# Comparing methods for gap filling ... Comparing methods for gap filling in historical snow depth...

date post

28-Jul-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Comparing methods for gap filling ... Comparing methods for gap filling in historical snow depth...

Comparing methods for gap filling

in historical snow depth time series

1 WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland 2 Federal Office of Meteorology and Climatology MeteoSwiss, Zurich, Switzerland

Contact: johannes.aschauer@slf.ch

Johannes Aschauer1, Mathias Bavay1, Michael Begert2, Christoph Marty1

2

Introduction

Switzerland has a unique dataset of long-term manual daily snow depth time series ranging back

more than 100 years for some stations. This makes the dataset ideal to be analyzed in a

climatological sense. However, there are sometimes shorter (weeks, months) or longer (years) gaps

in these manual snow depth series, which hinder a sound climatological analysis and reasonable

conclusions. Furthermore, ongoing efforts towards homogenization of these manual snow depth time

series (Resch et al. 2020, Buchmann et al. 2019) require continuous time series.

For historical time periods (first half of the 20th century), we are limited in the amount of available

meteorological variables and lack a dense grid of snow measurement stations for spatial

interpolations. In this study we focus on simple approaches to reconstruct historical snow depth time

series.

3

Objective

Evaluate the capability of different methods to reproduce snow depth (HS) data in a single winter of

missing data at different stations.

The basic workflow:

1. Create synthetic gap of one winter (Nov-Apr).

2. Fill the gap winter with different techniques.

3. Evaluate the method’s accuracy based on the withheld data.

We use three different types of methods which are (a) Regression based methods, (b) Spatial

interpolation methods and (c) Temperature index models. In case we need to train model parameters,

we use data from the gap winter’s three preceding winters. Model accuracy is calculated by the two

score metrics root-mean-square error (RMSE) and mean arctangent absolute percentage error

(MAAPE, Kim & Kim 2016).

Methodology

4

Figure 1: Spatial distribution of evaluation stations and predictor stations.

Data

• We use data from the period 11/1999 until 05/2018 from meteorological observation stations in

Switzerland (Figure 1)

• Manually observed daily snow depth data from stations operated by both MeteoSwiss and SLF

is used for spatial interpolation and regression methods.

• Daily homogenized mean temperature and precipitation data based on services from

MeteoSwiss is used for driving temperature-index models.

• Evaluation is performed at 19 target

stations that have a complete record

of both snow depth and meteorological

data.

5

Regression based Methods

Regression models are trained with three preceding winters of the simulated gap winter (for the

winters 2000-2003, the three following years are used for training). The relationship found between

the predictors (i.e. HS at neighboring stations) and predictand (i.e. HS at target station) is then used

to reproduce snow depth in the gap winter. The relation of neighboring stations and target station has

to therefore assumed to be constant over training and gap periods.

PCA Regression

The 5 best correlated neighboring stations are used as initial predictors with standard scaling. A

principal component analysis (PCA) is calculated on those predictors and two main components of

the PCA are then used as predictors for a multilinear regression model.

Lasso Regression

A linear model with LASSO regularization is used (Tibshirani1996). The 10 best correlated snow

depth time series of neighboring stations are given to the model as potential predictors. Predictors

and predictand (i.e. target station) are standard scaled during training. The regularization parameter λ

is selected during a 5-fold cross validation.

Elastic Net Regression

A linear model with elastic net regularization is used (Zou, & Hastie 2005). The 20 best correlated

snow depth time series of neighboring stations are used as predictors. Predictors and predictand are

standard scaled during training. Regularization parameters alpha and l1_ratio are optimized in a 5-

fold cross validation.

6

Distance Weighting Methods

The 15 closest neighboring stations within a vertical limit of ±300 m and maximum horizontal distance of 100 km are used to spatially interpolate snow depth in the gap winter at the target

station. The number of neighboring stations is reduced if less than 15 stations are satisfying these

conditions.

