Russia land cover mapping from VGT S10 data

1
Russia land cover mapping from VGT S-10 data
Global Land Cover 2000
Sergey Bartalev International Forest Institute,
Moscow, Russia Visiting Scientist in the JRC of
the European Commission, Ispra, Italy
Global Vegetation Monitoring Unit
2
Geographical extent of the study and used
SPOT4-VGT data
Global Land Cover 2000
Type of products used SPOT 4-VGT S10 products
including spectral and angular data Geographical
extent 420N - 750N and 50E -1800E with
particular attention to Russian territory and the
boreal zone of Eurasia Time window from end of
March 1999 until beginning of November 1999
Global Vegetation Monitoring Unit
3
Distribution of the forest in the World
Global Land Cover 2000
Boreal and Temperate Zone
World
Global Vegetation Monitoring Unit
4
The key elements being taken into consideration
to design the land cover mapping method

• requirements of users and particularly at
  national level
• the satellite data properties allow to
  distinguish the land cover types
• regional environment peculiar properties
• availability of auxiliary (non satellite) data
• practical realizability due to existing
  technical and other limitation

Global Land Cover 2000
Global Vegetation Monitoring Unit
5
The main options and obstacles to classify the
land cover types with SPOT4-VGT data

• Options
• spectral properties the land cover types
• spectral-phenological changes of the land cover
• the angular anisotropy of reflected radiation of
  land cover
• Obstacles
• presents of the pixels contaminated by
  clouds/shadow and snow
• presents of the pixels contaminated MIR defective
  detectors
• Sun/View directional dependence of spectral
  response
• dependence of phenological changes both from time
  of observation and geographical location

Global Land Cover 2000
Global Vegetation Monitoring Unit
6
Relation between main options and obstacles to
classify the land cover types with SPOT4-VGT data
Spectral features (single image)
Global Land Cover 2000
Contamination by clouds, snow and MIR defective
detectors
Global Vegetation Monitoring Unit
7
Separability of main land cover types with
spectral signatures derived from single S-10
SPOT4-VGT image
Global Land Cover 2000
Global Vegetation Monitoring Unit
8
Separability of different species forest with
spectral signatures derived from single S-10
SPOT4-VGT image
Global Land Cover 2000
Global Vegetation Monitoring Unit
9
Steps of the satellite data pre-processing

• Producing the mask of contaminated pixels related
  to snow, clouds and MIR channel defective
  detectors with the following steps procedure
• detection of utterly contaminated pixels with
  pre-specified thresholds
• detection of slightly contaminated pixels with
  pixel-by-pixel adaptive thresholds derived from
  time series of data
• Hot-spot factor normalization of spectral
  reflectance with BRDF model (subsidiary and
  optional step)
• Producing the seasonally optimised composites of
  spectral channels
• Producing of the spectral-angular parameters with
  two options are considered
• statistics of Sun-Earth-Sensor relative angular
  parameters derived under condition of maximum
  NDVI pixels selection
• BRDF model derived parameters estimation

Global Land Cover 2000
Global Vegetation Monitoring Unit
10
Clouds and Snow spectral properties. Normalised
Different Snow Indexes NDSI
Global Land Cover 2000
NDSI (CH1-CH4) / (CH1CH4)
From Hall et al., 1998 "Algorithm Theoretical
Basis Document (ATBD) for the MODIS Snow-, Lake
Ice- and Sea Ice-Mapping Algorithms. Version 4.0"
Global Vegetation Monitoring Unit
11
Detection of the contaminated pixels

• Step 1 Detection of the pixels utterly
  contaminated by snow and clouds with
  pre-specified thresholds
• Snow/Ice Cs (?, t) 1 ? ??R1 (?, t) gt 0.1
  AND NDSI (?, t) gt 0.1
• Clouds Cc (?, t) 1 ? ? ? R1 (?, t) gt
  0.1 AND - 0.1 lt NDSI (?, t) lt 0.1
• where
• ? - geographical co-ordinates t - fixed time of
  observation (decade)
• Ri ( i1?4) - reflectance in the channel i NDSI
  - snow indexes
• Cs - set of snow detected pixels Cc - set of
  clouds detected pixels
• CP1 ? Cs U Cc - set of contaminated pixels
  at the 1st step
• Steps 2?J Detection of the defective detectors
  and slightly contaminated by snow/clouds pixels
  with adaptive thresholds derived from time series
  of data
• CP'j (?, t) 1 ? t ? R4 (?, t) ? Mj (R4 (?) ?
  CPj-1(?, t) ? 1) 2SDj (R4 (?) ? CPj-1(?, t)
  ? 1)
• OR ? R4 (?, t) ? Mj (R4 (?) ? CPj-1(?,
  t) ? 1) - 2SDj (R4 (?) ? CPj-1(?, t) ? 1)
• Mj (R4 (?) and SDj (R4 (?) - mean and standard
  deviation of R4 (?, t) at the step j
• ? - fixed geographical co-ordinates
• CPj ? CPj-1 U CP'j - set of contaminated
  pixels at the step j (j2?J)

Global Land Cover 2000
Global Vegetation Monitoring Unit
12
Producing of the seasonal mosaics
Main zonal ecosystems of Russia
Global Land Cover 2000
Snow duration map derived from SPOT4-VGT S-10 data
Global Vegetation Monitoring Unit
13
The seasonal composites of S-10 images
spring
Global Land Cover 2000
summer
autumn
Global Vegetation Monitoring Unit
SWIR-NIR-R
14
Comparison of the summer and autumn seasonal
composites of S-10 images
Global Land Cover 2000
summer
autumn
Global Vegetation Monitoring Unit
NIR-SWIR-R
15
Producing of the spectral-angular parameters

