Summer Week 5: TESSERA brambles; (searching for) resting habitats; obstacles & ABM alternatives; temporal co-variates
Categories:EnvironmentAI/ML🦔
This week
The past five days, I was pretty locked in on the hedgehogs project. I started the week by making my ABM model a bit more complex by adding a mechanism that searches for “ideal” sleeping habitats as a hedgehog agent’s day approaches an end. Furthermore, I finally got to use TESSERA for bramble inference; however, I obtained mixed results. Then, I opened up two more directions: 1) after meeting with Dr. Petrovan, I made my return to battling convergence issues (amongst countless other kinds of issues) in my iSSA script (in R, using the amt library); and 2) actually using insurpassable barriers (walls, buildings, water surfaces - for now, despite hedgehogs alleged ability to swim) with OSM data, and switching from an approach that samples steps using appropriate step lengths and turning angles to one that is more considerate of the barriers (although I cannot say that’s a huge issue in the currently considered rural landscape).
Sleep habitat navigation
Using a mixed-effects logistic regression model, I calculated the coefficients for used sleeping habitats (for both sexes separately, of course).
I then used these coefficients with increasing weight whenever a hedgehog agent started approaching the end of its day/cycle. In the current implementation, whenever a hedgehog has 5 or fewer steps left in them, they considers the sampled average of the area extended by the number of steps left for each step.
I think overall this has worked quite well, the only problem is the occasional random behavior of some hedgehogs. This is particularly well visible in the step lengths, where my statistical models don’t capture the outliers too well, and the synthetically generated step lengths are then concentrated around the average values.
TESSERA x brambles
Another thing I did this week was using the foundational satellite imagery model TESSERA for inferring where brambles are. In short: I downloaded raw brambles locations from iNaturalist, and then filtered as much as I could to obtain as reliable and accurate records (for instance, only using ‘quality_grade’=’research’, assuming that should be generally superior to other records). I then decided to find the bounding box with the highest concentration of these filtered brambles, which happened to be somewhere vaguely near Blackpool, England. Afterwards, I spent a while trying to get the best classification model (bramble - not bramble) using TESSERA. To achieve the best results, I eventually went through all of the bramble records in the area (a little less than 400), and filtered it down to ~180 - as expected, the key issue was that brambles are so often located under trees and branches. For the negative examples, I experimented with: 1) manually selecting negative examples, trying to pick interesting and diverse cases (other greenery, built environments) - which happened to produce generally bad results; 2) randomly sampling points that were far enough from the known locations of brambles. With this approach, with sets between 1000 and 2000 negative examples, I received the best results.
However, it is obvious that most of the generated findings aren’t brambles. Despite that, I think this model is good at finding bramble-like bushes (which is actually very useful), but I doubt it’s great for filtering only brambles. It will be interesting to see how it compares to the hedgerows from the hedgerows/stonewall/woodlands dataset, once that is fully released (I was told that should be eow a few days ago).
Future work: time co-variates and different ABM approach
Next, I should try to figure out how to finally deal with the time co-variates (and perhaps be done with R once and for all), and switch the ABM model approach. I started working on a kind of wavefront expansion (first generating a restrained cloud of accessible locations), that would (I hope) work well with future applications (e.g., evaluating how making changes to accessibility could affect expansion).