Inverse Distance Squared

Simple inverse distance weighting is used as a benchmark without use of any lapse rate. The

inverse square of the distance of a neighboring station is used to calculate its weight.

GIDS

GIDS (gradient-plus-inverse-distance-squared) is a combination of inverse distance weighting and

multiple linear regression (Nalder & Wein, 1998). The method is also already used for

reconstructing snow depth time series in Austria (e.g. see contribution of Resch et al. here in the

same session).

7

Temperature Index Models

Two different temperature index models are used to calculate the snow water equivalent (SWE) of

the snowpack based on daily mean temperature and precipitation.

SNOW-17

An implementation of the SNOW-17 algorithm is used to model both snow depth (HS) and SWE

(Anderson 1976). SNOW-17 is a conceptual snow cover model in which the energy exchange at the

snow-air interface is calculated based on air temperature only.

SLF Temperature Index Model

A temperature index model that is run by the operational snow-hydrological service at the WSL

Institute for Snow and Avalanche Research SLF is used in order to model SWE at the target

stations. The model is driven by daily mean air temperature and precipitation.

8

Density model for transferring SWE to HS

In order to transfer SWE output from the temperature-index models to snow depth, we developed a

conceptual snow density model (SWE2HS). The model treats each increase of SWE during a day

as a new snow layer. The density 𝜌 𝑡 in each snow layer at day 𝑡 after deposition is calculated as

𝜌 𝑡 = 𝜌𝑚𝑎𝑥 + 𝜌𝑛𝑒𝑤 − 𝜌𝑚𝑎𝑥 ∗ 𝑒 −𝑡/𝜏

where 𝜌𝑛𝑒𝑤 is the density of new snow, 𝜌𝑚𝑎𝑥 is the density of settled old snow, and 𝜏 is a decay constant.

If SWE decreases during a day, snow

layers are removed from the top of

the snowpack in order to compensate

for the loss in SWE.

The model parameters 𝜌𝑛𝑒𝑤, 𝜌𝑚𝑎𝑥, and 𝜏 are optimized in order to match measured HS series. This is done in

the 3 preceding years of a simulated

gap winter. The best parameters

identified by the model are then used

to estimate HS in the simulated gap

winter from the modeled SWE series.

Figure 2: Density model SWE2HS employed at station DAV. SWE was

calculated with SNOW-17 .

9

Results

Figure 3: Examples of reproduced gaps with the different methods. Target station and hydrologic year of the gap-

winter are indicated in the subplots.

10

Figure 4: Boxplots of score values (RMSE left, MAAPE right) for the different methods evaluated over all target

stations and gap winters.

Results

11

Scores vs. mean snow depth in each gap winter.

Scores vs. Station altitude in each gap winter.

Results

Figure 5: RMSE (top) and MAAPE (bottom) of each filled gap winter plotted against target station altitude. Each column

depicts score values for a different method.

Figure 6: RMSE (top) and MAAPE (bottom) of each filled gap winter plotted against mean snow depth in the respective gap

winter. Each column depicts score values for a different method.

12

Conclusion

• Temperature-index and regression based methods are able to reconstruct historical HS data

gaps of one winter with RMSE scores below 20 cm in most example cases.

• Regression based methods do not guarantee good results for low elevation stations and

require suitable predictor stations.

• Temperature-index approaches should be preferred over spatial interpolation methods

whenever there is meteorological data available.

• When transferring these findings to gaps in the first half of the 20th century, it should be kept in

mind that we will have a much more sparse observation grid and accuracies of the spatial

interpolation methods are likely to decrease.

• Spatial interpolation methods can compete with the other approaches when introducing vertical

limits for the predictor stations.

Further investigations:

• a comparison of the here presented methods with the SNOWGRID model (Olefs et. al 2013),

which has been already applied for HS gap reconstruction in Austria.

• In case there is no meteorological data at a station, is it better to fir

*View more*