This study investigates the application of explainable AI (XAI) techniques to understand the deep learning models used for predicting urban conflict from satellite imagery. First, a ResNet18 convolutional neural network achieved 89% accuracy in distinguishing riot and non-riot urban areas. Using the Score-CAM technique, regions critical to the model’s predictions were identified, and masking these areas caused a 20.9% drop in the classification accuracy, highlighting their importance. However, Score-CAM’s ability to consistently localize key features was found to be limited, particularly in complex, multi-object urban environments. Analysis revealed minimal alignment between the model-identified features and traditional land use metrics, suggesting that deep learning captures unique patterns not represented in existing GIS datasets. These findings underscore the potential of deep learning to uncover previously unrecognized socio-spatial dynamics while revealing the need for improved interpretability methods. This work sets the stage for future research to enhance explainable AI techniques, bridging the gap between model performance and interpretability and advancing our understanding of urban conflict drivers.

ABSTRACTConflict, manifesting as riots and protests, is a common occurrence in urban environments worldwide. Understanding their likely locations is crucial to policymakers, who may (for example) seek to provide overseas travelers with guidance on safe areas, or local policymakers with the ability to pre‐position medical aid or police presences to mediate negative impacts associated with riot events. Past efforts to forecast these events have focused on the use of news and social media, restricting applicability to areas with available data. This study utilizes a ResNet convolutional neural network and high‐resolution satellite imagery to estimate the spatial distribution of riots or protests within urban environments. At a global scale (N = 18,631 conflict events), by training our model to understand relationships between urban form and riot events, we are able to predict the likelihood that a given urban area will experience a riot or protest with accuracy as high as 97%. This research has the potential to improve our ability to forecast and understand the relationship between urban form and conflict events, even in data‐sparse regions.

As both satellite imagery and image-based machine learning methods continue to improve and become more accessible, they are being utilized in an increasing number of sectors and applications. Recent applications using convolutional neural networks (CNNs) and satellite imagery include estimating socioeconomic and development indicators such as poverty, road quality, and conflict. This article builds on existing work leveraging satellite imagery and machine learning for estimation or prediction, to explore the potential to extend these methods temporally. Using Landsat 8 imagery and data from the Armed Conflict Location & Event Data Project (ACLED) we produce subnational predictions of the risk of conflict fatalities in Nigeria during 2015, 2017, and 2019 using distinct models trained on both yearly and six-month windows of data from the preceding year. We find that predictions at conflict sites leveraging imagery from the preceding year for training can predict conflict fatalities in the following year with an area under the receiver operating characteristic curve (AUC) of over 75% on average. While models consistently outperform a baseline comparison, and performance in individual periods can be strong (AUC > 80%), changes based on ground conditions such as the geographic scope of conflict can degrade performance in subsequent periods. In addition, we find that training models using an entire year of data slightly outperform models using only six months of data. Overall, the findings suggest CNN-based methods are moderately effective at detecting features in Landsat satellite imagery associated with the risk of fatalities from conflict events across time periods.

Human migratory decisions are driven by a wide range of factors, including economic and environmental conditions, conflict, and evolving social dynamics. These factors are reflected in disparate data sources, including household surveys, satellite imagery, and even news and social media. Here, we present a deep learning-based data fusion technique integrating satellite and census data to estimate migratory flows from Mexico to the United States. We leverage a three-stage approach, in which we (1) construct a matrix-based representation of socioeconomic information for each municipality in Mexico, (2) implement a convolutional neural network with both satellite imagery and the constructed socioeconomic matrix, and (3) use the output vectors of information to estimate migratory flows. We find that this approach outperforms alternatives by approximately 10% (r(2)), suggesting multi-modal data fusion provides a valuable pathway forward for modeling migratory processes.

The lack of systematic measurements of socioeconomic factors on a worldwide scale remains a significant challenge to our understanding of human well-being. A growing body of literature suggests that some of these measurement gaps can be filled using remote sensing, imputing human conditions on the ground based on the ways in which social groups have modified - or, not - their physical environment. In this article, we contribute to this growing body of literature by presenting a case study estimating school test scores based solely on publicly available imagery in both the Philippines (2010, 2014) and Brazil (2016). We contrast single image convolutional neural network (CNN) approaches to multi-source ensembles and find predictive accuracy for individual schools across years and regions ranging from 76% to 80%. Finally, we discuss broader considerations related to the operational use of CNN-based approaches for measuring socioeconomic factors, and provide open source computer code for community use.

