Time series forecasting

Hello,

we are four students who want to develop a Grasshopper-based tool to support architects and urban planners in their decision making.

Our idea for this is to forecast urban growth using classified images (in the form of a time series). At first, the idea was to take time series of satellite imagery and label them accordingly into urban and non-urban. Since this step alone seemed too time consuming for now, we took a different approach. We photographically documented the growth of the slime mold Physarum Polycephalum in a Petri dish (about 2000 images in 2-min increments) and then labeled them into fungal and non-fungal.

We would now like to use this labeled data (in the form of a list of black levels for each pixel) to make a prediction using Crow.
Our first consideration is to use as input the pixel values of time step n and as output those of time step n+X and to train the network in this way (X could be 1, i.e. 2 min in the future, or 10, i.e. 20 min in the future). We would then use a later time step as a test, which is not included in the training set. So we could then compare the real output with the generated one.

Now the questions:

How realistic is it to perform such a prediction using the backpropagation approach in Crow? Especially also with regard to the structure of input and output (both times long lists of pixel values)? Will this detect the patterns of the temporal component, even if they are not explicitly stored in the inputs and outputs?

Or in general, is it realistic to do this kind of time prediction with Crow?

Thank you to anyone in advance who might be able to help!