Conversion of time series data to images (ts2img) is a general problem that we have to deal with often for all kinds of datasets. Although most of the external datasets we use come in image format most of them are converted into a time series format for analysis. This is fairly straightforward for image stacks but gets more complicated for orbit data.

Orbit data mostly comes as several data values accompanied by latitude, longitude information and must often be resampled to fit on a grid that lends itself for time series analysis. A more or less general solution for this already exists in the img2ts module.

Possible steps involved in the conversion

The steps that a ts2img program might have to perform are:

  1. Read time series in a geographic region - constrained by memory
  2. Aggregate time series in time
    • methods for doing this aggregation can vary for each dataset so the method is best specified by the user
    • resolution in time has to be chosen and is probably also best specified by the user
    • after aggregation every point in time and in space must have a value which can of course be NaN
    • time series might have to be split into separate images during conversion, e.g. ASCAT time series are routinely split into images for ascending and descending satellite overpasses. This means that we can not assume that the output dataset has the same number or names of variables as the input dataset.
  3. Put the now uniform time series of equal length into a 2D array per variable
  4. A resampling step could be performed here but since only a part of the dataset is available edge cases would not be resolved correctly. A better solution would be to develop a good resampling tool which might already exist in pyresample and pytesmo functions that use it.
  5. write this data into a file
    • this can be a netCDF file with dimensions of the grid into which the data is written
    • this could be any other file format, the interface to this format just has to make sure that in the end a consistent image dataset is built out of the parts that are written.


The chosen first solution uses netCDF as an output format. The output will be a stack of images in netCDF format. This format can easily be converted into substacks or single images if that is needed for a certain user or project.

The chosen solution will not do resampling since this is better and easier done using the whole converted dataset. This also means that if the input dataset is e.g. a dataset defined over land only then the resulting “image” will also not contain land points. I think it is best to let this be decided by the input dataset.

The output of the resulting netCDF can have one of two possible “shapes”:

  • 2D variables with time on one axis and gpi on the other. This is kind of how SWI time series are stored already.
  • 3D variables with latitude, longitude and time as the three dimensions.

The decision of which it will be is dependent on the grid on which the input data is stored. If the grid has a 2D shape then the 3D solution will be chosen. If the input grid has only a 1D shape then only the 2D solution is possible.

Time Series aggregation

The chosen solution will use a custom function for each dataset to perform the aggregation if necessary. A simple example of a function that gets a time time series and aggregates it to a monthly time series could look like agg_tsmonthly

Simple example of a aggregation function

def agg_tsmonthly(ts, **kwargs):
    ts : pandas.DataFrame
        time series of a point
    kwargs : dict
        any additional keyword arguments that are given to the ts2img object
        during initialization

    ts_agg : pandas.DataFrame
        aggregated time series, they all must have the same length
        otherwise it can not work
        each column of this DataFrame will be a layer in the image
    # very simple example
    # aggregate to monthly timestamp
    # should also make sure that the output has a certain length
    return ts.asfreq("M")

Time series iteration

The function agg_tsmonthly will be called for every time series in the input dataset. The input dataset must have a iter_ts iterator that iterates over the grid points in a sensible order.

Interface to the netCDF writer

The netCDF writer will be initialized outside the ts2img class with a filename and other attributes it needs. So the ts2img class only gets a writer object. This writer object already knows about the start and end date of the time series as well as the target grid and has initialized the correct dimensions in the netCDF file. This object must have a method write_ts which takes a array of gpi’s and a 2D array containing the time series for these gpis. This should be enough to write the gpi’s into the correct position of the netCDF file.

This approach should also work if another output format is supposed to be used.

Implementation of the main ts2img class

The ts2img class will automatically use a the function given in agg_ts2img if no custom agg_ts2img function is provided. If the tsreader implements a method called agg_ts2img this function will be used instead.

class Ts2Img(object):

    Takes a time series dataset and converts it
    into an image dataset.
    A custom aggregate function should be given otherwise
    a daily mean will be used

    tsreader: object
        object that implements a iter_ts method which iterates over
        pandas time series and has a grid attribute that is a pytesmo
        BasicGrid or CellGrid
    imgwriter: object
        writer object that implements a write_ts method that takes
        a list of grid point indices and a 2D array containing the time series data
    agg_func: function
        function that takes a pandas DataFrame and returns
        an aggregated pandas DataFrame
    ts_buffer: int
        how many time series to read before writing to disk,
        constrained by the working memory the process should use.


    def __init__(self, tsreader, imgwriter,

        self.agg_func = agg_func
        if self.agg_func is None:
                self.agg_func = tsreader.agg_ts2img
            except AttributeError:
                self.agg_func = agg_tsmonthly
        self.tsreader = tsreader
        self.imgwriter = imgwriter
        self.ts_buffer = ts_buffer

    def calc(self, **tsaggkw):
        does the conversion from time series to images
        for gpis, ts in self.tsbulk(**tsaggkw):
            self.imgwriter.write_ts(gpis, ts)

    def tsbulk(self, gpis=None, **tsaggkw):
        iterator over gpi and time series arrays of size self.ts_buffer

        gpis: iterable, optional
            if given these gpis will be used, can be practical
            if the gpis are managed by an external class e.g. for parallel
        tsaggkw: dict
            Keywords to give to the time series aggregation function

        gpi_array: numpy.array
            numpy array of gpis in this batch
        ts_bulk: dict of numpy arrays
            for each variable one numpy array of shape
            (len(gpi_array), len(ts_aggregated))
        # have to use the grid iteration as long as iter_ts only returns
        # data frame and no time series object including relevant metadata
        # of the time series
        i = 0
        gpi_bulk = []
        ts_bulk = {}
        ts_index = None
        if gpis is None:
            gpis, _, _, _ = self.tsreader.grid.grid_points()
        for gpi in gpis:
            ts = self.tsreader.read_ts(gpi)
            ts_agg = self.agg_func(ts, **tsaggkw)
            for column in ts_agg.columns:
                except KeyError:
                    ts_bulk[column] = []

            if ts_index is None:
                ts_index = ts_agg.index

            i += 1
            if i >= self.ts_buffer:
                for key in ts_bulk:
                    ts_bulk[key] = np.vstack(ts_bulk[key])
                gpi_array = np.hstack(gpi_bulk)
                yield gpi_array, ts_bulk
                ts_bulk = {}
                gpi_bulk = []
                i = 0
        if i > 0:
            for key in ts_bulk:
                ts_bulk[key] = np.vstack(ts_bulk[key])
            gpi_array = np.hstack(gpi_bulk)
            yield gpi_array, ts_bulk