Nature communications - modeling armed conflict risk under climate change using machine learning and time series data

The article quoted ：（ The code of this article is attached in the message area ）

Ge, Q., Hao, M., Ding, F. et al. Modelling armed conflict risk under climate change with machine learning and time-series data. Nat Commun 13, 2839 (2022). https://doi.org/10.1038/s41467-022-30356-x

Abstract ：

Understanding the dangers of armed conflict is essential to promoting peace . Although for decades , Academia has been studying the relationship between climate variability and armed conflict , Using quantitative and qualitative methods on different spatial and temporal scales , But causality on a global scale is still poorly understood . ad locum , We adopt a quantitative modeling framework based on machine learning , Infer potential causality from high frequency time series data , And simulate 2000 - 2015 The risk of global armed conflict in . Our results show that , The risk of armed conflict is mainly affected by a stable background environment with complex patterns , The second is the covariate related to climate deviation . The inferred pattern indicates that , Positive temperature deviations or extreme precipitation are associated with an increased risk of armed conflict worldwide . Our results show that , Better understand the link between climate and conflict on a global scale , It can enhance the spatio-temporal modeling ability to deal with the risk of armed conflict .‎

---

Discuss ：

‎O'Loughlin Previous studies by et al , The risk of conflict is related to climate anomalies , But it is more influenced by politics , The impact of socio-economic and geographical environment , Especially in sub Saharan Africa .‎‎20‎‎,‎‎31‎‎. Our results show that , There are similar patterns around the world . for example , Stable background covariates （ See Supplementary information ） It has greatly promoted the temporal and spatial distribution of armed conflict events , The average relative contribution exceeds 96.0%（ Supplementary table ‎‎7‎‎）. Compared with stable background covariates , The normalized temperature index or normalized precipitation index has relatively little effect on the simulation results , However, the cumulative covariates related to climate deviation account for 2.5% above . This is why the simulated risk level of armed conflict events in local areas varies in different years . We interpret this as evidence of the impact of climate change on the risk of conflict . Supplementary table ‎‎3‎‎ indicate , Considering the two-year climate deviation , Can slightly improve BRT The performance of the model . This result can be seen as partially supporting other findings , I.e. normal climatic conditions （ Social adaptation ） The deviation for many consecutive years may affect the stability of the society in part or as a whole , Whether in history .‎‎32‎‎,‎‎33‎‎ And current time period ‎34‎. Supplementary table ‎‎7‎‎ Show , Long period climate deviation has a great impact on the risk level , The relative contribution value is 3.806%. This is in the lower range , But it is consistent with the judgment of experts in different disciplines , namely 3-20% The risk of conflict is related to climate change .‎9.

‎ Although more and more quantitative studies have found that climate change has an impact on the incidence of armed conflict , But the evidence about climate change and the outbreak of armed conflict is more scarce and controversial .‎‎10‎‎,‎‎23‎‎,‎‎35‎‎. therefore , Our research not only simulates the possibility of armed conflict , And further discussed the feasibility of simulating the outbreak of armed conflict . be based on van Weezel Adopted definitions ‎36‎, We constructed an incidence and an incidence index to represent the risk of conflict , The modeling and analysis are carried out respectively . The results further show that , Combine machine learning with high frequency time series data , It has great potential in predicting the risk of the outbreak of armed conflict on a global scale （ Supplementary drawing ‎‎4‎‎、‎‎17‎‎ and ‎‎18‎‎）. Besides , Our results also show that , On a global scale , The occurrence of armed conflict is more sensitive to climate change than the occurrence of armed conflict , Such as supplementary table ‎‎7‎‎ and ‎‎8‎‎ Shown .‎

