Summer Week 4: Walkability evaluation and brambles
Categories:EnvironmentAI/ML🦔Walkability
Back to walkability
A big chunk of my time this week went back into the walkability project. I reassembled the code I used for my thesis, and connected a pipeline solely dependent on the OSM data (as in the thesis local government datasets are used for greenery and green spaces). Collecting and parsing the OSM data is easy, but the question of the OSM coverage depends on location, which, I think, should be decided based upon the evaluation (as expanded upon below). Moreover, to be able to run some evaluations, I need to find a GPU, as running this on my Air is not conceivable.
In general, I think a better approach would be to focus on the walkability rather than the generated routes (in contrast to the thesis), as the walkability may, after all, be easier to evaluate. For instance, it would be easier to get some evaluation of a singular point than an entire route.
One exception (to some degree), is the possibility of using GPS traces, which has been done before to “evaluate” walkability, for instance by measuring the deviation from the shortest path (e.g. Miranda et. al. 2020). However, I think there are two problems with this. First, the GPS traces are difficult to obtain. Traces used in the example paper are obtained from a non-disclosed provided who collected them from personal mobile devices, and, already at the time the paper was published, were five years old. Second, I don’t think the deviation from the shortest path approach is great, as assuming that taking a longer path means the path is more walkable is not always the case (think morning, business, and other commutes where walkability may not be an important factor). Furthermore, I don’t the codebase implemented by the mentioned paper or the calculated walkability indices were released.
Another option is relying on point-wise environmental surveys. This can be easily established with street-map imagery, which can be fetched effectively from services like Mapillary. Furthermore, for human crowd-sourcing, there exist ready-to-use survey frameworks, that take care of both the front- and back-end. One such framework is introduced in Danish et. al., 2024. In this case, the researchers also collected their own survey data in Amsterdam (over 22k ratings with over 19k images), but those aren’t publicly accessible at the moment. I sent them an email but haven’t heard back yet, as it seems 4 out of the 5 authors are currently on a vacation… In any case, I think these could be extremely interesting for the walkability framework evaluation.
An alternative to human survey may be using LLMs. Both now and when I used LLMs to label points in my thesis, I found the results satisfying. Moreover, if I was able to run some larger model somewhere, I could use the street-view imagery as well. However, I’m not entirely sure if using LLMs as judges is up to the standard.
Hedgehogs
On top of the walkability, I also did some more work on the hedgehogs. I did some tunning to both the step selection analysis model (which seems to be a never-ending process) and the agent-based model (particularly some better use of the turning angle coefficients).
Also, I have been exploring using TESSERA for the bramble localization problem. I discovered iNaturalist, got all sightings of brambles in the UK, and then did some filtering (confidence, sources) and clustering. From that, I found a 10km x 10km square with the highest number of brambles, and we’re planning using that this week to see if TESSERA can work for bramble identification. Having seen a few brambles over the past few days, however, I’m a bit worried about the ability to see brambles in satellite imagery. In Cambridge, for instance, brambles generally seem to be covered by trees.