The MODIS Land Cover and Land Cover
Dynamics Products
A.H. Strahler (PI), Mark Friedl, Xiaoyang Zhang, John Hodges,
Crystal Schaaf, Amanda Cooper, and Alessandro Baccini
http://geography.bu.edu/landcover/
Center for Remote Sensing and Dept. of Geography
Boston University
1
MODIS Land Cover: Five Sets of
Labels
• IGBP:International Geosphere-
IGBP Biosphere Project labels
– 17 classes of vegetation life-
UMD LAI/FPAR BGC
form
• UMD: University of Maryland
land cover class labels
Plant Functional Types
– 14 classes without mosaic
classes
• LAI/FPAR: Classes for
• Plant Functional Types LAI/FPAR Production
(Future) – 6 labels including broadleaf
– Plant functional types to be and cereal crops
used with the community • BGC: Biome BGC Model
land model (NCAR, Bonan) Classes
– Exact classes TBD – 6 labels: leaf type, leaf
longevity, plant persistence
2
MODIS Land Cover: Where Does it Come
From?
• MODIS Data
– 16-day Nadir BRDF-Adjusted Reflectances
(NBARs) assembled over one year of
observations
– 7 spectral bands, 0.4–2.4 m, similar to
Landsat
– 16-day Enhanced Vegetation Index (EVI)
• Training Data
– >1,500 training sites delineated from high
resolution satellite imagery (largely Landsat)
• Classifier
– Uses decision tree classifier with boosting
3
MODIS Land Bands
Band Spatial Resolution Wavelength, nm
1 250 m 654–664
2 250 m 860–870
3 500 m 465–475
4 500 m 550–560
5 500 m 1234–1246
6 500 m 1632–1648
7 500 m 2120–2140
4
MODIS Geolocation
• Geolocation accuracy specification is 300 m (2 )
and
goal is 100 m (2 ) at nadir
• Geolocation goal driven by Land 250 m change
product requirements
• Goal is currently being met
Land: 550 CPs
from 126 TM Scenes
Ocean: 4600 island
points from SeaWifs
library
Ground Control Points—Land 5
MODIS Data Levels
• Level 1
– Radiometrically corrected, geolocated radiances
• Level 2
– Products derived from Level 1 data without geometric
resampling
• Level 2G (MODIS Land)
– Forward-binned into integerized sinusoidal projection
without resampling
• Level 3
– Products resampled using geolocation information to a
standard family of map grids; often multitemporal or
composited
• Level 4
– Products derived from multiple data sources by
modeling 6
IGBP Land Cover Units (17)
• Natural Vegetation • Developed and Mosaic
(11) Lands (3)
– Evergreen Needleleaf – Croplands
Forests – Urban and Built-Up
– Evergreen Broadleaf Lands
Forests – Cropland/Natural
– Deciduous Needleleaf Vegetation Mosaics
Forests
– Deciduous Broadleaf • Nonvegetated Lands
Forests
(3)
– Mixed Forests
– Snow and Ice
– Closed Shrublands
– Barren
– Open Shrublands
– Water Bodies
– Woody Savannas
– Savannas
– Grasslands
– Permanent Wetlands 7
The Land Cover Input Database
• 242 Features From MODIS:
– Temporal and spectral information; 16-day composites
• Uses Surface Reflectance (NBAR)
– View-angle corrected surface reflectance, 7 land bands
• And Enhanced Vegetation Index (EVI)
• Plus (in the future)….