Convolutional neural networks (CNNs) trained with satellite imagery have been successfully used to generate measures of development indicators, such as poverty, in developing nations. This article explores a CNN-based approach leveraging Landsat 8 imagery to predict locations of conflict-related deaths. Using Nigeria as a case study, we use the Armed Conflict Location & Event Data (ACLED) dataset to identify locations of conflict events that did or did not result in a death. Imagery for each location is used as an input to train a CNN to distinguish fatal from non-fatal events. Using 2014 imagery, we are able to predict the result of conflict events in the following year (2015) with 80% accuracy. While our approach does not replace the need for causal studies into the drivers of conflict death, it provides a low-cost solution to prediction that requires only publicly available imagery to implement. Findings suggest that the information contained in moderate-resolution imagery can be used to predict the likelihood of a death due to conflict at a given location in Nigeria the following year, and that CNN-based methods of estimating development-related indicators may be effective in applications beyond those explored in the literature.

We present the geoBoundaries Global Administrative Database (geoBoundaries): an online, open license resource of the geographic boundaries of political administrative divisions (i.e., state, county). Contrasted to other resources geoBoundaries (1) provides detailed information on the legal open license for every boundary in the repository, and (2) focuses on provisioning highly precise boundary data to support accurate, replicable scientific inquiry. Further, all data is released in a structured form, allowing for the integration of geoBoundaries with large-scale computational workflows. Our database has records for every country around the world, with up to 5 levels of administrative hierarchy. The database is accessible at http://www.geoboundaries.org, and a static version is archived on the Harvard Dataverse.

Since 1992, the Global Environment Facility (GEF) has mobilized over $131 billion in funds to enable developing and transitioning countries to meet the objectives of international environmental conventions and agreements. While multiple studies and reports have sought to examine the environmental impact of these funds, relatively little work has examined the potential for socioeconomic co-benefits. Leveraging a novel database on the geographic location of GEF project interventions in Uganda, this paper explores the impact of GEF projects on household assets in Uganda. It employs a new methodological approach, Quasi-experimental Geospatial Interpolation (QGI), which seeks to overcome many of the core biases and limitations of previous implementations of causal matching studies leveraging geospatial information. Findings suggest that Sustainable Forest Management (SFM) GEF projects with initial implementation dates prior to 2009 in Uganda had a positive, statistically significant impact of approximately $184.81 on the change in total household assets between 2009 and 2011. Leveraging QGI, we identify that (1) this effect was statistically significant at distances between 2 and 7 km away from GEF projects, (2) the effect was positive but not statistically significant at distances less than 2 km, and (3) there was insufficient evidence to establish the impact of projects beyond a distance of approximately 7 km.

Interdisciplinary use of geospatial data requires the integration of data from a breadth of sources, and frequently involves the harmonization of different methods of sampling, measurement, and technical data types. These integrative efforts are often inhibited by fundamental geocomputational challenges, including a lack of memory efficient or parallel processing approaches to traditional methods such as zonal statistics. GeoQuery (geoquery.org) is a dynamic web application which utilizes a High Performance Computing cluster and novel parallel geospatial data processing methods to overcome these challenges. Through an online interface, GeoQuery users can request geospatial data - which spans categories including geophysical, environmental and social measurements - to be aggregated to user-selected units of analysis (e.g., subnational administrative boundaries). Once a request has been processed, users are provided with permanent links to access their customized data and documentation. Datasets made available through GeoQuery are reviewed, prepared, and provisioned by geospatial data specialists, with processing routines tailored for each dataset. The code used and steps taken while preparing datasets and processing user requests are publicly available, ensuring transparency and replicability of all data and processes. By mediating the complexities of working with geospatial data, GeoQuery reduces the barriers to entry and the related costs of incorporating geospatial data into research across disciplines. This paper presents the technology and methods used by GeoQuery to process and manage geospatial data and user requests.

