4
$\begingroup$

I'm trying to find a way to predict/calculate how a shape (e.g. outline of a glacier) will change in the future—based on its history (previous shape) and additional factors (e.g. Δtemperature).

In my example: I have the shape/coordinates of a glacier and an average temperature at 1970, 1985, 2000, 2015. How can I give an estimate on how that shape will look like in 2030, based on the previous shapes and a predicted temperature?

The shapes would ideally come in a format similar to this:

[
  [
    [-113.74926783355818, 48.32440832757374],
    [-113.74926767330584, 48.32440813255405],
    [-113.74926748239692, 48.32440458296232],
    [-113.74926717549286, 48.32439887665663],
    [-113.74926686858888, 48.32439317035087],
    ...
  ],
  [
    [-113.75763099733634, 48.32877681033601],
    [-113.75763955455557, 48.32877660502744],
    [-113.75764546550784, 48.32878027200765],
    [-113.75765402272751, 48.32878006669797],
    [-113.75766003497893, 48.32878563609638],
    ...
  ]
]

But I imagine it could be necessary to convert this to coordinate offsets (Δx, Δy) similar to what sketch_rnn is doing.

Any pointers would help. Even that this is too ambitious. Thanks!


Disclaimer: I'm a beginner with this and hope the question isn't too naïve :)

$\endgroup$
  • $\begingroup$ If you could partition the area into points, where each point either appears of disappear, using sequential models to encode the evolution overtime $\endgroup$ – Fadi Bakoura May 23 '18 at 18:24
  • 2
    $\begingroup$ This a fabulous but ambitious question. I would start by thinking about shape representation, then learn the warp or deformation. Take a look at DeepWarp and Deformation Based Curved Shape Representation (code). Let the date be a feature. The only problem is that you need more than four data points, so go and get some. At least get pictures of other glaciers so your model can learn the manifold of glaciers if not their temporal dynamics. $\endgroup$ – Emre May 23 '18 at 22:19
0
$\begingroup$

Given that you only four data points (i.e., 970, 1985, 2000, 2015), there is no way to apply machine learning to make an accurate prediction.

If you get more data points, you can start with a sign prediction of surface area. You predict if the shape is increasing or decreasing in size.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.