– Spatial Texture from 250-m Band 2
• Standard deviation-to-mean ratio in Band 2 (near-infrared)
– Snow Cover
• MODIS Snow Cover Product, number of days with snow cover
– Land Surface Temperature
• MODIS Land Surface Temperature, maximum value composite
– Directional Information
• Bidirectional reflectance information from BRDF product
8
Global Composite Map of Nadir BRDF-Adjusted Reflectance (NBAR)
April 7–22 2001
no
datadata
No True color, MODIS Bands 2, 4, 3
10 km resolution, Hammer-Aitoff projection,
produced by MODIS BRDF/Albedo Team
MODLAND/Strahler et al. 9
MODIS Nadir BRDF-Adjusted Reflectance
May 25–June 9 2001
False Color Image
NIR–Red–Green
10
NBAR Time Trajectories
11
MODIS 500 m Vegetation Indices
(September 30 –
October 15, 2000
NDVI
MOD13A1 16 day
Composite
EVI
MODLAND/Huete et al 12
NDVI EVI
EVI shows better dynamic range, less saturation
13
Advanced Technology Classifiers
• Supervised Mode
– Use of supervised mode with training sites
– Allows rapid reclassifications for tuning
• Decision Trees—C4.5 Univariate Decision Tree
– Fast algorithm
– Uses boosting to create multiple trees and improve
accuracy, estimate confidence
• Neural Networks—Fuzzy ARTMAP
– Uses Adaptive Resonance Theory in building network
– Presently not in use. Too slow; does not handle missing
data well.
14
Decision Tree Classification
• Goal:
– Optimal prediction of class labels
from a set of feature values
• Basic Approach
– Supervised learning using training
data
• Key Attributes:
– Nonparametric
– Able to handle noisy or missing
features
– Adept at capturing nonlinear,
hierarchical patterns
15
DTs: Basic Theory
• Tree Structure
– Root node (all data), internal
nodes and terminal or leaf nodes Root
(predictions)
• Building the Decision Tree Internal nodes
– Recursive partitioning of training
data into successively more
homogeneous subsets
• Multiple Leaf Nodes per Class
– Leaf nodes identify class
assignment
– Sub-classes allocated individual
leaves
Leaf nodes16
Postclassification Processing
• Application of Prior Probabilities
– Use of priors to remove training site count biases
(sample equalization)
– Application of global and moving-window priors from
earlier products
• Increases accuracies, reduces speckle
– Use of external maps of prior probabilities to resolve
confusions
• Agriculture/natural vegetation confusion in some
regions
• Use of city lights DMSP data to enhance urban class
accuracy (to come)
• Filling of Cloud-Covered Pixels from Earlier Maps
– Use of at-launch (EDC DISCover v. 2) or provisional
product when there are not sufficient values to classify
a pixel with confidence
17
Using Priors to Classify Cereal and Broadleaf
Crops
Cereal Crop Intensity from USDA Statistics Broadleaf Crop Intensity from USDA Statistics
MODIS Map of Cereal Crops in the Continental United States MODIS Map of Broadleaf Crops in Continental United States
18
Provisional Land Cover Product June 01
MODIS data
from Jul 00–
Jan 01
19
Consistent Year Land Cover Product June 02—
IGBP
MODIS data
from Nov 00–
Oct 01
20
Consistent Year Land Cover Product, Nov 00–Oct 01
Mixed Forest
Evergreen
Needleleaf
Forest
Cropland/Natural
Vegetation Mosaic
Cropland
Deciduous Urban
Broadleaf
Forest
21
Classification Second Most-Likely
Confidence Map Class
Lower
Second choice
Confidence
omitted
with very high
High
confidence level
Confidence
22
Rondonia Comparison
• Note better
delineation
of land
cover
pattern
Consistent Year Confidence
EDC DISCover v.2 Provisional Product
23
Consistent Year Confidence
EDC DISCover v.2 Provisional Product
24
Consistent Year Confidence
EDC DISCover v.2 Provisional Product
25
Land Cover Validation
• Validation Plan Utilizes Multiple Approaches
• Level 1: Comparisons with existing data sources
– Examples
• Global AVHRR land cover datasets: DISCover, UMd
• Humid Tropics: Landsat Pathfinder
• Forest Cover: FAO Forest Resources Assessment
• Western Europe: CORINE
• United States: USGS/EPA MLRC
• United States: California Timber Maps (McIver and
Woodcock)
• MODIS and Bigfoot test site comparisons
26
Validation Levels, Cont.
