Predictions ahead of time

[Peter9192]

Currently we're predicting "the day of year that event X happened", given all the data (temperature, satellite, etc.) covering the entire growing season. This means you can only predict retrospectively.

Ideally, we'd also like to make forecasts, i.e. predictions ahead of time, as well. So at some point during the growing season, we want to predict, given the latest information available, when the event might occur (or perhaps it has occurred already).

To achieve this, we would like to make the following two changes to the training data:

1. Predicting "Days until/since event happened" instead of predicting the absolute values of the Day Of Year of the event.
2. Duplicate the training data several times, every time subtracting a different offset from the DOY. Then, for each offset, extract meaningful features that can be valid at any time during the growing season. For example, cumulative temperature since start of growing season, temperature during the past 10, 20 and/or 30 days, number of days temperature exceeded 15 degrees, etc.

[Peter9192]

@mkhzadeh it would be helpful to link to your example notebook here

[Peter9192]

The transposed columns are a bit in the way of implementing this neatly. Perhaps we could

• Use xarray for multidimensional data until the very last step when we convert it to dataframe
• Use daft.io to store timeseries in a single cell in dataframes
• Use a multiindex??

[Peter9192]

I created an example notebook to illustrate the main idea/steps in #185 @sverhoeven I think this could be relevant for the finalization/testing of the data load/transform/combine methods, we could discuss it some time.

[khzadeh]

Please consider "Duplication of the training data" on hold.