Analyzing Conflict from Space

The "la Caixa" Foundation supports ground-breaking research at the Barcelona School of Economics through the "la Caixa" Foundation Research Grants on Socioeconomic Well-being.

Project overview

With the ongoing refugee crisis in the Turkey, the Middle East and North Africa, there is a renewed interest in understanding the causes for fight and the triggers of conflict. At the same time, the Syrian civil war has provided formidable challenges for data gathering due to the extent of the violence and the politicization of violence data. This project tries to help in this effort through the generation of a new method to gather data on violence. Specifically, the project seeks to provide a new, automatized way to capture fighting and destruction using satellite imagery, build a panel dataset using this method, and then illustrate what can be learned from this alternative measure of violence.

Main results

  • Proof-of-concept study of automated destruction classification directly tailored towards data generation and implementation in unbalanced samples
  • First team to generate a completely new database of destruction data at a highly disaggregated dimension within a city
  • The team's data opens the door to the identification of media biases through the provision of a completely alternative dataset

Summary, output, and dissemination

Research summary

There is a huge interest in the methodology of automated destruction identification through satellite images. Despite considerable interest by governments, NGOs and international organizations, the problem has not been solved by any public actor yet. There are academic papers on the issue but no transparent implementation of an automated remote sensing in practice, and remote sensing in practical applications is conducted by hand. This project, therefore is truly breaking new ground.

The research team decided to develop its own program without relying on help from private firms or satellite companies. This turned out to be a challenging endeavor and resulted in a complex distribution of tasks both among team members and research assistants. Apart from the human resources and code development, a main bottleneck of the project was the availability of server and computation resources.

Initial results

The researchers now understand the dramatic disparity between successful academic papers on automated destruction detection and the use of hand-coding in practice. The problem faced by policy makers is that a very small share of the housing stock gets destroyed. Even in a city like Aleppo less than 3 percent of all patches show destruction in the ground-truth. This means that even if precision is very good in the 1:1 test sample, the predictions in actual applications can be essentially useless. Therefore, the team's objective in the given context is to maximize the precision of its method, which is the share of true positives (correctly predicted destruction) over the sum of true positives and false positives (incorrectly predicted positives). Based on recent results, the researchers anticipate that the precision performance of the method will be substantially better than those of existing academic article published on a similar problem so far.

Based on the current version of the method, the team have been able to classify destruction in Aleppo in the available imagery from Google Earth already. They have been able to generate geo-referenced destruction data for the whole of Aleppo for 19 distinct points in time during the Syrian civil war. In the next step they will combine generated ground truth data with the media reporting data to start exploring potential biases in war reporting about Aleppo.

Based on preliminary results of the media reporting data, the researchers find that it differs dramatically from the UNOSAT destruction scores across Syrian cities. Some cities like Homs and Hama had comparatively little media coverage whereas Aleppo is covered much more. This motivates the team's approach. If reporting differs so dramatically across cities, then methods that rely on reporting to measure violence will be biased.

Looking at different sources of media reporting from the two big alliances present in Syria (Russia and the Syrian regime vs. the USA and its Western allies), the researchers have identified very interesting patterns which may indicate the direction and strength of this bias: The main driver of news reporting, both for Russian and Western news sources, was engagement in the war. The capture of area control by Syrian government forces is negatively and significantly correlated with reporting from media sources of both alliances, but the magnitude of this effect is much stronger for Syrian allies than for Western allies’ news sources. On the other hand, when the opposition gains ground, there is a positive and significant correlation with news reporting on the side of the Syrian allies, while Western allies do not change the extensive margin of war reporting. The researchers are currently exploring how to correctly interpret these findings and believe that they reflect particular instances of reporting biases by media outlets from different sides of the conflict in response to specific war actions unfolding across Syrian localities.

Contributions to date

The main contribution to date is the proof-of-concept study of automated destruction classification which is directly tailored towards data generation and the actual implementation in unbalanced samples. The existing literature in remote sensing does not stress the problem of extreme unbalancedness sufficiently in our view. Typically 1:1 statistics are discussed without a critical evaluation of performance in heavily unbalanced samples where destruction appears only a few percent of all images. In Economics this is one of the first papers that incorporates this new technology of remote sensing. Currently, only working papers exist on this issue, and this is certainly the first project to apply this in the area of conflict studies – an area with considerable policy interest.

The team is the first to generate a completely new database of destruction data at a highly disaggregated dimension within a city. Existing data based on media reports is not able to distinguish the within-city spatial dimension of violence due to restrictions in the precision of the reports. Only very few media reports provide exact locations – typically if important buildings were hit. Here is where the team's approach holds particular merit: they have achieved tracking of destruction across space.

The team's data opens the door to the identification of media biases through the provision of a completely alternative dataset. They have analyzed the reporting data and have found that reporting differs across time and sources. With the data they can show which of the sources is closer to ground truth measures in quantitative terms. This adds to the existing literature on media biases.

Apart from that, the new data which the team's methods generate also holds considerable potential to be employed in the context of other applications within and beyond the Aleppo case study, such as the one of forced migration and internal displacement as well as the evolution of conflict at the micro-level. The work during the last two years has laid the foundation for further studies along these lines and the researchers are certain that it will lead to rich output in terms of academic publications.

National and international funding
  • Faculty Opportunity Fund, Chapman University (United States)
  • Computer Vision Centre (CVC) at UAB and Chapman University both donated server resources to the project
  • Additional support from IAE-CSIC and UAB Department of Economics and Economic History
Conferences, workshops, and seminars


  • AI for Development conference organized by CEGA and the World Bank Development Impact Evaluation Group (DIME), March 2018
  • Empirical Studies of Conflict (ESOC) Annual Meeting organized by Princeton University, May 2018 
  • New York R conference, April 2018 [video]
Awards and Recognitions

La Caixa Foundation. Together, we bring projects to life.