[algorithm discussion] beware! it is a temporal-spatial input problem

At first, I thought this is a simple, image segmentation problem. But on a closer look, not all satellite images contain useful information.

Hence you need to consider temporal-spatial input. My model is feature extraction from the segmentation model (e.g. unet) + transformer for fusing features for 12 months and predicting a single yearly agbm. From my experiments, i have some interesting observations:

Selection-033
Selection-032

2 Likes

here is another trick. a printout of agbm values show that it is actually “discrete” (similar values).
e.g. in chip_id = ‘01bc24f4’ below

 291.44000244140625,
 291.44000244140625,
 291.44000244140625,
 291.44000244140625,
 291.3500061035156,
 291.3500061035156,
 291.3500061035156,
 291.3500061035156,
 291.17999267578125,
 291.17999267578125,
 291.05999755859375,
 291.05999755859375,
 291.05999755859375,

further,

Selection-038

1 Like

I’m really excited about this competition - thank you for adding your data insights here :slight_smile:

Very cool approach @hengcherkeng, it seems to me finding a clever way to use the temporal information might be key here.

What do you pass to the transformer? The output of the entire UNet, or only an embedding at some point?

you should think of transformers as a switch to select what information to pass through.

all cases are possible, but with different performance

predict(x,y) =  some_function ( select best ... for input (month,c,x,y) for each ... )
 - you can select the best month or best (month,c) 
 - over each image or over each patch(x,y) or over each patch(c,x,y)


1 Like