Estimation of Construction Year from Building Images: An Evaluation of AI Against Human Expertise
Additional geographical or spatial references | Lower Saxony | |
Countries to which the data refer | GERMANY | |
Type of the data | Dataset | |
Type of the data | Image | |
Type of the data | Text | |
Total size of the dataset | 1613685059 | |
Author | Soot, Matthias | |
Author | Kretzschmar, Daniel | |
Author | Eberwein, Johannes | |
Author | Zaddach, Sebastian | |
Author | Teuber, Andreas | |
Author | Weitkamp, Alexandra | |
Upload date | 2026-11-19T12:31:25Z | |
Publication date | 2026-11-19T12:31:25Z | |
Data of data creation | 2024-11 | |
Publication date | 2026-11-19 | |
Abstract of the dataset | The construction, operation, and demolition of buildings account for a significant proportion of global energy demand and CO2 emissions. The preservation of existing buildings and the promotion of refurbishment are therefore essential for sustainability. The year of construction is a primary indicator when targeting specific sustainability measures, but detailed information about the construction year of a particular building is often unavailable. This study explores the possibility of using large pre-trained visual language models (VLMs) to estimate the year of construction based on images. Their accuracy is compared with that of real estate experts. A comprehensive survey of experts is conducted, and ChatGPT prompts are applied to a dataset of building images. The results show that ChatGPT makes more accurate statements about the year of construction of buildings than individual experts, although the collective intelligence of many experts provides better results than ChatGPT in terms of the mean absolute error. This suggests that VLMs should be used more extensively than at present to determine the construction year of buildings, and especially to verify expert assessments. Using this knowledge, urban planning can be tailored more specifically to extend the life cycles of buildings and close loops in line with the principles of circularity. | |
Public reference to this page | https://opara.zih.tu-dresden.de/handle/123456789/3038 | |
Public reference to this page | https://doi.org/10.25532/OPARA-1619 | |
Publisher | Technische Universität Dresden | |
Licence | Attribution-NonCommercial-NoDerivatives 4.0 International | en |
URI of the licence text | http://creativecommons.org/licenses/by-nc-nd/4.0/ | |
Specification of the discipline(s) | 4::45::410::410-01 | |
Title of the dataset | Estimation of Construction Year from Building Images: An Evaluation of AI Against Human Expertise | |
Research instruments | Survey | |
Underlying research object | Buildings | |
Funding Acknowledgement | This work was supported by the federal state of Lower Saxony as well as from Deutsche Forschungsgemeinschaft (DFG) under grant number 530964147. |
Files
Original bundle
- Name:
- ergebnisChatGPT2.xlsx
- Size:
- 22.39 KB
- Format:
- Microsoft Excel XML
- Description:
- File contains a summary of all runs by ChatGPT-4.o, ChatGPT-5.6 and Mistral with the estimation of the year of construction (Columns B) and uncertainty estimates (Column U) for all 100 images. The ChatGPT GUI runs are excluded in the publication as the temperature was not contolled.
- Name:
- reference.csv
- Size:
- 1.18 KB
- Format:
- Comma-Separated Values
- Description:
- File contains the reference year of construction for all 100 images
- Name:
- results-survey915272-anonym.csv
- Size:
- 220.01 KB
- Format:
- Comma-Separated Values
- Description:
- File contains the results of the survey for all participants. Columns Bild<n>[1] contains the estimate of the year of construction for image <n>, Bild<n>[2] contains the uncertainty estimate for image <n>. Complete text of questions are in the file results-survey915272-fragentext-anonym.csv
- Name:
- results-survey915272-fragentext-anonym.csv
- Size:
- 256.64 KB
- Format:
- Comma-Separated Values
- Description:
- File contains the results of the survey for all participants. Contains the complete text of every question.
- Name:
- GUI-Results.zip
- Size:
- 384.17 MB
- Format:
- Description:
- Restults from the GUI of ChatGPT 4.o (not included in the paper as the temperature was not controlled) and Mistral.
- Name:
- API-Results.zip
- Size:
- 1.13 GB
- Format:
- Description:
- Results from all API runs using OpenAIs API for ChatGPT 4.o and ChatGPT 5.6 including the pdf file with the response and reasoning for each real estate. The folder also includes the script (prompting) to use the API and the raw data (images).
License bundle
- Name:
- license.txt
- Size:
- 3.86 KB
- Format:
- Item-specific license agreed to upon submission
- Description:
Collections

