1 snow cover mapping using multi-temporal meteosat-8 data martijn de ruyter de wildt jean-marie...

16
1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy and Photogrammetry, ETH Zürich, Switzerland * MeteoSwiss, Zürich, Switzerland ** now at: ESA-ESRIN, Directorate of Earth Observation, Rome, Italy A fellowship, in cooperation with

Upload: sophie-fitzgerald

Post on 05-Jan-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

1

Snow cover mapping usingmulti-temporal Meteosat-8 data

Martijn de Ruyter de WildtJean-Marie Bettems*

Gabriela Seiz**Armin Grün

Institute of Geodesy and Photogrammetry, ETH Zürich, Switzerland

* MeteoSwiss, Zürich, Switzerland

** now at: ESA-ESRIN, Directorate of Earth Observation, Rome, Italy

A fellowship, in cooperation with

Page 2: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

2

Introduction

Objective: to obtain accurate snow cover maps for the numerical weather

prediction model of MeteoSwiss (aLpine Model, aLMo).

Main problem: discrimination between ice clouds and snow.

• Use high temporal frequency of MSG (15 minutes) in addition to spectral

capabilities (12 channels) to improve separation of clouds and snow

• in real-time, fully automatic

• usable over alpine terrain

Page 3: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

3

Data

Areas of interest:model domains of aLMo (western and central Europe). Resolution: 7 and 2.2 km.

Training and validation periods: 8 - 10 March, 2004 23 - 24 February, 2005(only day-time images)

8+1 spectral bands used: 1 VIS 0.635 m 2 VIS 0.81 m 3 NIR 1.64 m 4 IR 3.92 m 7 IR 8.70 m 9 IR 10.80 m 10 IR 12.00 m 11 IR 13.40 m 12 HR-VIS 0.70 m

Page 4: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

4

r1.6

BT3.9 - BT10.8BT10.8

Spectral image classification: “traditional” features (10-3-2004, 12:12 UTC)

r0.81

snow

ice cloud

snow

snowsnow

ice cloud

ice cloud

ice cloud

Page 5: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

5

BT3.9 - BT10.8

BT3.9 - BT13.4

Improved spectral classification II

BT3.9 - BT10.8: snow is as dark as or darker than ice clouds;

BT3.9 - BT13.4: snow is as dark as or brighter than ice clouds;

=> the following feature should enhance the contrast between snow and ice clouds:

13.43.9

10.83.9

BTBT

BTBT

−−

(BT3.9 - BT10.8) / (BT3.9 - BT13.4 )

snow

ice cloud

snow

ice cloud

snow

ice cloud

Page 6: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

6

Spectral classification

classification result:

white : snowdark gray : cloudslight gray : snow-free landblack : sea

UTC:200403101212

clouds

snow

Page 7: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

7

Temporal classification

Temporaltest

snow

Page 8: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

8

∑ ∑−= −=

σ=1

1i

1

1jj,i,mm wd

Temporal classification

Temporal variability can be quantified for each channel m with:

where ( )∑−=

−=2

2

2

,4

1

tmtmm IIσ

more ice more water more ice more water

Page 9: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

9

Temporal classification

The temporal standard deviations of the 9 used channels form a 9-dimensional parameter space,

where some of the parameters are correlated with each-other.

Reduce data redundancy: principal components analysis (PCI); when applied to the difference

between two images, the change information is concentrated into fewer dimensions (Gong, 1993).

Here:

- standardised PCI (applicable to data with variables at different scales)

- applied to the 9 temporal standard deviations

Normalised eigenvalues of the 9 new components, averaged over all training data:

1 0.5872 0.2883 0.0794 0.0245 0.0136 0.0067 0.0028 0.0019 0.000

Change information

noise

Page 10: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

10

First principal component of thetemporal standard deviation(10-3-2004, 12:12 UTC):

Second and third componentsare also useful for detectingclouds.

more ice more water

clouds

snow

Page 11: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

11

white : snowdark gray : cloudslight gray : snow-free landblack : sea

UTC:200403101212

UTC:200403101212

temporal

spectral

temporal cloudmask is ‘liberal’, only used to check snowy pixels for misclassifications:

spectral/temporal

Spectral and temporal classification

UTC:200403101212

UTC:200403101212

Page 12: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

12

Composite snow map, March 10th, 2004, 07:00 - 12:00 UTC

March 10th, 2004, 12:12 UTC

white: snow dark gray: clouds light gray: snow-free land black:sea

spectral/temporal

UTC:200403101212

Composite snow map, March 8th - March 10th

spectral/temporal

spectral/temporal

Composite snow maps

Page 13: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

13

Composite snow maps: spectral vs. spectral/temporal

March 10th, 2004, 07:00 - 12:00 UTC

white: snow dark gray: clouds light gray: snow-free land black:sea

spectral spectral/temporal

Page 14: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

14

High resolution visible (hrv) channel

RGB image, red= rhrv, green= r1.6 (low res.), blue= (low res.)

red pixels: surface snow OR ice clouds 13.43.9

10.83.9

BTBT

BTBT

−−

Page 15: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

15

Classification of hrv channel

Use low resolution cloud mask and temporal variability in hrv channel to detect clouds.

Composite snow map, March 10th, 2004, 07:00 - 12:00 UTC

Page 16: 1 Snow cover mapping using multi-temporal Meteosat-8 data Martijn de Ruyter de Wildt Jean-Marie Bettems* Gabriela Seiz** Armin Grün Institute of Geodesy

16

Conclusions:

• new spectral feature detects more clouds than

BT3.9 - BT10.8 alone and is less influenced by the solar zenith angle

• spectral classification separates snow and clouds reasonably well,

but: some clouds have the same spectral signature as snow

• using temporal information, most of these clouds can be detected

• temporal classification classifies snow in a conservative way

(somewhat too little snow detected, but with high certainty)

• high frequency strongly reduces cloud obscurance

• snow mapping also possible in hrv channel

• start of implementation at MeteoSwiss this winter

13.43.9

10.83.9

BTBT

BTBT

−−