5209
11.5319
Unetformer [ 35 ]
51.11
50.07
33.45
43.85
42.33
14.9417
10.4037
ABCNet [ 16 ]
49.31
47.82
31.54
40.96
39.65
15.0633
10.7217
Segformer [ 38 ]
54.73
53.59
36.77
48.27
47.02
13.9652
9.6142
Image only
CMNeXt [ 44 ]
35.68
35.73
21.99
33.92
33.97
15.1417
11.4695
Table 3: Comparison on the Hong Kong dataset of Accuracy, mF1, and mIoU at the pixel level and Accuracy, mF1, RMSE, and MAE at the instance level.
Method
Pixel
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Unet [ 27 ]
52.68
50.42
34.44
Unetformer [ 35 ]
58.95
58.08
41.13
ABCNet [ 16 ]
61.57
60.80
43.99
Segformer [ 38 ]
44.32
34.42
23.36
CMNeXt [ 44 ]
46.61
33.72
23.00
Table 4: Comparison on the Amsterdam dataset of Accuracy, mF1, and mIoU at the pixel level.
Figure 8: Performance comparison across different building age groups on the Hong Kong dataset. (a) Accuracy at the pixel level. (b) F1 at the pixel level. (c) Accuracy at the instance level. (d) F1 at the instance level.
Figure 9: The visual comparisons on the Hong Kong dataset. (a) Satellite image. (b) Ground truth. (c) Unet. (d) Unetformer. (e) ABCNet. (f) Segformer. (g) CMNeXt. (h) TransUnet. (i) CMTFNet. (j) CMT. (k) Deeplab. (l) A2FPN. (m) CMX. (n) Asymformer. (o) MFNet. (p) Ours.
Figure 10: Performance comparison across different building age groups on the Amsterdam dataset. (a) Accuracy at the pixel level. (b) F1 at the pixel level.
Table 5: Parameter Analysis of encoder and decoder dimensions on the Hong Kong dataset
Figure 12: The visual comparisons of cosine similarity features and expert responses in the CMoE module with and without DSM data. (a) Satellite image. (b) Features after the cosine similarity without DSM. (c) Features after the first expert without DSM. (d) Features after the second expert without DSM. (e) DSM data. (f) Features after the cosine similarity with DSM. (g) Features after the first expert with DSM. (h) Features after the second expert with DSM.
Figure 13: Accuracy distribution across different building pixel counts.
Ratio
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
0%
63.45
61.86
44.94
54.94
53.77
10.6374
7.3914
20%
63.67
62.08
45.17
55.28
54.08
10.5328
7.3158
40%
63.80
62.22
45.32
55.54
54.36
10.5312
7.3077
60%
64.21
62.61
45.73
56.59
55.35
10.2738
7.1202
80%
64.45
62.78
45.91
56.93
55.57
10.2249
7.0769
Table 6: Impact of DSM at different proportions on age estimation performance.
Method
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
DSM + Property
43.71
44.87
29.06
42.61
44.20
12.0516
8.9089
Image only
62.47
59.84
43.06
53.91
52.86
10.9541
7.6076
Image + Property
63.30
61.71
44.77
54.63
53.50
10.6200
7.3926
Image + DSM + Property
65.01
63.14
46.29
57.88
56.64
9.7749
6.7446
Image + LU + Property
63.45
61.86
44.94
54.94
53.77
10.6374
7.3914
Table 7: Ablation study of multi-source geospatial data on the Hong Kong dataset.
Set
Number
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
Training set
23,134
82.00
80.92
68.22
67.75
67.17
8.5639
5.8254
Filtered test set
3,361
61.10
60.35
43.28
59.12
58.31
9.4499
6.5159
Full test set
6,824
65.52
63.63
46.80
58.37
57.07
9.6919
6.6978
Table 8: Performance evaluation under different data subsets, where Number denotes the number of buildings.
Method
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
Unetformer [ 35 ]
36.84
33.44
20.34
34.48
31.04
17.3992
12.8084
Segformer [ 38 ]
34.23
32.85
19.78
33.66
31.56
18.1527
13.4193
CMTFNet [ 37 ]
36.16
34.48
21.03
35.48
33.87
17.0559
12.4664
A2FPN [ 15 ]
35.95
33.85
20.54
34.84
32.59
17.6333
12.8944
CMX [ 43 ]
32.69
31.92
19.26
32.61
30.22
18.7380
13.9637
Table 9: Comparison on the Hong Kong dataset under spatial split.
PEM
IEM
TOCL
Instance
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
✓
58.70
57.36
10.4868
7.0139
✓
56.33
55.13
11.2805
7.6128
✓
✓
59.36
58.02
10.2441
6.8792
✓
✓
✓
58.37
57.07
9.6919
6.6978
Table 10: Ablation study of the whole framework. PEM and IEM represent the pixel-level and instance-level estimation modules, respectively.
PCAdaIN
AP
CMoE
Pixel
Instance
Acc (%) ↑
mF1 (%) ↑
mIoU (%) ↑
Acc (%) ↑
mF1 (%) ↑
RMSE ↓
MAE ↓
62.27
60.05
43.21
55.08
53.56
10.2712
7.1624
✓
64.06
62.31
45.39
56.49
55.23
10.0003
6.9430
✓
63.89
62.11
45.23
56.35
54.94
9.9073
6.9147
✓
62.53
61.18
44.16
55.96
55.08
10.1195
7.0564
✓
✓
64.34
62.69
45.78
56.48
55.30
9.9104
6.8473
Table 11: Ablation study of the model structure on the Hong Kong dataset. AP represents Attention Pooling.
