Advances in mapping land use/land cover (LULC) and irrigated areas Exploring Ganges basin using continuous time series MODIS Data

1
Advances in mapping land use/land cover (LULC)
and irrigated areas Exploring Ganges basin using
continuous time series MODIS Data

• Hugh Turral, Prasad Thenkabail, and Mitch Schull

2
Hugh OVERVIEW
3

• Advances in Mapping LULC and Irrigated Areas
• OVERVIEW of This LECTURE
• Goals
• Advanced Sensor Characteristics
• Study Area
• Data Preparation and normalizations
• Methods
• Ground-truth Data
• Results
• Acurracy Assessment
• Conclusions

4
GOALS
5
Advances in Mapping LULC and Irrigated Areas GOALS
Mapping land use/land cover (LULC) and irrigated
area classes in the Ganges and Indus river basins
using near-continuous time-series 500-m MODIS
data.
New Mega dataset concepts New Suite of
Methods, techniques, and approaches Linkages
between LULC and land cover (LC) percentages
and Fuzzy Classification Accuracy Assessment.
6
STUDY AREA
7
Advances in Mapping LULC and Irrigated Areas
STUDY AREA
8
Few more slides
9
Prasad Advances in Mapping LULC and Irrigated
Areas
10
Advances in METHODS, TECHNIQUES, and
APPROACHES FirstDATA CHARACTERISTICS
RELATED
11
Advances in Mapping LULC and Irrigated Areas
First..DATA CHARACTERISTICS RELATED

• ADVANCES in SENSOR CHARACTERISTICS
• Band-width Narrow bands (lt30 nm) compared to
  broad bands (gt30 nm)
• Radiometric 12-bits (4096 levels) compared to
  8-bits (256 levels)
• Spatial 250-1000 m (multiple spatial
  resolutions)
• Wavebands High-spectral (100s of bands) vs.
  multispectral (4-7 bands)
• Temporal near continuous (daily, 8-day
  composites).
• ADVANCES in DATA PROCESSING and in providing DATA
  PRODUCTS
• Science quality data (e.g., normalizations for
  acquisition dates, surface reflectance)
• Guaranteed data in regular time intervals (e.g.,
  product delivered for entire globe)
• ADVANCES in Data Accessibility
• free downloads over internet

