Gapfill {gapfill}

Main Function for Gap-Filling

Description

The function fills (predicts) missing values in satellite data. We illustrate it with MODIS NDVI data, but it can also be applied to other data, that is recorded at equally spaced points in time. Moreover, the function provides infrastructure for the development of new gap-fill algorithms. The predictions of the missing values are based on a subset-predict procedure, i.e., each missing value is predicted separately by (1) selecting a subset of the data to a neighborhood around the missing value and (2) predicting the values based on that subset.

Usage

Gapfill(
  data,
  fnSubset = Subset,
  fnPredict = Predict,
  iMax = Inf,
  nPredict = 1L,
  subset = "missing",
  clipRange = c(-Inf, Inf),
  dopar = FALSE,
  verbose = TRUE,
  ...
)

Arguments

data

Numeric array with four dimensions. The input (satellite) data to be gap-filled. Missing values should be encoded as NA. When using the default Subset and Predict functions, the data should have the dimensions: x coordinate, y coordinate, seasonal index (e.g., day of the year), and year. See the ndvi dataset for an example.

fnSubset

Function to subset the data around a missing value. See Subset and Extend for more information.

fnPredict

Function to predict a missing value based on the return value of fnSubset. See Predict and Extend for more information.

iMax

Integer vector of length 1. The maximum number of iterations until NA is returned as predicted value.

nPredict

Integer vector of length 1. Specifies the length of the vector returned from fnPredict. Values larger than 1 may increase memory usage considerably.

subset

If "missing" (default), all missing values of data are replaced by predictions. If "observed", all observed values of data are replaced by predictions. If a logical array of dimension dim(data) or a vector with positive integers, the elements data[subset] are replaced by predictions. Note that independently of the specified subset all provided values in data are used to inform the predictions.

clipRange

Numeric vector of length 2. Specifies the lower and the upper bound of the filled data. Values outside this range are clipped accordingly. If nPredict is larger than 2, only the first return value of fnPredict will be clipped.

dopar

Logical vector of length 1. If TRUE, the %dopar% construct from the R package foreach is used. This allows the function to predict several missing values in parallel, if a parallel back-end (e.g., from the R package doParallel or doMpi) is available. See the example below and foreach for more information.

verbose

Logical vector of length 1. If TRUE (default), messages are printed.

...

Additional arguments passed to fnSubset and fnPredict.

Details

The predictions of the missing values are based on a subset-predict procedure, i.e., each missing value is predicted separately by (1) selecting a subset of the data to a neighborhood around it and (2) predicting the values based on that subset. The following gives more information on this subset-predict strategy.
Missing values are often unevenly distributed in data. Therefore, the size of a reasonable subset may be different depending on the position of the considered missing value. The search strategy to find that subset is encoded in fnSubset. The function returns different subsets depending on the argument i. The decision whether a subset is suitable and the prediction itself is implemented in fnPredict. To be more specific, the subset-predict procedure loops over the following two steps to predict one missing value:

(1)

The function fnSubset is provided with the argument i = i (where i <- 0 in the first iteration) and returns a subset around the missing value.

(2)

The function fnPredict decides whether the subset contains enough information to predict the missing value. If so, the predicted value is returned. Otherwise, the function returns NA and the algorithm increases i by one (i <- i + 1) before continuing with step (1).

The procedure stops if one of the following criteria is met:

• fnPredict returns a non-NA value,

• iMax tries have been completed,

• fnSubset returns the same subset two times in a row.

Value

List of length 4 with the entries:

• fill contains the gap-filled data. If nPredict = 1, fill is an array of dimension dim(data), otherwise the array is of dimension c(dim(data), nPredict).

• mps integer vector of length equaling the number of predicted values. Contains the (1 dimensional) indices of the predicted values.

• time list of length 4 containing timing information.

  • start start date and time.

  • end end date and time.

  • elapsedMins elapsed minutes.

  • elapsedSecsPerNA elapsed seconds per predicted value.

• call call used to produce the object.

Note

The default Predict function implements the prediction of the missing value and can also return lower and upper bounds of an approximated 90% prediction interval. See the help page of Predict for more information on the prediction interval. The example section below shows how the prediction interval can be calculated and displayed.

To tailor the procedure to a specific dataset, it might be necessary to adapt the subset and/or the prediction strategy. On the one hand, this can be done by changing the default arguments of Subset and Predict through the argument ... of Gapfill. See the help of the corresponding functions for more information about their arguments. On the other hand, the user can define a new subset and predict functions, and pass them to Gapfill through the arguments fnSubset and fnPredict. See Extend for more information.

The current implementation of Subset does not take into account that values at the boundaries of data can be neighboring to each other. For example, if global data (entire sphere) are considered, data[1,1,,] is a neighbor of data[dim(data)[1], dim(data)[2],,]. Similar considerations apply when data are available for an entire year. To take this into account, the Subset function can be redefined accordingly or the data can be augmented.

There are two strategies to run the gap-filling in parallel. The first one is to set the argument dopar of Gapfill to TRUE and to use an openMP or MPI parallel back-end. The parallel back-end needs to be setup before the call to Gapfill. An example using the R package doParallel is given below. Note that there exist other parallel back-ends implemented in other packages; such as, e.g., the package doMpi. Some parallel back-ends are platform dependent. While this approach shortens the process time by distributing the computational workload, it does not reduce the memory footprint of the procedure. The second strategy, which also reduces memory usage, is to split the data into several independent chunks. Whether data chunks are independent or not depends on the function provided to fnSubset. For example, the default Subset function never includes data that is further apart from the missing value than 1 seasonal index. Hence, data[,,1:3,] can be used to gap-fill data[,,2,].

Author(s)

Florian Gerber, [e-mail].

References

F. Gerber, R. de Jong, M. E. Schaepman, G. Schaepman-Strub, and R. Furrer (2018) in IEEE Transactions on Geoscience and Remote Sensing, pp. 1-13, doi: 10.1109/TGRS.2017.2785240.

See Also

Extend, Subset-Predict, Image.

Examples

## Not run: 
out <- Gapfill(ndvi, clipRange = c(0, 1))

## look at input and output
str(ndvi)
str(out)
Image(ndvi)
Image(out$fill)

## run on 2 cores in parallel
if(require(doParallel)){
  registerDoParallel(2)
  out <- Gapfill(ndvi, dopar = TRUE)
}

## return also the prediction interval
out <- Gapfill(ndvi, nPredict = 3, predictionInterval = TRUE)

## dimension has changed according to 'nPredict = 3'
dim(out$fill)

## clip values outside the valid parameter space [0,1].
out$fill[out$fill < 0] <- 0
out$fill[out$fill > 1] <- 1

## images of the output:
## predicted NDVI
Image(out$fill[,,,,1])
## lower bound of the prediction interval
Image(out$fill[,,,,2])
## upper bound of the prediction interval
Image(out$fill[,,,,3])
## prediction interval length
Image(out$fill[,,,,3] - out$fill[,,,,2])


## End(Not run)