• Objective to investigate the possibilities to
  derive from time series of SPOT4-VGT data the
  parameters that are sensitive to the structure
  (forest density, height and etc.) of observed
  surface based on the angular anisotropy of
  reflected radiation
• Two options are considered
• statistical analysis of Sun-Earth-Sensor
  relative angular parameters derived under
  condition of maximum NDVI pixels selection
• parameters derived from BRDF model

Global Land Cover 2000
Global Vegetation Monitoring Unit
16

Averaging of Sun-Earth-Sensor relative angular
parameters from SPOT4-VGT S-10 time series
products
Color composite of M(PHA) and M(VZA) and M(SZA)
Global Land Cover 2000
M (?) Mj (? (?, t) ? t ? CP(?, t) ? 1) where
? - one of Sun-Earth-Sensor angular parameter t
- time of observation (decade) ? - fixed
geographical co-ordinates CP(?, t) - mask of
contaminated pixels M (?) - mean of ?
Z
SZA
N
SAA
VAA
VZA - view zenith angle SZA - Sun zenith
angle PHA - phase angle
17
Comparison of Sun-Earth-Sensor time-averaged
angular parameters from SPOT4-VGT S-10 products
with forest map
Color composite of mean values of PHA and VZA and
SZ
Global Land Cover 2000
Z
SZA
N
Forest map
SAA
VAA
18
BRDF model based approach to derive
spectral-angular parameters from time series of
S10 products
The linearized MRPV model of BRDF
Global Land Cover 2000
- Sun Zenith, View Zenith and Phase angles
respectively
where RNi - reflectance in the spectral channel i
normalized for the hot-spot factor

RNi A RNj B, where i and j - spectral
channels and i?j An estimations of A and B
coefficients of linear equation are considering
as parameters sensitive to surface structure
Global Vegetation Monitoring Unit
19
An estimation of the linear equation coefficients
between pairs of normalized reflectances in two
channels of S10 products
An estimation of slope A and interception B with
moving time window along profiles of NIR and MIR
channels
Global Land Cover 2000
Maximum is R2 0.93
Global Vegetation Monitoring Unit
20

Color composite of Slope and Interception values
of linear equation derived with pairs of RED-NIR
and NIR-MIR channels
RED-NIR Slope - Slope - Interception
Global Land Cover 2000
Z
SZA
N
SAA
VAA
NIR-MIR Slope - Slope - Interception
21

An example of relationship between Slope values
of linear equation derived with NIR-MIR channels
and forest density
The forest density values are estimated in the
set of grid cells based on forest inventory data
base
Global Land Cover 2000
Z
SZA
N
SAA
VAA
Standard deviation of Slope values for each class
of forest density is estimated in range of 0.4-0.5
22
Possible steps of the satellite data thematic
classification

• Eco-regional stratification
• Ecostrata-by-ecostrata unsupervised clustering of
  spectral-seasonal composites 3 mosaics x RED,
  NIR and MIR channels.
• Two different clustering algorithm are
  considering
• ISODATA ERDAS
• ELBG
• Clusters labeling (if needs with additional
  splitting of clusters using auxiliary
  non-satellite data) into thematic classes with
  using
• existing forest maps and forest inventories data
• high-resolution satellite imagery
• digital elevation model and topomaps
• Splitting of forest related classes to 2-3 cover
  density categories with spectral-angular
  parameters derived from satellite data (have to
  be investigated additionally)

Global Land Cover 2000
Global Vegetation Monitoring Unit
23
Eco-regional stratification
Global Land Cover 2000
Global Vegetation Monitoring Unit
24
The lessons learned from Russia land cover
mapping exercise with SPOT4-VGT data

• Pre-processing
• the procedure of contaminated pixels detection
  is developed and applied with satisfactory
  results
• the developed multi-temporal data composting
  procedure allows to produce the seasonal mosaics
  that are almost free from contaminated pixels and
  angular effect and gives possibility to use
  phenological changes of land cover as an
  additional option for thematic classification
• the possibility to retrieve from
  spectral-angular data an information on surface
  structure has been demonstrated. An practical
  benefit of spectral-angular data for land cover
  mapping are limited with using S-10 product and
  still have to be clarified. It is very likely
  that better result can be obtained with daily
  data (S-1 and P products)

Global Land Cover 2000
Global Vegetation Monitoring Unit
25
The lessons learned from Russia land cover
mapping exercise with SPOT4-VGT data

• Thematic classification
• Conclusions on clustering algorithm
• the ELBG gives significantly better results then
  ERDAS ISODATA algorithm when both of them are
  applied at the continental level
• on the eco-regional level ERDAS ISODATA
  algorithm is allow to perform the clustering with
  acceptable results
• the ecological stratification is critical issue
  to classify the forest at the level of main group
  of species (for example dark and light
  coniferous). Without applying any stratification
  it is foreseeable to classify the coniferous,
  deciduous and mixed forest.
• the efficient way to integrate seasonal mosaics
  into the classification procedure have to
  developed
• the possibility to classify the forest according
  to its density is required to be investigated
  additionally

Global Land Cover 2000
Global Vegetation Monitoring Unit