[ad_1]
A new map of almost all of Africa shows exactly where the continent's 1.3 billion people live, which could help everyone from local governments to aid organizations. The map joins others like Facebook created by executing satellite images through a machine learning model.
It's not exactly the mystery of where people live, but the degree of accuracy matters. You may know that one million people live in a given area and that almost half are in the largest city and another quarter in several cities. But this leaves only hundreds of thousands of people explained in the most vague way.
Fortunately, you can still look at the satellite imagery and spot the locations of small villages and isolated houses and communities. The only problem is that Africa is big. Really big. Manually tagging satellite images, even of a single medium-sized country like Gabon or Malawi, would require a lot of time and effort. And for many data applications, such as coordinating the response to a natural disaster or distributing vaccines, the time lost is life lost.
Better to do it all at the same time, right? This is the idea behind Facebook's population density map project, which had already mapped several countries in the last two years before the decision was made to extend to the entire continent. African.
Zoom in and you can see the difference between new and old maps. It's quite significant.
"Facebook's cards ensure that we focus the time and resources of our volunteers on where they are most needed, thereby improving the effectiveness of our programs," said Tyler Radford, Executive Director of Humanitarian OpenStreetMap Team, a project partners.
The central idea is simple: combine census data (number of people living in a region) with structural data derived from satellite imagery to get a better idea of the situation of these people.
"With the only census data, the best you can do is badume that the population lives everywhere in the district – buildings, fields and forests," said Facebook engineer James Gill. "But once you know the location of the buildings, you can ignore the fields and the forests and only affect the population to the buildings. This gives you very detailed population maps of 30 meters by 30 meters. "
This is many times more accurate than any existing population map of this size. The badysis is performed by an automated learning agent trained from OpenStreetMap data from around the world, where people have labeled and described buildings and other features.
First of all, it was necessary to remove the enormous surface of Africa devoid of any structure, which made it possible to reduce the space which it had to evaluate by a factor of a thousand or more. Then, using an algorithm specific to the region (because the situation is very different on the Moroccan coast compared to the center of Chad), the model identifies the patches containing a building.
The map data, two images at the top left, are processed to find buildings, two at the bottom left; in the end, large tracts of land can be labeled as populated or not, as can be seen on the right.
Throughout this process, humans perform numerous checks to ensure there is no bias or regional tendency to erroneously confuse. The team has been doing this for a while, so it's not their first rodeo, but the "one country" versus "all of Africa" scale is a bit different. Fortunately, the artificial intelligence of society wrote in an explanatory blog:
We were able to simplify the problem into a simple binary clbadification task … Now, with an input image, a single neural network predicts whether the given image contains a building. This clbadification approach is also significantly less expensive in computing than a segmentation-based approach, as it allows us to use smaller neural networks and produce outputs with a smaller memory footprint.
Greater efficiency, in this case, also comes with greater precision, since the algorithms will have learned from their previous attempts and more data is included to avoid false positives and negatives. The team found that out of 1,000 plots labeled as containing buildings, 996 were correct. This kind of error rate seems to me quite acceptable and is certainly better than the existing tools, which only gave you a vague "out there somewhere" when you ask questions about a small community or an off-grid village.
If you're wondering why Facebook does it first, it's about the efforts that have been made in recent years to identify people with poor connectivity so that they can then send them to the Internet through lasers or whatever. It's a pretty low priority right now, with the many problems in society right now, but the tools it was building obviously had humanitarian applications and it's nice to see that the baby has not been thrown with bath water.
Source link
