Using Facebook's ML derived settlement data for tasking my mapping activity in Banton island, Philippines

Posted by maning on 20 June 2019 in English.

A couple of weeks back, I was aimlessly panning around the Philippines in the OSM.org website and saw Banton island right at the center of the archipelago north of the Sibuyan Sea. I loaded the data in JOSM to check what was actually mapped in the island. It looks like it has decent coverage for roads and very few buildings. So I started adding a few buildings here and there and then it struck me, what would it take to systematically map the whole island remotely?

Banton Island in Romblon as viewed from a pump boat arriving from Marinduque. Photo credit: Lawrence Ruiz, Wikimedia Commons

This quest has begun …

Choosing the imagery

First, I checked if there are good imagery available other than Bing, I found that DigitalGlobe (DG) Premium and Standard looks more recent compared to Bing. Some areas are cloudy in Premium while clear in Standard and vice versa. I figured I can interchange these two depending on where I am mapping.

I then checked if there are GPS traces in OSM I can use to align DG imagery in case it has an offset. I found no traces in OSM, not even a sigle dot. Strava has no coverage too. I guess I will have to use the default referencing from DG. I confirmed if there are any large offset by comparing to Bing. Other than a few meters, I did not find large offsets.

Using Facebook’s ML derived settlement data in Maproulette

The island is mostly covered by thick vegetation of forest, scrub and coconut orchards. Just panning around the island trying to find a small hut underneath the foilage is too laborious. So I thought of using Facebook’s High Resolution Settlement Layer (HRSL) as a guide to identify the location of settlements. According to HRSL’s metadata, the settlement raster is a “30m pixel where buildings where detected with imagery”.

I converted the HRSL raster to vector polygons. This layer gave me rough indication where to map or at least pan around the surrounding area of the grid. I then created a Maproulette challenge using the extracted HRSL polygon. With Maproulette, I can easily track which parts of the island I have not visited/traced buildings.

HRSL grids showing porbably location of settlements.

It took me about a week of casual mapping to finish the whole challenge. In additon to buildings, I also added and aligned roads, updated the coastline and moved the village nodes closer to the center of the settlements. In areas where HRSL was absent but settlements exist in the imagery I buildings as well.

Outcomes

Improved data coverage

Here’s the before and after data for buildings (cyan).

Before	After

highway (km) - 41.7 building (n) - 82	highway (km) - 53.9 building (n) - 1,862

With a few hours of mapping for a week, I was able to improve road coverage by 1.2x and buildings by 22x.

Tasking efficiency

Using FB’s HRSL data as a rough guide for tasking prioritization definitely helped me organize the work. The whole island covers ~31,000 30x30m grids. Using HRSL narrowed it down to only 535 grids (1.7% of the whole island).

Comparison of 30x30m grid where HRSL data detected settlements and buildings mapped in OSM.

Category	N (%)	Color
HRSL grid with OSM building traced	361 (~24.4%)	blue
Non-HRSL grid with OSM building traced	944 (~63.8%)	orange
HRSL grid without OSM building traced	174 (~11.8%)	red
Total	1,479

For large scale mapping effort especially during crisis response, HRSL can be a good proxy to direct mappers areas to focus to get maximum mapping coverage. However, there are areas in the island where I traced buildings but were not detected by FB’s algorithm (~63.8%). Possible reasons are:

imagery used for detection is outdated (2012-05-08) compared to DG’s imagery I used for tracing, unfortunately, I cannot confirm the actual date of DG’s imagery in JOSM;
depending on the date of capture, many isolated houses are obscured by the thick vegetation and thus not detected by their algorithm, I experienced the same problem when swtiching from DG premium and standard.

Here’s a few zoom-in images for detailed comparison.

Barangay Poblacion. The largest concentration of settlements in the island, the blue grids confirms presence of settlements in HRSL and presence of buildings in OSM.

Barangay Tongonan. Only a few grids of settlement were detected by HRSL within the surrounding areas of Tongonan Elementary School. Most of the isolated buildings were mapped from DG imagery (yellow grids).

Barangay Yabawon. The false positive (red grids) along the coast from HRSL were mostly high reflectance surfaces such as bare rock, beaches and sea waves which are similar to roofs of houses

In summary, FB’s HRSL data is a great resource for scoping initial areas to conduct mapping in the absence of ground truth information. In most cases, it correctly identified contiguous areas of settlement. During HOT crisis activation, for example, information from the ground is scarse during the onset of the event, we can use HRSL to define initial areas of interest (AOI) for coordinating mapping. However, it does not guarantee completeness (at least in the island I used for this test) because depending on the imagery (recency and quality), isolated settlements maybe missed by the detection.

Is it complete? Definitely not, but that is as far as what I can do given the the current sources available. Of course this was done remotely, I hope this is a good starting point for future mappers who will visit this island.

Finally, there are 11 mappers who mapped this island before this quest. Thank you to droki, JAT86, burts, PelleB, Shaun Austin, Evelyn Desouzaa, yawor_89, LordOfMaps, 5EE, behappy24, ed_waypointsdotph!