Governments use a variety of policies to increase the impact of foreign investment on economic growth. An increasingly popular policy is to require that foreign companies provide public goods near the communities where their commercial investments are sited. This approach seeks to crowd in additional investments, create clusters of interconnected firms, and set in motion economic agglomeration processes. Post-2006 Liberia represents an ideal empirical setting to test the effectiveness of this approach. We construct a new dataset that measures the precise locations of 557 natural resource concessions granted to investors. We then merge these data with a remotely sensed measure of nighttime light growth at the 1 km x 1 km grid cell level and analyze it using a matched difference-indifferences strategy. We find heterogeneous treatment effects across sectors and investor types: mining (specifically iron-ore) investments projects have positive growth effects, while agriculture and forestry investment projects do not; furthermore, concessions granted to Chinese investors have positive growth effects while those given to U.S. investors do not. These patterns of heterogeneous treatment effects across sectors and investor types are consistent with the theory of change underpinning the government's development corridor strategy. (C) 2018 Elsevier Ltd. All rights reserved.

The policy choices of local governments are highly relevant today, but we know relatively little about how or when local governments choose to respond to a given issue and why this might vary between policy areas. A key variant for local governments is the proximity of policy issues: they are engaged in solving local, regional, and global problems. Using evidence from the United States on the policy issues of social inclusion, watershed management, and climate change, we demonstrate that the drivers of policy response vary with the proximity of the problem. When an issue is highly local, policy response is influenced by problem severity; when an issue is global, policy response is influenced by local political leanings; and when an issue is regional, policy response is driven by the actions of neighboring and state level governments. Local governments consider different factors and respond to different cues when engaging with different types of policy issues. Our findings provide a more nuanced understanding of sustainability policy adoption in local governments, and further our understanding of the domain-contingent nature of policy response in local governments and the structuring role of problem proximity.

We evaluated the local impacts of World Bank development projects on sites of recognized conservation significance (Important Bird and Biodiversity Areas [IBAs]) using tree cover change data and in situ state, pressure, and response monitoring data. IBAs adjacent to World Bank project locations and a matched set of IBAs distant from World Bank project locations had similar rates of tree loss and similar in situ measurements of conservation outcomes. Thus, we did not detect any significant net negative impacts of World Bank projects on tree cover or conservation outcomes. These results are encouraging because 89% of World Bank projects that are close to IBAs are environmentally sensitive projects (so-called Category A and Category B projects) subjected to the organization’s most stringent safeguards. However, the limitations of our evaluation design do not allow us to rule out the possibility that World Bank projects had positive or negative effects that were undetectable.

There has been considerable debate regarding the efficacy of international aid in meeting the dual goals of human development and environmental sustainability. Many donors have sought to engage with this challenge by introducing environmental safeguard and monitoring initiatives; however, evidence on the success of these interventions is limited. Evaluating aid is a particular challenge in the case of donors that do not disclose information on the nature, geographic location, or extents of their interventions. In such cases, new methods that extract and geoparse data on the activities of opaque donors through the manual interpretation of thousands of news and other articles allow us to investigate the impacts of these activities. However, residual spatial uncertainty in these data remains a potential source of bias. In this article, we apply and discuss a Geographic Simulation and Extrapolation (GeoSIMEX) approach to mitigate the spatial imprecision inherent in geoparsed data. In conjunction with GeoSIMEX, we test and contrast multiple approaches to reducing the imprecision of aid, including high-assumption cases in which other covariates (i.e., nighttime lights) are leveraged to allocate aid. In our application, we find that methods which do not account for spatial imprecision find statistically significant relationships between Chinese aid and vegetation change; after accounting for spatial uncertainty, findings are similar for Rwanda and inconclusive for Burundi.

Quantifying the impact of an intervention or treatment in a real setting is a common and challenging problem. For example, we would like to calculate the environmental implications of aid projects in third world countries that target economic development. For causal inference problems of this kind, the Rubin causal model is one of several popular theoretical frameworks that comes with a set of algorithmic methods to quantify treatment effects. However, for a given data set, we neither know the ground truth nor can we easily increase the size of the data set. So, simulation is a natural choice to evaluate the applicability of a set of methods for a particular problem. In this paper, we report findings of a simulation study with four causal inference approaches, namely two single tree approaches (transformed outcome tree, causal tree), and two random forest versions of the former.

