Calculate center of mass of simulation result

[SteffenME]

Currently we only provide a distribution of likely positions of drifting particles at a specific time. When the SAR vessel arrives, the locations have already changed.

We should factor the position and speed of the SAR vessel into the simulation and provide an efficient search pattern.

[julled]

Good idea!

As we have 100 points at the input, which one would we choose to be the one to be intercepted?

[SteffenME]

Filter outlayer and then calculate the center of mass. Its more or less an addon but it could help reduce the overhead of the user that has to deal with 100 points on the map..

[svenseeberg]

Simply setting a course to the center of the mass is not really helping. This can easily be done by extracting the rough center of the coordinates. But this is only scratching the surface. There are much bigger issues with this approach: we would have to know the position of the SAR vehicle and its cruise/max speed. We would then simulate how long it takes for the ship to catch up with the most likely position of the drifting object, while the object is still drifting / changing direction. Otherwise a course is leading the ship to the wrong location.

IMHO for the time being we should limit this project to the drift simulation aspect, not a fully fledged SAR search pattern creation tool ;-)

[svenseeberg]

This is far beyond the MVP: As a center of mass is probably not sufficient to find a person, we should actually calculate a (the most efficient?) search pattern (https://owaysonline.com/iamsar-search-patterns/). We should also get external expertise here.

[SteffenME]

The idea was more to condense the information of 1000 points to a single point, so that its easier to make sense of the Data. The ship can then decide on a search pattern or next steps. I think it will be sufficient if the Dirft Model is accurate, you can spot boats in a 3-4km Radius, that will include most of the possible points in the Output (depending on time and weather conditions). Finding a person in the water is a lot harder and requires a more strategic pattern. You might not be able to see a person from less than 100m.

So for MVP i would actually like to include some sort of center of mass, to condense the information.

[svenseeberg]

So for MVP i would actually like to include some sort of center of mass, to condense the information.

Should be possible as well for an MVP. However, maybe simulation run with 1 particle with a radius of 0 is also a good approach?

*edit: probably not:

[SteffenME]

1 particle and radius zero would be an easy approach but the power of this model is to run many simulations with slightly differnt params in parallel which is required for a more accurate result. So i do not expect the radius 0, 1 particle to be always in the center of mass.

However I will implement such a functionality for finding boats in the satellite images anyways. I will share it as soon as i have it.

[svenseeberg]

By chance I stumbled upon a simulation where the center of mass would be the wrong place to go:

Some of the drifters are being stranded. The result is that right in the middle of the current cloud of drifters there is place where there is a smaller drifter density. Starting a search in this place would be wrong.

We may have to work with density plots, similar to the last image in https://opendrift.github.io/gallery/example_leeway.html#sphx-glr-gallery-example-leeway-py

IMHO we should postpone this issue and not tackle this in the MVP.

[svenseeberg]

It may be useful to limit the scope of this tool as much as we can. Instead of doing further calculations here, we may want to set up a dedicated SAR management tool. For that we should implement features like #50 and #48 .

[julled]

A additional radius / std-deviation @ 2 sigma might also help the user to understand the distribution of the data and usability of the center of mass.

this measure could be also calculated for x/y N/E axis separatly.

[SteffenME]

I am not that convinced that just the point cloud is a good representation, it is not really a quantitative measure. Its impossible to see how many are on top of one another. The density is closer to being quantitative and it comes out of the box.

[SteffenME]

Also the main idea was to condense the information, and ease the use of the data. From the feedback it looked like they do not really want/need that.

[svenseeberg]

I think the k-means clustering described in #84 is the way to go ;-)