This paper analyses the evolution of French urban areas from a historical perspective. Using building footprints extracted from collections of digitised historical maps covering mainland France in the 18th, 19th and 20th centuries, we define consistently urban areas and analyse their trajectories along the urban hierarchy. Our findings highlight increasing urbanisation with fewer and larger urban areas, consistent with agglomerations economies. Disaggregate analysis reveals significant heterogeneity, with the emergence, persistence, disappearance of urban areas.

This paper develops methods to estimate the unobserved knowledge networks from observed similarities in visual artworks. Deep neural auto-encoders have the ability to project high-dimensional data points such as images, into latent spaces where the distances between each pair of points is well defined. This property can be exploited to derive meaningful measures of distance between artworks and to quantify accurately similarity and by extension, novelty. Using large datasets of digitised historical paintings, we analyse the change in the spatial patterns of knowledge diffusion in artistic styles in response to policy and transport shocks.

This paper analyses paintings as expressions of their times and contexts. We exploit large repositories of paintings with time and location information and develop an interpretable convolutional network extracting the emotions conveyed through paintings. The geography described by emotions sheds light on economic development and structural changes experienced by those locations. We use the estimated mapping to predict political turmoil and economic change, and provide novel proxies for economic development, inequalities and uncertainty over time and across locations.

Existing data on building destruction in conflict zones rely on eyewitness reports or manual detection, which makes it scarce, incomplete, and potentially biased. We introduce a new siamese convolutional transformer network architecture measuring destruction in high-resolution satellite images, while explicitly taking into account the spatial and temporal structure of the data, while relying on comparison with pre-conflict images to increase predictive performance. Our approach allows measuring conflict intensity with unprecedented scope, resolution, and frequency.

Extracting features such as buildings, land-use or transportation networks depicted in historical maps provides novel insights for urban and spatial economics. Using image segmentation techniques and random forest models, these features are extracted automatically using the spectral and spatial information available at the pixel level. Our methodology is successfully applied to a collection of 19th century maps covering mainland France, and proves robust to considerable amount spectral heterogeneity both across maps and within classes. 

This paper studies the economic geography of the North American fur trade between 1670 to 1870. European hat makers used felt made from beaver pelts to produce hats for the elite. With the Eurasian beaver nearly extinct, Europeans turned to North America, trading beaver pelts with Indigenous communities in exchange for iron goods and textiles. We introduce a computational spatial model that integrates a biological beaver growth model with detailed transport cost calculations across North America’s extensive river system, forming a "beaver gravity model". The model illustrates that the over-exploitation of beavers drove the westward expansion of the fur trade, simultaneously shedding light on spatial dynamics of Indigenous populations.

With Stephan Heblich and Stephen J. Redding

There is a growing interest among applied economists in using machine learning models for predictive applications. Particularly promising implementations revolve around the collection of original research data for subsequent economic analysis. In this perspective, a family of models called neural networks have proven particularly effective in dealing with high-dimensional and unstructured data, such as text or images. Despite the appeal, the lack of familiarity with algorithmic modelling has largely discouraged their application in applied economic research. This paper introduces neural networks using an approach and notation familiar to the applied economist. It provides a comprehensive understanding of their mechanisms, the problems to which they can be usefully applied, and their limitations.

This paper contributes to the emerging literature on the accurate and consistent definitions of cities by providing, for 53 African countries, a detailed comparison of the delineations of urban areas resulting from the application of two of the frontrunners among the new methods. We provide a detailed comparison of results from the application of the “degree of urbanization” approach (Dijkstra and Poelman 2014) with results from the "dartboard" (de Bellefon et al. 2021) approach for delineating urban areas.

This project repository contains the scripts to delineate urban areas using the methodology developed in Combes et al. (2021) as well as some recent developments including a much more scalable and computationally efficient implementation. These methods have since been used by the Banque de France and the World Bank, among others, to delineate urban areas in a large number of countries, using population and building densities.

This project computes historical boundaries of French communes and arrondissements for every year during the period 1793 - 2006. The resulting geometries closely approximate the borders depicted on historical maps, while remaining fully compatible with current projections. The numerous records of the Cassini project contain the disaggregated administrative changes for every commune that existed during this period. This information is collected automatically to track the administrative status of every record at any point in time. Existing records are then matched to current geometries and processed to compute the most precise geometries using the information available. 

The increasing availability of large-scale and open databases offers a strong potential for original economic analysis. OpenStreetMap is a geographical database of the world containing a wide array of public spatial datasets and user-contributed information, harmonised and available through a single platform. Such sources can be used to collect disaggregated data for economic analysis on a large scale. Possible operations include geocoding, the computation of accessibility measures or the collection of city-specific data on amenities, among others.