• Level 2: Quantitative studies of output and training
data
– Per-pixel confidence statistics
• Aggregated by land cover type and region
• Describe the accuracy of the classification process
– Test site cross-comparisons
• Confusion matrices globally and by region
• Provides estimates of errors of omission and
commission
• Level 3: Sample-based statistical studies
– Random stratified sampling according to proper statistical
principles
– Costly, but needed for making proper accuracy statements
• CEOS Cal-Val Land Product Validation Land Cover
27
Confidence Values by Land Cover Type
(Preliminary)
IGBP Class IGBP Class
Confidence Confidence
1 Evergreen Needleleaf 9 Savanna 67.8
68.3 10 Grasslands 70.6
2 Evergreen Broadleaf 11 Permanent Wetlands
89.3 52.3
3 Deciduous Needleleaf 12 Cropland 76.4
66.7 14 Cropland/Nat. Veg’n.
4 Deciduous Broadleaf 60.7
65.9 15 Snow and Ice
5 Mixed Forest 87.2
65.4 16 Barren 90.0
6 Closed Shrubland 60.0 Overall Confidence 76.3
7 Open Shrubland 75.3
Includes adjustment for prior probabilities. Urban and Built-Up (13), Water(17) classes omitted. Pixels filled
from prior data omitted. Based on preliminary data, subject to change.
8 Woody Savanna 64.0
28
Confidence Values by Continental Region
(Preliminary)
Region Confidence,
percent
Africa 79.4
Australia/Pacific
83.2
Eurasia 76.8
North America 71.9
South America 78.5
Overall Confidence 76.3
Includes adjustment for prior probabilities. Urban and Built-Up (13), Water(17) classes omitted. Pixels filled
from prior data omitted. Based on preliminary data, subject to change.
29
Cross Validation with Training Sites
• Cross-Validation Procedure
– Hide 10 percent of training sites, classify with
remaining 90 percent; repeat ten times for ten
unique sets of all sites
– Provides ―confusion matrix‖ based on unseen
pixels where whole training site is unseen
– Not a stratified random sample, but a
reasonable indication of within-class accuracy
30
Confusion Matrix (Preliminary)
Global Test Site Confusion Matrix—Consistent Year Product,
After Priors
Site Class Classification Outcome
Class Name 1 2 3 4 5 6 7 8 9 10 11 12 14 15 16 Total
1 Ev e rgr ee n Ne e dle le af 14 60 42 18 11 26 6 7 9 17 23 10 15 21 2 0 0 19 01
2 Ev e rgr ee n Broadle af 31 48 89 0 14 14 11 18 79 23 17 4 38 10 0 1 51 49
3 De ciduous Ne e dle le af 87 0 10 4 25 11 8 0 0 4 0 0 0 10 0 0 0 34 8
4 De ciduous Broadle af 22 56 16 38 4 27 8 0 3 11 1 3 0 47 82 0 0 90 3
5 M ixe d Fore st 40 5 63 94 14 8 13 55 3 1 27 7 8 40 41 17 0 0 22 09
6 Close d Shrubland 34 35 2 12 5 14 0 12 4 29 15 30 2 15 8 19 0 8 61 3
7 Ope n Shrubland 10 12 3 9 1 41 10 02 33 45 20 3 0 21 0 6 0 21 3 17 88
8 Woody Sav anna 62 13 3 0 16 11 0 11 10 4 57 7 14 1 71 0 22 1 22 0 3 14 71
9 Sav anna 10 53 1 0 21 18 48 93 44 0 43 1 25 2 79 0 16 10 75
10 Grasslands 2 16 0 2 20 4 17 9 6 10 1 63 2 0 24 9 13 0 36 3 15 87
11 Pmnt Wtlnd 63 24 0 5 28 23 1 2 36 2 89 1 7 0 0 28 1
12 Cropland 6 75 2 7 16 8 61 42 13 2 13 3 2 51 68 18 3 0 18 58 53
14 Cropland/Natural Ve gn 2 13 3 0 48 28 2 8 16 66 8 1 32 0 83 2 0 7 14 71
15 Snow+ice 1 0 0 0 0 1 2 0 0 0 5 1 0 12 97 5 13 12