Objective: Cross-national studies provide inconclusive results as to the effectiveness of foreign health aid. We highlight a novel application of using subnational data to evaluate aid impacts, using Malawi as a case study. Design: We employ two rounds of nationally representative household surveys (2004/2005 and 2010/2011) and geo-referenced foreign aid data. We examine the determinants of Malawi's traditional authorities receiving aid according to health, environmental risk, socioeconomic and political factors. We use two approaches to estimate the impact of aid on reducing malaria prevalence and increasing healthcare quality: difference-in-difference models, which include traditional authority and month-of interview fixed effects and control for individual and household level time-varying factors, and entropy balancing, where models balance on health-related and socioeconomic baseline characteristics. General health aid and four specific health aid sectors are examined. Results: Traditional authorities with greater proportions of individuals living in urban areas, more health facilities and greater proportions of those in major ethnic groups were more likely to receive aid. Difference-in-difference models show health infrastructure and parasitic disease control aid reduced malaria prevalence by 1.20 (95% CI -0.36 to 2.76) and 2.20 (95% CI 0.43 to 3.96) percentage points, respectively, and increased the likelihood of individuals reporting healthcare as more than adequate by 12.1 (95% CI 1.51 to 22.68) and 14.0 (95% CI 0.11 to 28.11) percentage points. Entropy balancing shows similar results. Conclusions: Aid was targeted to areas with greater existing health infrastructure rather than areas most in need, but still effectively reduced malaria prevalence and enhanced self-reported healthcare quality.

Since 1945, over $4.9 trillion dollars of international aid has been allocated to developing countries. To date, there have been no estimates of the regional impact of this aid on the carbon cycle. We apply a geographically explicit matching method to estimate the relative impact of large-scale World Bank projects implemented between 2000 and 2010 on sequestered carbon, using a novel and publicly available data set of 61,243 World Bank project locations. Considering only carbon sequestered due to fluctuations in vegetative biomass caused by World Bank projects, we illustrate the relative impact of World Bank projects on carbon sequestration. We use this information to illustrate the geographic variation in the apparent effectiveness of environmental safeguards implemented by the World Bank. We argue that sub-national data can help to identify geographically heterogeneous impact effects, and highlight many remaining methodological challenges.

Concerns over the expropriation of and encroachment on indigenous communities' lands have led to greater formalization of these communities' rights in a number of developing countries. We study whether formalization of indigenous communities' land rights affects the rate of deforestation in both the short and medium terms. Beginning in 1995, the Government of Brazil formalized the rights of several hundred indigenous communities whose lands cover more than 40 million hectares in the Amazon region and provided support for these rights’ enforcement. We study the program's impacts using a long time-series of satellite-based forest cover data. Using both plausibly exogenous variation in the timing of formalization and matched samples of treated and comparison communities, we find no effect of these protections on satellite-based greenness measures. This is true even for communities that received support for surveillance and enforcement of these rights. Notably, we observe low counterfactual rates of deforestation on communities' lands between 1982 and 2010, suggesting that indigenous land rights programs should not uniformly be justified on the basis of their forest protection, at least in the medium term.

There is a large and growing set of literature examining how different classes of models can integrate information on spatial imprecision in order to more accurately reflect available data. Here, we present a flexible approach geoSIMEX which can provide parameter and error estimates while adjusting for spatial imprecision. We illustrate this approach through a case study leveraging a novel, publically available dataset recording the location of Chinese aid in Southeast Asia at varying levels of precision. Using a difference-in-difference modeling approach, we integrate Chinese aid information with satellite derived data on vegetation (NDVI) to examine if Chinese aid has caused an increase or decrease in vegetation. Followingmultiple approacheswhichdonot incorporate spatial imprecision,we find that Chinese aid had a negative impact on vegetation; once spatial imprecision was incorporated into our estimates through the geoSIMEX procedure no evidence of impact is found.