## What are you working on: "visual Iced"
- CNN trained on satellite imagery
- specifically a CNN "visual Iced", input are images 40km2, in high-res increments of 250m
- CNN classifies sea ice & polar features
- CNNs are great at learning scale invariant features
- uses concurrent multispectral satellite imagery in addition to radar images to overcome challenges like cloud cover

## What is current output/goal?
- Visual Iced feeds into SDA vessel
- Martin visited vessel and talked to crew to better understand how they will be using the image data :)
- SDA vessel
- the route planning algorithm optimizes for fuel consumption and time usage. It requires various "base layers" (eg sea ice concentration, currents, wind vectors) to inform its route
- for sesa ice vessel currently relies on coarse radar images (25km resolution), but a new image is produced every day.
- this coarse dataset is open data provided by open space agency, NASA (AMSR/osisaf)

- Visual Iced will offer higher res but has the limitation that images are not produced every day (but could be produced when prompted), so the aim is not to replace the original base layer but add a new one when needed. eg vessel mostly relies on coarse predictions, but when traversing ice-infested waters, they could switch to high-res.
- the DT feedback loop of the vessel feeds back onto the route planning, not necessarily into the CNN.

**DT feature relevant for this project**: DT infrasturcture is set up to be agnostic with regards to where data comes from, as long as it provides the needed base layers

**Possibly relevant for TRICDT** Martin's CNN is a nice case study for how a complete ingress work flow already could look like and what the challenges are.

## Current state of the project:
- They are assessing how the route recommendations are changing based on the CNN input and how that differs from the coarser input data.
- Challenges: With coarser data, route was adjusted every 25km, but now it is adjusted every 250m. This is not desirable, we want the boat to find a smooth trajectory around an ice sheet, so in that sense while the DT infrastructure is agnostic, the planning algorithm (most crucial decision making tool of DT) is not truly agnostic.

### Future plans:
- The coarse resolution dataset has some additional issues, it currently under/overestimating sea ice
- Martin has put in proposal to validate the current algorithms producing this data. Proposal was put in January 2023, he is currently getting people on board who have done the pre-work, if it would be funded this would start beginning 2024
- This coarse datset is the same one that is currently used for sea ice forecasting using IceNet
- Andrew McDonald will be working on work on turning the current IceNet forecasting into more real-time forecasting (aim is a scale of 10 days)

## What are your stakeholders?
- SDA vessel (operational output, from Ms perspective somewhat secondary but cool)
- Martin is also very interested in the environmental perspective:
- Envrionmental NGOs working on biodiversity and wildlife habitat monitoring might be stakeholders for ice forecasting (eg the caribou project)
- But he also mentions some challenges in actually getting those collaborations going. Many are interested but it rarely ends up being used.
- why?
- they don't have time to scope out other datasets
- lack of trust or understanding of machine learning
- who are they even? There might be more potential stakeholders but hard to identify them **Can we help??**
- example Caribou: WWF canada, indigenous communities, warning system re caribou crossing
- but for Caribou project a problem is that sea ice cannot be detected from images due to darkness during winter, so here M's proposal for working with the coarser dataset would come in.