Models
Complexity ↓
Memory ↓
Parameters ↓
Speed ↑
(GFLOPs)
(MB)
(M)
(FPS)
Unet
6551.17
6441.47
39.42
7.780
Unetformer
238.03
499.17
11.72
76.613
ABCNet
312.52
507.22
13.39
86.644
Segformer
1135.27
5079.99
27.35
11.922
CMNeXt
1162.14
3625.27
58.67
9.249
Table 12: Computational complexity of the compared methods.
Figure 14: The left figure shows the overall distribution of building ages across Hong Kong by aggregating footprint-level age information into 100 m spatial grids, while the three subfigures on the right illustrate building age distributions in three representative areas at the individual building instance level. To visually represent these aggregated values, the six predefined age categories are evenly distributed along a continuous color bar, where the gradual color transitions between adjacent categories reflect the continuous intermediate category values generated through spatial aggregation rather than an exact continuous age. In contrast, the three subfigures on the right show individual building footprints, where each building is assigned to one of the six discrete age categories and no intermediate values are generated.
Determining the age distribution of the urban building stock is crucial for sustainable municipal heat planning and upgrade prioritization. However, existing approaches often rely on datasets gathered via sensors or remote sensing techniques, leaving inconsistencies and gaps in data. We present a multi-agent LLM system comprising three key agents, the Zensus agent, the OSM agent, and the Monument agent, that fuse data from heterogeneous sources. A data orchestrator and harmonizer geocodes and deduplicates building imprints. Using this fused ground truth, we introduce BuildingAgeCNN, a satellite-only classifier based on a ConvNeXt backbone augmented with a Feature Pyramid Network (FPN), CoordConv spatial channels, and Squeeze-and-Excitation (SE) blocks. Under spatial cross validation, BuildingAgeCNN attains an overall accuracy of 90.69% but a modest macro-F1 of 67.25%, reflecting strong class imbalance and persistent confusions between adjacent historical cohorts. To mitigate risk for planning applications, the address-to prediction pipeline includes calibrated confidence estimates and flags low-confidence cases for manual review. This multi-agent LLM system not only assists in gathering structured data but also helps energy demand planners optimize district-heating networks and target low-carbon sustainable energy systems.
Kundan Thota, Thorsten Schlachter, Veit Hagenmeyer
Institute for Automation and Applied Informatics (IAI), Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
Building extraction from optical remote sensing (RS) imagery is fundamental to urban mapping, yet existing methods are often dataset-specific and generalize poorly to unseen domains. Their practical use is also limited by insufficient detail recovery and weak geometric regularization, leading to blurred boundaries, irregular shapes, and merged adjacent buildings. To address these issues, we propose UniBuild, a unified building extraction framework for multi-source RGB optical RS imagery. First, a unified multi-dataset training scheme is constructed over heterogeneous RGB optical datasets to learn transferable building representations across sensors and resolutions. Second, a novel detail-preserving HR-DPT decoder is designed to integrate high-level semantic features with high-resolution spatial features, enhancing building detail recovery. Third, geometry-aware regularization is introduced through a structure-tensor-based direction-aware loss for boundary direction consistency and a saddle-aware loss for suppressing false activations in narrow inter-building gaps under low-resolution conditions. We train and evaluate UniBuild on multi-source RGB optical datasets, including 10 public high-resolution datasets and two self-collected low-resolution datasets. Experiments show that UniBuild consistently improves building-region accuracy, boundary sharpness, and adjacent-building separation across diverse datasets. It also generalizes well to unseen domains and supports practical building extraction from RGB optical RS imagery up to 10,m resolution. The predicted masks can be further converted into GIS-compatible building footprints through simple polygonization. The trained model and inference code are released at https://github.com/zhu-xlab/UniBuild.
Wei Huang, Chenying Liu, Yilei Shi +1
Chair of Data Science in Earth Observation, Technical University of Munich, 80333 Munich, Germany
Sentinel-2 imagery offers open access, global coverage, and frequent revisit times, making it attractive for practical building mapping at scale; however, its native 10m resolution makes building vs non-building classification challenging, particularly for small or sub-pixel buildings, and performance can vary with both seasonality and the heterogeneity of built-up environments. This paper introduces a Sentinel-2 building-detection framework designed to systematically quantify these effects and to support more formalised, practice-oriented model selection. We construct a dedicated multi-temporal Sentinel-2 dataset over the Warsaw region and derive binary ground-truth masks by rasterising official Polish topographic database (BDOT10k) building footprints onto the Sentinel-2 pixel grid. Using two established convolutional segmentation backbones (U-Net and DeepLabV3+), we first perform scene-specific fine-tuning to select a robust architecture and identify the best monthly models for L1C and L2A products separately. We then conduct cross-temporal inference by applying each best monthly model to all scenes, enabling an assessment of (i) which months provide favourable training and inference conditions, (ii) how performance transfers between seasons, (iii) the impact of processing level, and (iv) how these effects differ across built-up typologies. Based on these results, we provide practical guidance for routine Sentinel-2 building classification under varying acquisition periods and settlement characteristics.
Michał Romaszewski, Kamil Drejer, Katarzyna Kołodziej +7
Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Bałtycka 5, 44-100 Gliwice, Poland · Space Research Centre, Polish Academy of Sciences Bartycka 18a, 00-716 Warszawa, Poland