Streamflow reconstruction with mass balance adjustment

library(mbr)

Introduction

The package has two main functions, mb_reconstruction() for reconstruction, and cv_mb() for cross-validation. This vignette will demonstrate these functions using the two built-in data sets. First, let’s look at the built-in data, taken from Nguyen et al (2021).

The data frame p1Seasonal contains the reconstruction targets, namely the dry season, wet season, and water year streamflow for the Ping River at station P.1 (Chiang Mai, Thailand). The data span from 1922 to 2003.

p1Seasonal
#>      season year       Qa
#>   1:     NJ 1922  576.020
#>   2:     NJ 1923  629.170
#>   3:     NJ 1924  583.900
#>   4:     NJ 1925  680.410
#>   5:     NJ 1926  564.320
#>  ---                     
#> 242:     WY 1999 1338.252
#> 243:     WY 2000 1235.269
#> 244:     WY 2001 1898.075
#> 245:     WY 2002 1770.417
#> 246:     WY 2003 1853.401

As paleoclimate proxies, we use the principal components (PCs) of the Southeast Asian Dendrochronology Network. A set of PCs has been derived for each target (see details in Nguyen et al, 2020). These are provided in pc3seasons.

str(pc3seasons)
#> List of 3
#>  $ NJ: num [1:254, 1:4] -1.2142 -0.2256 -0.0803 1.6726 3.1718 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : NULL
#>   .. ..$ : chr [1:4] "PC1" "PC5" "PC6" "PC7"
#>  $ JO: num [1:254, 1] 0.51 -0.417 2.113 2.968 1.692 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : NULL
#>   .. ..$ : chr "PC1"
#>  $ WY: num [1:254, 1] -0.242 -0.439 2.016 3.242 2.508 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ : NULL
#>   .. ..$ : chr "PC1"

The tree ring data spans from 1750 to 2003. Let us look at the first 10 rows of each principal component matrix.

lapply(pc3seasons, head, n = 10)
#> $NJ
#>               PC1        PC5        PC6        PC7
#>  [1,] -1.21420806  0.7473268 -0.5783063 -0.7674267
#>  [2,] -0.22561311 -0.2710709 -0.5699502  0.7896157
#>  [3,] -0.08031591 -0.5816381 -0.7841998  0.9286036
#>  [4,]  1.67255722 -1.5043708 -1.2338918 -1.3938897
#>  [5,]  3.17182495 -0.7129998  0.7000585 -0.9904258
#>  [6,]  1.02307742 -1.4861251 -0.1496151 -0.2286487
#>  [7,]  1.30705152 -0.5124227 -0.2053901 -0.2480695
#>  [8,]  0.88386663  1.0840077 -0.1134354 -0.8353086
#>  [9,] -0.03419247  0.7223128  1.0412121  1.4769790
#> [10,] -1.11164309 -1.5384658  1.8846923 -0.4261086
#> 
#> $JO
#>              PC1
#>  [1,]  0.5098988
#>  [2,] -0.4166514
#>  [3,]  2.1134018
#>  [4,]  2.9683959
#>  [5,]  1.6922489
#>  [6,] -0.5483287
#>  [7,]  2.1975923
#>  [8,]  1.0627828
#>  [9,]  0.3118631
#> [10,] -0.7735868
#> 
#> $WY
#>              PC1
#>  [1,] -0.2417374
#>  [2,] -0.4387672
#>  [3,]  2.0158877
#>  [4,]  3.2423019
#>  [5,]  2.5075143
#>  [6,] -0.3573139
#>  [7,]  1.7650888
#>  [8,]  0.8676295
#>  [9,] -0.7693806
#> [10,] -0.1852180

Reconstruction

We build a reconstruction with the full data set.

fit <- mb_reconstruction(
  instQ = p1Seasonal,
  pc.list = pc3seasons,
  start.year = 1750,
  lambda = 1,
  log.trans = 1:3
)

We need to provide the instrumental data (instQ) and the PC list (pc.list). Since the PC data do not have a time column, we need to provide start.year, 1750 in this case.

For the mass balance adjustment, we need to provide a penalty weight lambda. The default value is 1 and it works in this case. But for other applications you may need to test a few values for lambda to figure out the optimal value.

Finally, the argument log.trans provides the indices of the targets that need to be log transformed. Here we transform all three targets.

Let’s look at the results.

fit
#>      season year         Q lambda
#>   1:     NJ 1750  596.9214      1
#>   2:     NJ 1751  666.2604      1
#>   3:     NJ 1752  697.5015      1
#>   4:     NJ 1753 1146.8719      1
#>   5:     NJ 1754 1171.9384      1
#>  ---                             
#> 758:     WY 1999 1866.2336      1
#> 759:     WY 2000 2139.2197      1
#> 760:     WY 2001 2388.0273      1
#> 761:     WY 2002 1763.4657      1
#> 762:     WY 2003 1815.3354      1

Cross-validation

Let us now cross-validate the model with a hold-out-25% scheme. The cross-validation folds can be created with the function make_Z().

# Create hold-out chunks
set.seed(24)
cvFolds <- make_Z(
  obs = 1922:2003,
  nRuns = 50, 
  frac = 0.25,
  contiguous = TRUE
)
# Run cross validation
cv <- cv_mb(
  instQ = p1Seasonal,
  pc.list = pc3seasons,
  cv.folds = cvFolds,
  start.year = 1750,
  lambda = 1,
  log.trans = 1:3,
  return.type = 'metric means'
)
# Round up to two decimal places
cv[, (2:6) := lapply(.SD, round, digits = 2), .SDcols = 2:6][]
#>    season   R2   RE   CE nRMSE  KGE
#> 1:     NJ 0.46 0.52 0.43  0.24 0.51
#> 2:     JO 0.43 0.39 0.28  0.26 0.42
#> 3:     WY 0.46 0.49 0.39  0.22 0.46

References

Nguyen, H. T. T., Galelli, S., Xu, C., & Buckley, B. (2020). Multi-Proxy, Multi-Season Streamflow Reconstruction with Mass Balance Adjustment. Earth and Space Science Open Archive, 22. https://doi.org/10.1002/essoar.10504791.1