DATA-150

Arthur Kim

Professor Brewer

21 Nov 2021

Gender Disaggregated CDR’s and Climate Change: Data Science Methods

Word Count: 1611

Because climate change is a dire issue affecting all of humanity, much research has gone into understanding how it affects populations and human development, as well as forecasting climate-related disasters so that we can act accordingly. In order to determine and deal with people’s numerous vulnerabilities to climate disasters, we have developed sophisticated means of forecasting and using geospatial population data in order to allow for appropriate warnings and action. For one, we use call detail records (CDR’s) in order to generate a spatially-accurate map of population dynamics (2). Additionally, in order to accurately forecast disasters such as floods and droughts, we have combined AI with Earth System Model (ESM) data (4), as well as used satellites (GRACE) to directly measure changes in mass distributions (7). Despite our understanding and technical prowess, we have not perfected our ability to evaluate vulnerability and make policies accordingly. For one, due to the high amount of complexity in climate science, it is near-impossible to perfectly determine the effects of climate change (5). Additionally, only recently have we begun to understand the importance of considering the role of sex and gender in climate change vulnerability (6). Luckily however, a recent study has shown that gender-disaggregated CDR’s are a feasible data-gathering method. Therefore, they can be used to evaluate gendered vulnerability to climate change, giving us a better idea of the needs that will need to be fulfilled. By ensuring that people receive the specific aid that they need, we can help reduce gendered climate change vulnerability, promoting greater gender equality.

CDR’s have proven to be an effective means of evaluating population dynamics. When a cell-phone user makes a phone call, they are localized to a Base Transceiver Station (BTS) (2). CDR’s include information on the BTS location of a caller, the time and duration of the call, and of importance for this study, the gender of the caller (3). When callers are localized to different BTS’s over time, we get a rough idea of their trajectories (Figure 1) (2). Because of the spatial and temporal elements of CDR’s, when the CDR’s of many callers are aggregated, they provide an accurate map of population dynamics and offer insight about human activities (2). Rahul Goel et al. in their Plos One article, “Understanding gender segregation through Call Data Records: An Estonian case study” disaggregated the CDR’s by gender in order to measure the dynamics of men and women separately (3). As mentioned previously, this can be applied in the study of climate change vulnerability in order to reduce disproportionality by gender. By knowing who is located in areas forecasted to be hit hard by climate disasters, appropriate warnings can be issued in order to bolster adaptive capacities and minimize harm. Ultimately, CDR’s are considered secondary data because the data are generated by the cell companies only to be used later by researchers.

image (2)

Figure 1: Based on the call patterns of individuals over time and the BTS’s they are localized to, we can get a rough sense of their trajectory (2). Figure A also showcases a Voronoi Tessellation, used in order to localize all points in space to the nearest cellular antenna (2).

As with most disciplines in the modern world, ML and AI have found their way into climate science and disaster forecasting. ML can find correlations in big ESM data, which allow for forecasts of disasters to be made. Using this information, AI can issue timely automated warnings, improving societal adaptive capacities (4). The ESM data itself is the aggregation of meteorological data, specifically the relationships between climate and geological cycles (4). The outputs of numerous different ESM’s have been aggregated into a single database called Coupled Model Intercomparison Project Phase 5 (CMIP5), which has been invaluable for modeling the complex climate (4). Unfortunately however, CMIP5 has had issues due to discrepancies between the different ESM’s. ML and AI, thanks to their analytical prowess, have helped mend this issue (4).

Notable sources of ESM data include the GRACE and GRACE Follow-on (GRACE-Fo) satellite programs (7). The original GRACE program lasted from 2002 to 2017. Thanks to its success, the GRACE-Fo program was started and two new and improved satellites were launched and are currently in operation (7). The objective of the GRACE program was to directly measure changes in the hydrological mass distribution (7). In order to do so, GRACE consisted of two satellites orbiting in tandem with a constant distance between them (7). Large variations in the terrestrial mass distribution caused by large water sources like glaciers, terrestrial bodies of water, and large reservoirs cause slight perturbations in the Earth’s gravity field (7). When the GRACE satellites traveled through the warped gravitational field, the distance between them changed slightly (7). The distance could be precisely determined and by using physics calculations, the hydrological mass distribution was able to be quantified (7). Changes in the hydrological mass distribution were determined when the GRACE satellites flew over the same area multiple times and experienced different perturbations during each flyover (7). These changes in the mass distribution directly influence the climate, and can therefore be used to forecast disasters well in advance (7). For example, the authors mention how with the Mississippi River, GRACE’s measurements of wetness allowed for a flood warning six weeks prior to the river actually flooding in 2011 (7).

In order for CDR’s to be useful for mapping the movement of people, it is necessary for every location to be associated with a specific antenna. Csáji et al. in their Physica A: Statistical Mechanics and its Applications article, “Exploring the Mobility of Mobile Phone Users” accomplished this using a Voronoi Tessellation (Figure 1A), which divides space into specific areas based on the closest antenna to each point in space (1). The adjacent areas around the most frequently used antennae were grouped together using Delaunay Neighborship in order to designate “frequent locations” (1). In order to work out overall calling patterns of individuals, k-means clustering was used with k=3, which revealed a clear spatio-temporal pattern of caller activity and dynamics (1). This pattern nicely followed the independent statistics from the Portuguese National Institute of Statistics (INE) (Figure 2), validating the use of CDR’s as a means of assessing the overall population dynamics (1).

image (1)

Figure 2: The pattern of call dynamics worked out using k-means clustering with k=3 (1). Figure b shows that they fall nicely with INE data, supporting their validity (1).

