Historical Image Analysis with Machine Learning: The WWII Raid Damage Evaluation

• In Event: Images as Data

Abstract

This study demonstrates the methodological utility of geographical image analysis and machine learning techniques in quantifying damages induced by political violence from historical imagery, using the Bombing of Tokyo during the World War II as a motivating example. Although a growing number of studies utilize satellite imagery and aerial photography to extract geographic information of their interests, most studies rely on contemporary data sources (Donaldson and Storeygard, 2016). In addition, the recent advances in conflict research have increasingly utilized historical archives to measure historical and local-level variations of wartime violence, but they typically exploit text-based records of political violence (e.g., Zhukov, 2017). Despite its promises, historical geographic imagery has hardly been utilized as a data source for political science study. This study is among the first attempts that utilize image analysis and machine learning techniques to capture the historical geography of wartime violence and explore its long-run impact on contemporary sociopolitical attitudes and behavior.

In Harada and Ito (n.d.), we made a first attempt to fill this paucity of the literature through an empirical investigation into the long-run impact of the Bombing of Tokyo on contemporary socioeconomic outcomes, which is the most destructive political violence with conventional weapons in terms of its civilian casualties (Saotome et al., 1981). We utilized the arguably exogenous, as-if random variations in the neighborhood-level raid damages that vary mainly due to differences in the landscape and microclimates on the bombing days. Our analysis exploits the exogenous variations, quantified with image analysis techniques, to isolate the long-term causal effects of the WWII political violence. Specifically, we developed an original dataset by utilizing the aerial imagery of 1946-1948 Tokyo, historical maps, and related historical archives. Relying on the image-based information and a human-coding scheme, the dataset codes the level of raid damages at the Cho-chomoku or neighborhood level, of which average size is only 1.6% of the typical “village-level” units of observation such as Vietnam’s hamlet with a 2km radii. We also compiled the information on possible treatment assignment mechanisms, or the geodesic distances from the aiming points of the US strategic bombers and the Imperial Palace, from the declassified documents of the US Air Force.

This paper extends the initial data development effort in two ways. First, while the initial dataset relies on a human-coding scheme, we introduce the geographic object-based image analysis (GEOBIA) techniques from the remote-sensing literature to develop a machine-coded version of the raid damage dataset. The GEOBIA approach mimics human visual perception to classify objects within images (Blaschke et al., 2014). The classification process first segments and groups small pixels together into vector objects (image segmentation). It then assigns individual objects into different land cover classes such as built-up areas using the information of colors and shapes of individual objects and their neighbors (object classification). This approach is particularly valuable when extracting relevant information from historical imagery with relatively limited information of resolution and colors.

Next, we evaluate the performance of the object-based, machine-coded dataset of the raid damage against the human-coded evaluation. We evaluate the machine-coding performance in two ways. First, since no historical ground truth is available at this moment, we follow the machine-coded event data development efforts (e.g., King and Lowe 2003) to compare the correspondence between the human-coding and machine-coding evaluations and how the latter successfully replicates the coding of the former. Second, we compare the performance of the two coding approaches by comparing the effect estimates of raid damages on contemporary socioeconomic outcomes. The main dependent variables derive from our original geocoded and web-based survey on political attitudes and behavior as well as precisely disaggregated and geocoded version of the Japanese Census and crime rates in contemporary Tokyo. If the machine-coded raid damage evaluation contains a greater magnitude of measurement errors, we would observe attenuated coefficients for the effect estimates. The preliminary results show that the effect estimates of the machine-coded raid damage evaluation yield the same coefficient signs as the human-coded evaluation while the coefficient sizes of machine-coded evaluation remain smaller in absolute terms, indicating greater measurement errors. These results demonstrate how the machine-enabled image-classification techniques help to extract relevant information from historical imagery in evaluating the historical geography of political violence.

Authors

• Masataka Harada, Fukuoka University

• Gaku Ito, Hiroshima University