12
Advances in Mapping LULC and Irrigated Areas
BAND-WIDTH Broad-bands ETM vs. Narrow-bands
MODIS
13
Advances in Mapping LULC and Irrigated Areas
Spatial-Resolution 4-m vs. 500-m
IKONOS 4-m NDVI
MODIS 500-m NDVI
14
Advances in Mapping LULC and Irrigated Areas
TEMPORAL RESOLUTION Near-continuous frequency
MODIS data
Rabi, 2001
Summer, 2001
Khariff, 2001
Khariff, 2002
Rabi, 2002
Summer, 2002
15
Advances in METHODS, TECHNIQUES, and
APPROACHES SecondMETHODS
16
Mega files and their role in advancing LULC and
Irrigated Areas
A. 100s or even 1000s of wavebands B. Near
continuous
Building single mega files is a must for
time-series continuous analysis
NDVI
And so on100s1000s of bands
17
Advances in Mapping LULC and Irrigated
Areas Tassel caps for single dates (TC-SD)
18
Tassel caps for single dates (TC-SD), class
identification, and labeling
Barren Type 4
Mixed grasslands (floodplain)/ Irrigated
crops (moist)
Soil line
Spectral properties of classes or pixels analyzed
based on their distribution in brightness-greennes
s-wetness (BGW) 2-dimensional feature space
(2d-FS) Purest classes occupy the BWG
end-members, All other classes form some
linear or non-linear combination of the 3 purest
classes.
Mixed irrigated crops/ riparian vegetation
Crop type 5
Crop type 4
Crop type 3
Barren Type 5
Mixed Natural Veg. (open)/dry rain fed ag.
Mixed Barren/rain fed crop
Crop type 7
Natural Vegetation (floodplain)
Barren Type 3
Crop type 6
Mixed Barren/Irrigated crop
Mixed Natural Veg. / crops
Agriculture (floodplains)
Barren Type 2
Mixed barren/ fallow crops
Mixed open forest/ crops
Mixed Natural veg. (open)/ supplemental ag.
Mixed Forest/ sugarcane rice
Mixed Riparian vegetation (moist),
wetlands/built-up
Wetlands
Barren Type 1
Crop type 1
Mixed Natural Veg. / Irrigated crops
Forest Type 5
Seasonal Snow Type 1
Mixed Rangelands open areas/ rain fed crops
Forest Type 4
Mixed water / barren land
Forest Type 3
Barren Type 6
TC-SD provides Opportunity to study spectral
properties of a class during each date (e.g.,
each of the 42 dates in 2001-2002 for
Ganges) ..TC-SD for 1 of 42 shown in the
plot here.
Forest Type 2
Water Type 2
Forest Type 1
Water Type 1
Forest Type 6
Barren Type 7 Moist Soils
Water type 3
Barren Type 8 Very Bright Soils
Snow Type 3
Snow type 1
Snow type 2
19
Tassel cap for multiple dates (TC-MD)
Soil Line
Juxtaposition of spectral classes from several
TC-SDs in a single plot Track changes in
magnitude and direction of spectral classes in
2-d FS
20
Advances in Mapping LULC and Irrigated Areas
Space-time spiral curves (ST-SC)
ST-SCs quantifies, visualizes, and tracks LULC
class changes in near-continuous mode.
ST-SCs Detects, and maps in 2-d FS subtle and
not-so-subtle changes continuously over time and
space.
21
class signatures based on NDVI (CS-NDVI) and New
Information Generation
Unique class signatures, based on NDVI (CS-NDVI)
and their intra- and inter-seasonal and intra-
and inter-year characteristics.
New Information Generation Examples Beginning of
cropping season Peak of cropping season
duration of cropping season dormancy
Land Use/Land Cover (LULC) and Irrigated Area
Mapping using Continuous Streams of MODIS Data
IWMI-data storehouse pathway http//www.iwmi.org/
rs-gis-center/dsp
22
Advances in Mapping LULC and Irrigated Areas
class signatures based on NDVI (CS-NDVI)
Duration of Khariff in 2001 160 days
Khariff in 2001 end November 15
Khariff in 2001 begin June 11
Peak of Khariff in 2001116 days
The threshold NDVIs and NDVI signatures over time
help us determine 1. Onset of a cropping
seasons (e.g., Rabi and khariff) 2. Duration
of the cropping seasons such as for khariff
and Rabi 3. Magnitude of the crops during
different seasons and years (e.g., drought vs.
normal years) 4. End of cropping season
(senescence)
23
Advances in Mapping LULC and Irrigated Areas
Drought vs. Normal year
Irrigated class 21
Rainfed class 18
24
Advances in Mapping LULC and Irrigated Areas
Fuzzy Classification Accuracy Assessment
fuzzy principal absolutely correct, mostly
correct, correct, incorrect, mostly incorrect,
absolutely incorrect) Actual LULC data
available from 3 by 3 pixel area gathered during
ground truthing, were compared with actual
classes mapped. Correct pixels Fuzzy accuracy
class 100 (9/9 pixels) absolutely
correct 75-99 mostly correct 51-74
correct 40-50 incorrect 25-50 mostly
incorrect lt25 absolutely incorrect
25
Mitch 25 minutes on specifics (unsupervised
100 classes, intitial class identification
process, ground truth database, 100 to 30
classes, class signatures, class characteristics
table-area, cover percentage, classification
accuracies, how could we have done better)
26
DATA PREPARATION
27
Advances in Mapping LULC and Irrigated Areas
DATASET PREPARATION Ganges Mega files

• I. Mega-file of reflectance A total of 294
  bands (42 images 7 bands)
• 21 images each of 7 bands from year 2001
• 21 images each of 7 bands from year 2002
• a single mega file of approximately 7 GB.
• II. Mega-file of NDVI A total of 42 bands (one
  NDVI band for each date)
• 21 images for 2001
• 21 images for 2002
• a single mega file of approximately 1.7 GB.