The main technique the ESM ML and AI focused on was dimension reduction, which is a technique in linear algebra that squishes information down to a lower-dimensional subspace of the original dataset at the cost of losing some of the data (4), which is useful for seeing how different variables are related to one another. This technique can be applied to differential equations, which are the primary means of mathematically modeling the climate (4). By focusing on a smaller subset of the data, it’s much easier to find connections between different factors in the climate, improving our models and reducing uncertainty (4). By combining AI and ML with the technique of dimension reduction, it is much easier to pick out relationships, make new discoveries, and mitigate uncertainty, improving our understanding of climate science (4).

As stated previously, by detecting changes in the distance between the satellites, the GRACE programs give us an accurate picture of how hydrological mass distributions are changing over time, allowing us to make accurate predictions about future climate disasters (7). Specifically, because the satellites orbit side-by-side, whenever they approach a perturbation in the gravitational field, they experience acceleration at slightly different times (Figure 3) (7). This causes the distance between the satellites and their velocities to change in proportion to the gravitational potential (U) (Figure 3) (7). This allows us to quantify the distribution of mass. The change in the distance is precisely determined with a dual-frequency K-band inter-satellite ranging system (7). Variations in the distance resulting from non-gravitational causes are accounted for using an accelerometer (7). When the GRACE satellites travel over the same area multiple times, the rate of mass change can be determined (7).

image (7)

Figure 3: This graphic shows how anomalies in the mass distribution result in velocity differences, causing the distance between the satellites to change depending on how much mass there is. (7)

When Csáji et al. regressed their CDR-based estimations of the positions of frequent locations with independent data from the INE, the correlation was 0.92 (1). This means that the CDR-based measurements very closely followed those generated from independent statistics, validating CDR’s as a spatio-temporal measurement of overall population dynamics. Therefore, it follows that using gender-disaggregated CDR’s are an accurate measure of the overall dynamics of the different genders as well as gender distributions, allowing for their use in vulnerability assessments.

Tapley et al. mention how GRACE calculated the average rate of mass loss in Greenland to be -258 ± 26 gigatons per year during the years of its operation, with 26 representing two standard deviations (Figure 4) (7). The measurements of Antarctica show much greater uncertainty (Figure 4), however there is clearly still a trend of ice mass decreasing over time. Because we have the ability to measure changes in the mass distribution with decent precision, AI and ML can draw much more clear connections and make better predictions, improving our forecasts. With our improved forecasts, we can more readily know the areas that will suffer from climate disaster and use CDR’s in order to know who to warn.

image (7)

Figure 4: Measurements recorded by GRACE show that the ice mass in Greenland and Antarctica have decreased over time, which directly relates to sea levels rising. (7)

Despite the fact that we have the ability to make disaster forecasts far in advance and warn people in a timely manner using CDR’s and AI, the CDR’s used for this purpose were never gender-disaggregated. By not disaggregating the CDR’s by gender, Dujardin et al. failed to provide insight into how the dynamics of different genders could reveal and contribute to disproportionate vulnerability (2). Although gender-disaggregated CDR’s have been used in studying segregation, it is clearly possible for them to be used for finding gender disparities in vulnerability assessments. By evaluating the gender dynamics and gender distributions in vulnerable areas, we can more effectively prevent unequal suffering and allocate appropriate aid, improving the well-being and equality of all genders. I intend to apply gender-disaggregated CDR’s to climate disaster vulnerability assessments in order to examine the extent to which these benefits can be realized.

References

  1. Csáji, B. C., Browet, A., Traag, V. A., Delvenne, J.-C., Huens, E., Van Dooren, P., Smoreda, Z., & Blondel, V. D. (2013). Exploring the mobility of mobile phone users. Physica A: Statistical Mechanics and Its Applications, 392(6), 1459–1473. https://doi.org/10.1016/j.physa.2012.11.040
  2. Dujardin, S., Jacques, D., Steele, J., & Linard, C. (2020). Mobile Phone Data for Urban Climate Change Adaptation: Reviewing Applications, Opportunities and Key Challenges. Sustainability, 12(4), 1501. https://doi.org/10.3390/su12041501
  3. Goel, R., Sharma, R., & Aasa, A. (2021). Understanding gender segregation through Call Data Records: An estonian case study. PLOS ONE, 16(3). https://doi.org/10.1371/journal.pone.0248212
  4. Huntingford, C., Jeffers, E. S., Bonsall, M. B., Christensen, H. M., Lees, T., & Yang, H. (2019). Machine Learning and Artificial Intelligence to aid climate change research and Preparedness. Environmental Research Letters, 14(12), 124007. https://doi.org/10.1088/1748-9326/ab4e55
  5. Pugatch, T. (2019). Tropical Storms and Mortality Under Climate Change. World Development, 117, 172–182. https://doi.org/10.1016/j.worlddev.2019.01.009
  6. Rautio, A., Kukarenko, N., Nilsson, L. M., & Evengard, B. (2021). Climate change in the Arctic—the need for a broader gender perspective in data collection. International Journal of Environmental Research and Public Health, 18(2), 628. https://doi.org/10.3390/ijerph18020628
  7. Tapley, B. D., Watkins, M. M., Flechtner, F., Reigber, C., Bettadpur, S., Rodell, M., Sasgen, I., Famiglietti, J. S., Landerer, F. W., Chambers, D. P., Reager, J. T., Gardner, A. S., Save, H., Ivins, E. R., Swenson, S. C., Boening, C., Dahle, C., Wiese, D. N., Dobslaw, H., … Velicogna, I. (2019). Contributions of GRACE to understanding climate change. Nature Climate Change, 9(5), 358–369. https://doi.org/10.1038/s41558-019-0456-2