Fortress America: How isolated are U.S. embassies in Africa?
ChatGPT’s Code Interpreter, Python, and QGIS help confirm a personal observation.
One in a series of projects highlighting my progress as a self-taught programmer.
I began teaching myself Python, SQL, and GIS in spring 2023 — starting from zero. I therefore welcome feedback on these projects and review for errors. And I’d be interested in taking a crack at your own data and geospatial questions too. Please get in touch.
28 July 2023
Background
U.S. embassy, Freetown, Sierra Leone | Matthew Muspratt
During the many years I lived in sub-Saharan Africa, a running joke—or commentary—held that U.S. embassies resembled fortresses. While the diplomatic posts of other nations might sit in the heart of a capital city’s CBD, the U.S. embassy would be situated on the outskirts, set behind high blast-proof walls and a deep perimeter. The 1998 embassy attacks in Tanzania and Kenya—and 9/11—explain the intense security, but I never thought it a good image, and the foreign policy crowd is well aware the architecture hardly projects freedom and opportunity, or wins hearts and minds.
While the standoff-ish, imposing effect might have been obvious, I always wanted to quantify and visualize “fortress America.” With Google’s Open Buildings and openrouteservice’s (ORS) isochrone tools, I now can.
Project outputs
Here is a Google Maps satellite view of Freetown, Sierra Leone, noting the whereabouts of the American, German, Ghanaian, British, and Indian diplomatic posts:
And here is a map I produced in QGIS that features building footprints (or, really, roof-prints, from Open Buildings) of all the buildings across the city as well as 15-minute-walking “reachability areas” (catchment polygons) from the five aforementioned embassy compounds:
Sure enough, as indicated in the table, while upwards of roughly 50-60,000 people live within a 15-minute walk of the German, Ghanaian, British, and Indian diplomatic posts, less than a third as many—only 15,744—live that close to the U.S. embassy. Likewise, there are significantly fewer buildings in the U.S. embassy neighborhood.
The population figures, I should note, are automatically generated by ORS’s QGIS isochrone plug-in, which pulls data from the European Commission’s Global Human Settlement Layer. They seem reasonable to me, given my knowledge of Freetown neighborhoods, but I in no way dug into the process by which the tool tallies populations within isochrones.
As for the buildings, I performed a polygons-within-polygon count with QGIS’s Join Attributes by Location (Summary) tool, a fairly heavy lift for my computer given I was working with a 870 MB Open Buildings file that defined 3.5 million buildings (i.e. polygons). I vector-clipped down to the 470,000 buildings within the Western Area region of Sierra Leone, which still proved a slog to process and render, but it was neat to see how well Google’s deep learning model identified rooftops. Here’s the Open Buildings layer on top of Google Maps:
Below, I outline some of the Python and GIS techniques used to convert data into the above maps—including a cross-check of work performed exclusively by ChatGPT’s Code Interpreter—but first let’s look at diplomatic posts in another West African capital, Accra, Ghana:
Once more, the U.S. embassy, though centrally located in Accra, finds itself with the least populated 15-minute-walk “radius.” And, again, no other embassy neighborhood contains fewer buildings.
I executed this project with a touch of smugness. Across Africa I was proud to live without the high-security show, proud that my workplaces did not feel compelled to keep neighbors at bay with tank- and riot-proof ramparts. In Accra, in fact, anyone could have walked right up to our wooden front door and knocked. But then I checked the isochrone:
Compared to my neighborhood of Kanda, the leafy streets of Cantonments, where the U.S. embassy is situated, appears half as dense in terms of buildings, but the population is apparently no different! A visual check explains the numbers: My home backed up to a highway dividing middle class Kanda from the much poorer and very dense Nima neighborhood; in the other direction, a 15-minute walk would get me to the vast, barely occupied grounds of Jubilee House and the Presidential Villa:
I may not have hunkered down in a fortress, but I lived right at the node between density and space.
Techniques
As alluded to, much of my work in QGIS involved the kinds of data joins, polygon dissolves, clips, and counts, third-party isochrone plugins, and rule- and expression-based labeling and filtering that I’ve experimented with elsewhere. The large building polygon datasets, multiple geographies, and multiple isochrone calculations—along with my focus on specific buildings within the data—made me appreciate the constant attention to layer management (and underlying file management) required even for projects as limited in scope as this one.
Still, I found the initial data preparation especially interesting. For starters, using Nicolas Raoul’s dataset of embassies and consulates, I tried preparing the csv containing embassy location data three ways:
I talked ChatGPT’s Code Interpreter through all the filtering I required, uploading the original csv to our chat, never writing any code myself, and asking ChatGPT to export the revised csv file. It worked incredibly well, and I was especially surprised how easy it was to ask Code Interpreter to investigate questions about the data. Here are snippets of the conversation:
Code Interpreter went on to successfully export (via a download link) a csv containing only embassies in Africa and the minimum columns that interested me (which, I have found, makes future data joins in QGIS easier to check in the attribute table).
2. With the csv export a success, I wanted to cross-check Code Interpreter’s work… by asking Code Interpreter to translate our conversation to Python:
This, as expected, perfectly delivered the same results—though it revealed ChatGPT’s list of African countries would fail to filter for “Democratic Republic of the Congo,” “Republic of the Congo,” and “The Gambia.” I asked my assistant to simply include any country with “Congo” and “Gambia,” re-exported my csv and Python code, and went on to my third data processing angle.
3. Fail. I asked Code Interpreter to edit the above Python script for execution in QGIS’s Python Console and Editor. First, it was reluctant to perform some of the filtering it had done with the above methods; then it struggled with the latitude and longitude columns; and after multiple prompting and re-prompting fails, I decided I had the csv I needed and moved into QGIS.
A final technical note. When working with the Open Buildings file that contained rooftop data from Ghana, I encountered a problem: It was 3.7 GB, defining millions of polygons, far too big for my computer to handle in QGIS, where I planned to clip the data with a polygon spanning central Accra. Unable to even load the file, let alone perform vector-clipping, I turned to Python and the pandas, geopandas, and shapely libraries. My script ably clipped the Open Buildings file to my Accra boundary polygon and exported a manageably-sized GeoJSON file.
Data sources
In addition to the GBs of building footprints from Google’s Open Buildings and Nicolas Raoul’s dataset of embassies and consulates, I downloaded boundary files for African countries from geoBoundaries. A QGIS plugin pulled Google Maps satellite imagery into my project.