28
Advances in Mapping LULC and Irrigated Areas
CLOUD REMOVAL 2 ALGORITHMS

• Blue band minimum reflectivity threshold
• If (blue band gt 21 reflectance) then null
  else I

B. Visible band minimum reflectivity
threshold If (blue band gt 22
reflectance and green band gt 21 reflectance and
red band gt 23 reflectance) then null else I
Results of the Second Algorithm
Before cloud Algorithm
Before cloud Algorithm
After cloud Algorithm
After cloud Algorithm
29
Advances in Mapping LULC and Irrigated
Areas Tassel caps for single dates (TC-SDs),
Class identification, and Labeling
Classes were labeled based on 1. Geocover 2.
NDVI dynamics 3. Image interpretations (e.g.,
tassel cap) We took these preliminary labeling
to the field and verified/updatedduring GT
Similar tassel caps are possible for each of the
42 dates.we plotted many dates
30
GROUND TRUTH
31
GT Mission October 2-22, 2003
MODIS FCC of April, 2002. (RGB) 2,1,6 in
background.
196 locations, 8000 kms

• Land use/land cover (LULC) classes
• Land cover types (percentage)
• Crop types (e.g.. Rice, sugar cane..)
• Cropping pattern
• Cropping calendar
• Irrigated, rainfed, supplemental irrigation
• 311 Digital photos hot linked _at_ 196 locations

32
Advances in Mapping LULC and Irrigated Areas
Ground Truth Data and Methods

• Stratified random sampling approach stratified
  by road-network, randomized by locating site
  every few minutes or few kilometers
• Representativeness key factor homogeneous
  pixel for a class or a
• representative pixel of a class.

33
Advances in Mapping LULC and Irrigated Areas
Sample Site Locations
Forests
Irrigated
MODIS NDVI image. Red High NDVI, Blue lowest
NDVI, Yellow moderate NDVI
34
RESULTS
35
Advances in Mapping LULC and Irrigated
Areas Tassel caps for multiple dates (TC-MDs)
Jan 1, 2001
May 9, 2001
Sept. 6, 2001
36
Space-time spiral curves (ST-SC)
e.g., Snow-story within and across years
Melting snow due to rain vegetation growth
Inter-mingling in Day 153, mid-Monsoon
Bright snow, high reflectance in Day 57, Summer
Wet-bright snow, snow accumulation, more wet in
day 345, winterSummer
Images illustrated for MODIS 2001 in FCC (RGB)
1,4,3 (red,green,blue)
37
Class signatures depicted in MODIS NDVI
38
Advances in Mapping LULC and Irrigated Areas
FINAL 29 CLASSES
39
Signatures of 6 irrigated area classes in the
Study Area
When conditions are wet, the NDVI threshold is
generally lower for same biomass levels
Rice in Ganges has substantially higher biomass
when comaped with rice in Indus in terms of MODIS
NDVI
40
CONCLUSIONS
41
Advances in Mapping LULC and Irrigated Areas
CONCLUSIONS
Cultivated Irrigated Canal Tube
well supplemental Total
Irri. () () () () ()
() as of total irrigated area as
of total area Ganges 95 52.9 12.9 58.
5 28.5 41.7 (90 Mha) (37.6
Mha) 24 Mha rice in
khariff Ganges Indus 63 46.6 28.5 53.5
27.2 34.1 (133 Mha) (45.4
Mha) GangesIndus basin area 211 Mha
All 6 irrigated areas in dark green
42
CONCLUSIONS Ganges River Basin Irrigated Areas
from Different Studies
Ganges-Classification Irrigated area
deviation from this study USGS (1992-1993)
32,255,630 (ha) 16.6 () less than ours
GLC (2000) 56,466,954 (ha)
33.4 () more than ours IWMI (2001-2002)
37,602,567 (ha) 0
Change in a decade is 16.6 additional area.
USGS (1993)
Just look in Ganges area
GLC has 2 irrigated area classes_class 32 and 33.
Class 33 in actual is a mix of rainfed and
supplemental.
IWMI (2004)
GLC (2000)
43
CONCLUSIONS QUANTITATIVE FUZZY CLASSIFICATION
ACCURACY ASSESSMENT (QFCAA)
Accuracy Range Overall accuracy 17
classes between 80 to 100 () Of 29 classes 6
classes between 70 to 80 () 6 classes
between 56 to 70 () Irrigated classes 5 of 6
classes between 75-100 () 6 irrigated classes 1
class 56 percent Rainfed 69-88 () Forest
63-75 ()
Note Most classes intermixed within the theme
(e.g., forests within forest classes, irrigated
within irrigated classes..so when
accuracies are low for some classes it is not a
huge issue.