For instance, it will be much help to know more about the input squence:
1. What are expected range of each bands? How to normalize the raw data?
2. What is the meaning for the 48 observations? The example uses 8 days as the interval. Is that expected? Or given the test data is like in two consecutive years, each observation should be like a half month, so 12*2*2=48?
Normal scaling has already been done. Factor of 0.0001
The model is built on top of PatchTST architecture. It takes in a fixed sequence length, for this model it is 48 . The example dataset was a dummy dataset so no it wasn't 8 days intervals. Let me share a document on thishttps://www.amini.ai/research/self-supervised-representation-learning-on-remote-sensing-pixel-time-series-with-patch-based-masking
Thanks for your reply.
Yes, the test data has been normalized somehow. Each band has different range, like 0 to 1.xxx, 0 to 2.xxx or 0 to 3.xxx. There is only one 0 for each band, which could imply the original min value. But the ranges mean they are not min-max or z-score normalization. Can you share the way of your normalization? Because we may need to process the train data with the same method.
yes, only scaling with a factor of 0.0001 was done on the sensor reading to get these values.