Question about the prediction course

Course Q&A TensorFlow Developer Professional Sequences, Time Series and Prediction
,
1

Hello to the whole community!!!

First of all I would like to thank you for the valuable knowledge about this particular course and in general about everything you teach, I find it very interesting.

My name is Mario and I am a PhD student, I am currently working in the environmental and remote sensing sector and I have a question about time series forecasts applied to my sector and I would like to know if someone knows how to answer it.

My question is how to forecast categorical data such as land cover data in which there is always 1 discrete label per class (e.g. forest 0, water 1, urban 3, etc…) based on historical data from satellite images or land cover datasets, assuming that these have some seasonality over time and may be related to demographic variables, climate, etc… What do you think is the best model option for such a problem?

On the one hand, it is interesting to classify the land cover using images of the land cover itself, and on the other hand, the forecast itself for future horizons.

Thank you very much for your time reading this post, I look forward to your valuable comments.

Bests regards

Mario

2

Depending on how you choose to model, you can include historical predictions as part of input features to make a prediction for output timestep(s). One thing to be careful is when including images is to consider the augmentations required to make the model generalize well.

This would depend on the features used and model hyperparameters. Sorry for such a generic response but the model performance have a lot of choices tied to it. Try deep learning specialization (courses 2 and 3) to get more insight into model construction / tuning.

3

Thank you so much for your reply, @balaji.ambresh! I will consider your helpful comments.