HUM data lab > Workflows > Spatial and Temporal Patterns of the Photograph Collection of the EFA

Spatial and Temporal Patterns of the Photograph Collection of the EFA

The aim of this workflow is to study the spatial and temporal patterns of the photograph collection of the Estonian Folklore Archives (EFA) and to make a large and difficult-to-grasp dataset easier to explore through an interactive visualisation tool. The project focuses on the first 10,000 black-and-white photographs from the EFA photograph collection, which contains approximately 88,000 photographs in total, together with their metadata. On the basis of this material, the distribution of photographs is analysed across time, space, keywords, collection projects, persons, and photographers. The metadata used for the analysis were obtained from the digital photograph collection of the Estonian Folklore Archives as CSV exports. In terms of location, particular attention is paid to historical parishes, since the location information for many photographs is linked to a parish, or geolocation has only been possible at the level of the parish centre point.


The workflow combines digital humanities methods, metadata analysis, and machine learning. Since the application does not include the images themselves due to copyright, data volume, and technical limitations, the machine learning component focuses on the analysis of keywords associated with the metadata. For this purpose, the CLIP library was used to investigate whether the machine can identify human-assigned keywords and assign categories to images without keywords. The result of the project is an analytical report and an interactive Streamlit dashboard application, available at the following website: https://era-dashboard.streamlit.app/..

Workflow steps

Keywords: Discovering, contextualizing

The structure and metadata of the EFA photograph collection are examined: the available CSV files, the meanings of the fields, data formats, and data quality are analysed. Particular attention is paid to the spatial and temporal information available for the photographs and to the limitations arising from historical collection practices. On this basis, the research questions are formulated: how are the photographs distributed across time and space, which keywords recur in different regions, and whether the dataset reveals collection waves or regional specificities.

Keywords: Gathering, organizing

The metadata of the photograph collection were located in several tables. During the aggregation process, one large table was created, which was later grouped into related CSV tables, such as the main photograph data, keywords, persons depicted in the photographs, and locations. The tables are linked using each photograph’s unique identifier — the PID identifier, a sequence of numbers and letters. Column names, data types, and data structures are standardised. In addition, new variables are created for the analysis, such as categories for persons and keywords. A unified logic is created for handling spatial data: where precise geolocation is missing, the coordinates of the centre points of historical parishes are used.

Keywords: Data cleansing, annotating

Data cleaning is carried out iteratively throughout the workflow. Dates, place names, personal names, and the spelling of keywords are standardised. A large part of the cleaning is done in Google Colab and manually in the OpenRefine environment. The original 517 keywords were manually grouped in OpenRefine into 19 broader thematic categories in order to reduce noise caused by spelling variation and to enable the analysis of more general content-related patterns. The categories were created on the basis of recurring themes in the dataset, grouping keywords with similar semantic fields into unified thematic groups. For example, the category “person” includes performer, collector, child, mother-in-law, and others.

The workflow is not linear but iterative: problems that emerge during analysis and visualisation create the need to return to the data cleaning stage.

Keywords: Analyzing, spatial analysis, discovering

The distribution of the photograph collection is analysed across time, space, and themes: which regions and periods contain the largest number of photographs, which keywords dominate, and how photographers and collection projects are distributed. Pattern detection is based on descriptive statistics, temporal distributions, spatial visualisations, and filter-based comparative analysis.

Keywords: analyzing

The CLIP library is used to assess how well the existing 19 keyword categories can be distinguished through machine learning. CLIP is a machine learning model created by OpenAI that links images and textual descriptions within a shared semantic space. The purpose of the machine learning component is not to create a final automatic tagging system, but to compare the existing metadata with the model’s predictions and to assess for which themes CLIP is best able to identify visual patterns. The results help evaluate the consistency of the metadata and provide a basis for the future development of automatic tagging solutions.

The machine learning results were not automatically used in the filtering system of the Streamlit application, but were treated as a separate experimental analysis for assessing the quality and consistency of the existing metadata.

Keywords: Data visualization, design

An interactive dashboard application is created in Streamlit. Streamlit is a Python-based tool for creating interactive web applications for data analysis. The application allows users to explore the photograph collection through filters: photographer, keyword category, keyword, person depicted in the photograph, genre, and time period. The central element of the application is a parish-based map that visualises the spatial distribution of the photographs using parish centre points. Temporal distributions and statistical summaries are also displayed. The application makes it possible both to analyse the research results and to present them publicly without displaying the original photographs.

Keywords: Contextualizing, publishing

The results of the analysis are interpreted in a cultural-historical and archival context: how collection practices, historical events, and regional specificities shape the patterns found in the dataset. The limitations of the dataset are also critically assessed, including collection bias and incomplete metadata.

The completed workflow, code, and analysis are documented on GitHub and in the Estonian Literary Museum’s Webrepo.. A report and presentation are prepared on the main results: spatial patterns, temporal distribution, keyword occurrence, and the results of the machine learning analysis.