1 Space-time models for soil moisture dynamics Valerie Isham Department of Statistical Science University College London [email protected],

1

Space-time models for soil moisture dynamics

Valerie Isham

Department of Statistical ScienceUniversity College London

[email protected], http://www.ucl.ac.uk/stats/

2

Collaborators:

David Cox

Nuffield College, Oxford

Ignacio Rodriguez-Iturbe

Civil and Environmental Engineering, Princeton

Amilcare Porporato

Civil and Environmental Engineering, Duke

3

Introduction

Temporal models of soil moisture at a single-sitepoint rainfall (ie concentrated at discrete time points)

Spatial-temporal models of soil moisturespatially-distributed rainfall (at a point or temporally

distributed in time)variable vegetationproperties at a point and averaged over space-time

Coupled dynamics of biomass and soil moisturetemporal process at a single site

Summary and future directions

Overview

4

Fundamental problem of hydrological interest…

Soil moisture (and its spatial and temporal variability)

is the dynamic link between climate, soil and vegetation,

and impacts processes at a range of spatial scales.

Point scale: infiltration, plant dynamics, biogeochemical cycle

Hillslope: controlling factor for slope instability and land slides

Basin: drought assessment, flood forecasting

Region/continent: interaction with atmospheric phenomena

5

R u n o f f

IN P U T : R A IN F A L L( in te rm it te n t-

s to c h a s t ic )

t

h

E v a p o -tr a n s p ira t io n

T r o u g h fa l l

Z r

E f fe c t iv e p o ro s i ty , n

Z r

E f fe c t iv e p o ro s i ty , n

L e a k a g e

R u n o f f

Soil moisture……• increases due to precipitation• decreases due to evapotranspiration

and leakage and is dependent on • soil properties• vegetation

6

We consider dynamics

• at a daily time scale (no effects of diurnal fluctuations in temperature on evapotranspiration)

• within a single season

• on relatively small spatial scales (no feedback between soil moisture and rainfall).

The impact on the vegetation as well as of the vegetation is of interest.

La Copita, Texas;courtesy of Amilcare Porporato/ Steve Archer

7

0.05

0.15

0.25

150 200 250 300 350

Julian Day

q (%

)interspace

canopy

0

5

10

15

20

Pre

cip

itat

ion

(mm

day

)

Sevilleta, New Mexicocourtesy of Amilcare Porporato/Eric Small

Precipitation and soil moisture

8

Temporal process of soil moisture

Modelling approach

We use

piecewise deterministic Markov processes (Davis

1984) in continuous time: sample paths have

• periods of deterministic change governed by a

differential equation

• random jumps occurring at random times

9

S(t)

X2 X3

X4

X1

T1 T2 T3 T4 t

S(t): the Takács virtual waiting time process for a M/G/1 queue

ie the service requirement of all the customers in the system at t,

Alternatively: S(t) is the content of a store (reservoir)

* replenished by random amounts at random times

* subject to depletion at a constant rate when non-empty

A very simple such process ……

Times: a Poisson process, rate

Jumps: iid, density gDecay: constant rate

10

Let

and let S have density for s > 0

Forward equation:

Many properties of the process can be determined

Special case: Xi ~ exp( )

Equilibrium: if

11

Other properties and extensions (Cox and Isham, 1986)

• transient solution: Laplace transform (wrt to t) of the moment generating function

• expansions determining convergence to equilibrium

• autocovariance function, in equilibrium

• slowly varying arrival rate

(small )

12

For soil moisture• state-dependent decay

losses depend on current soil moisture level

• boundedness of soil moisture

excess rainfall runs off saturated soil

state-dependent jumps, density g(x,s)

13

Soil moisture balance equation:

n soil porosity

Zr depth of root zone

I (random) rate of infiltration (dependent on ground cover)

E rate of evapotranspiration (dependent on vegetation)

L rate of leakage (dependent on soil properties)

Standardise I, E, L

14

Losses…approximated by

0 s* s1 1.0

15

distribution of infiltration…Assume that standardised infiltration I*(s,t) has an

exponential ( ) distribution, truncated at 1- s

The excess rainfall is lost as surface run-off.

16

Forward equation… for density of S(t)

(no atom at origin since ).

Equilibrium distribution (Rodriguez-Iturbe et al 1999) has the form

Use piecewise linear form of continuity of p(s) at s* and s1.

Normalise to 1 to find c.

17

Note: the atom of probability at 1-s in the state dependent jumps is not used explicitly in the derivation. Soil saturation only affects the restricted range over which p(s) is normalised – an effect of the Markov nature of the soil moisture process.

properties, impact of parameters on properties etc

Note: Equilibrium distribution is for linear evapotranspiration

18

Impact of climate, soil and vegetation on equilibrium distribution

Parameters chosen to represent

a) tropical climate and vegetation, frequent moderate rainfall, deep soil;

b) hot arid region, shallow sandy soil, mixture of trees and grasses;

c) cold arid region; d) forested temperate

region.