‎ Our procedures allow the quantification of the relationship between covariates and armed conflict on a global scale . Overall speaking , Discovered patterns from large amounts of data are complex . That's why , Because of different weather 、 Geography 、 The political and socio-economic background may make human beings adapt to environmental pressures differently .‎‎37‎‎,‎‎38‎‎, As a result, the social stability response to climate change varies . however , There are several general patterns , As shown in the supplementary figure ‎‎6‎‎、‎‎7‎‎、‎‎9‎‎ and ‎‎10 Shown ‎‎. for example , The positive correlation between the level of conflict risk and ethnic diversity shows that , The greater diversity of politically related races leads to a higher risk of conflict , This is consistent with several previous studies .‎‎21‎‎,‎‎39‎‎,‎‎40‎‎,‎‎41‎‎. meanwhile , There is a positive correlation between the level of conflict risk and urban accessibility , This shows that the transportation hub can easily become the outbreak of conflict , Because they play a key role in controlling territory and conflict Logistics .‎‎42‎‎,‎‎43‎‎. For the covariates related to climate bias , Some research shows that , Negative temperature deviations in temperate regions may lead to various forms of conflict , Negative precipitation deviation coincides with social instability ‎‎23‎‎,‎‎44‎‎,‎‎45‎‎,‎‎46‎‎. However , Due to the improvement of technology adaptability and the increase of social structure complexity , The adaptation of modern humans to climate change is much higher than that recorded in historical studies . However , Climate change is still likely to exceed specific regions （ for example , When they are remote and dependent on Agriculture ） Or group （ for example , When they are poor and politically excluded ） The ability to adapt . This makes from a single case （ For example, the Syrian civil war ） The inference of the global scale drawn from . chart ‎‎2‎‎ And supplementary pictures ‎‎8‎‎、‎‎11‎‎ and 12 indicate , from 2000 Year to ‎‎2015‎‎ year , Positive temperature deviations or extreme precipitation are associated with an increased risk of global armed conflict . This proves that the results of other studies are correct , for example Hsiang Et al .‎19‎, Mach et al .‎9‎, And Herman and Zach ‎47‎. Besides , Our results show that , Compared with the global precipitation deviation , The non-linear impact of temperature rise on the incidence of armed conflict and the risk of armed conflict outbreak is greater .‎

‎ Based on high-dimensional data sets and a large number of occurrence records , We make use of BRT Model , Under four strategies , The global incidence of armed conflict and the incidence of armed conflict are simulated at the grid year level （0.1°×0.1°） The risk of . On a global scale ,2000 - 2015 The distribution of annual conflict risk shows obvious spatial agglomeration characteristics , These models can well simulate . The simulation results depend on the distribution of the samples . In order to improve the simulation accuracy , Reduce the impact of low-risk samples , We repeat the random selection of low-risk samples 20 Time , And based on each sample set BRT The process of modeling . The uncertainty level maps associated with these simulations are based on 20 Integration BRT Generated in the model for the calculated standard deviation value of each grid , These models are shown in the supplementary figure ‎‎21-28‎‎ Described in the .‎‎ The uncertainty level diagram shows that the simulation uncertainty is low .‎

‎ In this study , There are some warnings . First , Media reports represent UCDP GED A data source for , The well-known media bias may increase the uncertainty of our results to some extent . Although several measures have been taken （ Triple check ） To ensure the high quality of the final data set ‎4‎,UCDP It can't be solved completely GED Prejudice in , And include all incidents of armed conflict in its data set . secondly , Our analysis is based on a global scale multidimensional spatiotemporal refined data set . Due to the lack of sophisticated data sets for cultural and historical factors , Our training of machine learning models is limited in quantifying the effects of these variables . however , As more samples of machine learning models are available for training , Our sophisticated analysis on a global scale can help models capture more reliable relationships . Third , There is no general theory to explain the global climate - The causal mechanism of conflict connection , But our modeling framework may be helpful for early warning of conflict risk . Comparative analysis shows that , The predicted risk of armed conflict events in Africa （ Supplementary drawing ‎‎29b‎‎） And Hegre The risk level estimated by et al. Is generally the same .‎26‎. in general , Our research provides a global climate - A better understanding of conflict connections , And enhanced the spatio-temporal modeling capability of global armed conflict risk .

translate ：Google translator‎

Fig. 1: Validated performance on a time scale of the boosted regression tree models.

Fig. 2: Marginal effect curves of each climate deviation related covariate.

Fig. 3: Maps of the global simulated risk of armed conflict incidence at 0.1° × 0.1° spatial resolution.

Statement ： Welcome to reprint 、 Forward the original content of this number , You can leave a message in the message area or contact the editor in the background ( WeChat ：gavin7675) To authorize . Meteorologists official account for information dissemination. , The content is the responsibility of the author , Doesn't represent this point of view . Some of the pictures in this article come from the Internet , If it involves the content of the work 、 Copyright and other issues , Please contact Xiaobian in the background for processing .