Predicting violent conflict in Africa - Leveraging open geodata and deep learning for spatiotemporal event detection
Violent conflicts endanger human lives, the social cohesion of societies and the natural environment. While the number of intensive international conflicts has remained on a low level during the 21st century, civil wars are on the rise. Since the 1990s, research engages in predicting the outbreak of violence. However, findings on the role the natural environment plays in the emergence of violence remain mostly inconclusive. In order to contribute to the discussion this thesis sets out to compare the predictive performance of deep learning models using data from the Uppsala Conflict Data Program (UCDP) on civil conflict between 2001 to 2019. The data is simultaneously aggregated on administrative districts and sub-basin watersheds and combined with socio-economic and environmental predictors. The hyperparameters of CNN-LSTM architectures are optimized employing a Bayesian Optimization strategy. The results in terms of F2-score suggest significant improvements for aggregating predictors on sub-basin watersheds (+7.16,p=3.4e-11) as well as integrating environmental predictors (+3.98,p=5.9e-05) for a combined conflict class. For other conflict classes, the results tend to the same direction but are not significant. Through the comparison to existing conflict prediction tools, the thesis exposes the sensitivity of prediction models to spatial scale and units of aggregation. It is argued that in order to fulfill the requirements of effective conflict prevention efforts, prediction research will have to fully integrate modern deep learning frameworks and constant data streams on different earth processes in the future.
This thesis was submitted to the Department of Geography, University of Marburg, in partial fulfillment of the requirements for the degree of M.Sc. Phyisical Geography. It was written by customizing the {huwiwidown} template of the Berlin School of Buisness and Economics, HU Berlin. The online version of the thesis just wraps the original R Markdown files which were written to produce LaTex and builds a workflowr project out of it. By the conversion to html some outputs my render not totally as expected. However, you can download the original pdf version from here. All code available in this repository is licensed under GPL-3.