19

Spatial-temporal soil moisture

Soil moisture is spatially dependent, because of• correlated rainfall input• ground topology causing run-off from one location to

affect nearby locations• correlated vegetation cover We assume • a stochastic process of rain cells with random spatial

extents• a flat landscape to avoid run-off problems, eg savannah • a) a homogenous vegetation, or b) a stochastic process of trees with random canopies in a

grassy landscape

20

The simplest model…• temporally instantaneous rainfall (ie daily timescale) at random times Tk

• linear losses (hot arid region, cf Fig (b)) ( will be vegetation and soil-dependent)

• ignore bound on soil moisture

In this case

• proportional interceptionstandardised infiltration for rainfall

• heterogeneous soil and vegetation and depend on location

21

Equilibrium distribution for….

hot arid region, shallow sandy soil, mixture of trees and

grasses; s* = 0.45

22

S(t)

t

Shot-noise process

Linear losses ( ) and no saturation

exponential decay:

if there is no input in (0,t). In this case

and S(t) has no atom at 0.

23

Rainfall process…

• Poisson process of rain cell origins, rate in space-time

• circular cells, random radii (iid)

• rainfall is instantaneous in time over the cell, depths Y (iid)

• at a fixed location, A say, rain events occur in a temporal Poisson process of rate

• events occur at locations A and B, d apart, in a temporal Poisson process of rate

Here

is the area of overlap of two unit discs, centres u apart.

24

Marginal distribution …

• Transient distribution and its properties

• Equilibrium distribution

where is the mgf of the rain depth Y, with

• If Y ~ exp( ), S ~

• For general infiltration, replace integrand by

where is the mgf of infiltration from a rain depth Y

25

Joint distribution: sites A and B, d apart…• SA (t) - rain events before t that only affect A, rate

- rain events before t that affect both A and B, rate

• SB (t+h) - events before t+h that only affect B, rate

- events in (t, t+h) that affect both A and B, rate

- events before t that affect both A and B, rate • Properties of transient distribution• Equilibrium distribution

26

In particular

For general infiltration

Joint equilibrium mgf:

27

As before, assume • Poisson process of rain cell origins, rate in space-time,

and circular cells, random radii (iid) Assume • rain cell duration D, with constant intensity V (iid)

Observe their superposition where

Soil moisture

(assuming, as before, linear losses, proportional interception and ignoring bound on soil moisture)

Alternative model: rain cells with exponential durations…

28

formal solution…

In particular, the covariance properties of

(assuming D ~ exp( ) ) imply those of S (via Campbell’s Th)

The corresponding covariance for the pulse rainfall model is

29

Properties for homogeneous vegetation

• Correlation as a function of the spatial and temporal lags

• Effect of spatial averaging (different spatial scales). Analytic results can be obtained by using a Gaussian approximation to

• Effect of spatial and temporal averaging (different scales)

30

Correlation as a function of spatial and temporal lags

(rainfall parameters fitted to data from 17 gauges in Southern Italy, two values for soil porosity-root depth factor)

nZr=100mm nZr=500mm

31

Standard deviation of spatially averaged field relative to standard deviation at a point

Here is the mean rain cell radius.The ratio depends only on and the spatial area

32

Standard deviation of spatially and temporally averaged fields

33

Heterogeneous vegetation – trees in a grassy landscape

A model for tree crowns…..

• Poisson process of tree locations, rate in space

• Circular canopies, random radii (iid)

• No. of trees covering location, A say,

• No. of trees covering A and B, d apart,

• P(neither A nor B covered)

• P(A is covered, B is not)

• P(both A and B are covered)

34

a realisation of the vegetation process…

35

Use probabilities to remove conditioning of previous results on vegetation cover, and determine corresponding properties with random vegetation

eg variance of spatially integrated soil moisture

10-6

10-4

10-2

100

102

104

106

10-4

10-3

10-2

10-1

Area (km2)

Var

ian

ce

HeterogeneousAll TreeAll Grass

Slope -0.915

36

Biomass and soil moisture…temporal process

For water-limited ecosystems, a simple model for the coupled system of biomass B and soil moisture S is

Assume (within a growing season)

37

transient solution…

moments, eg

Equilibrium:

(deterministic)

38

Summary and scope for further work• Single-site, temporal models of soil moisture

• Spatial-temporal models of soil moisture *simplifications - flat landscape

- linear evapotranspiration - ignore bound at s = 1

*spatially-distributed rainfall instantaneous distributional results temporally distributed second order results

(proportional interception only)variable vegetationareally-averaged properties

• Coupled dynamics of biomass and soil moisture *single site, temporal process