16 Barr e n 0 2 1 0 0 1 16 2 4 5 12 6 3 56 5 14 35 37 39 16
Total 21 95 55 33 24 1 68 1 22 60 27 0 17 22 94 0 10 35 12 86 16 2 67 93 12 77 13 11 41 71 29 877
31
Accuracies—Consistent Year Product
(Preliminary)
Based on Global Test Site Confusion Matrix
Dataset Training Site
Accuracy
Before priors 78.6 %
After priors 71.0 %
After priors, first two classes 84.0 %
32
Overall Accuracies
• Proper accuracy statements require
proper statistical sampling
• AVHRR state of the art has been 60–70
percent, depending on class and region
• MODIS accuracies are falling in 70–80
percent range
• Most ―mistakes‖ are between similar
classes
• Land cover change should NOT
be inferred from comparing
successive land cover maps 33
Land Cover Dynamics
• Primary Objectives:
– Quantify interannual change
• Uses change vectors comparing successive years
• Identifies regions of short-term climate variation
• Under development with Eric Lambin, Frederic Lupo
at UCL, Belgium
– Quantify phenology
• Greenup, maturity, senescence, dormancy
• Values of VI, EVI at greenup and peak, plus annual
integrated values
• Uses logistic functions fit to time trajectories of EVI
34
Land Cover Dynamics:
Defining Phenological Attributes
35
QuickTime™ and a
Video d ecompressor
are neede d to see this picture.
36
QuickTime™ and a
Video d ecompressor
are neede d to see this picture.
37
Web Site: http://geography.bu.edu/landcover
38
References
• Friedl, M.A., D. Muchoney, D.K. McIver, A.H. Strahler, and J.C.F. Hodges 2000:
Characterization of North American land cover from AVHRR Data, Geophysical
Research Letters, vol. 27, no. 7, pp. 977-980.
• Friedl, M.A., C. Woodcock, S. Gopal, D. Muchoney, A.H.Strahler, and C. Barker-
Schaaf 2000. A note on procedures used for accuracy assessment in land cover
maps derived from AVHRR data, International Journal of Remote Sensing, vol. 21,
pp.1073-1077.
• Friedl, M.A., Brodley, C.E. and A.H. Strahler 1999: Maximizing land cover
classification accuracies at continental to global scales, IEEE Transactions on
Geoscience and Remote Sensing, vol. 37, pp. 969-977.
• Friedl, M.A. and C.E. Brodley 1997: Decision tree classification of land cover from
remotely sensed data, Remote Sensing of Environment, vol. 61, pp. 399-409.
• Mciver, D.K. and M.A. Friedl 2002. Using prior probabilities in decision-tree
classification of remotely sensed data, Remote Sensing of Environment, Vol. 81, pp.
253-261.
• McIver, D.K. and M.A. Friedl 2001. Estimating pixel-scale land cover classification
confidence using non-parametric machine learning methods, IEEE Transactions on
Geoscience and Remote Sensing. Vol 39(9), pp. 1959-1968.
• Muchoney, D., Borak, J, Chi, H., Friedl, M.A., Hodges, J. Morrow, N. and A.H.
Strahler 1999: Application of the MODIS global supervised classification model to
vegetation and land cover mapping of Central America, International Journal of
Remote Sensing, Vol 21, no 6 & 7, pp. 1115-1138.
• Muchoney, D. M., and Strahler, A. H., 2001, Pixel and site-based calibration and
validation methods for evaluating supervised classification of remotely sensed
data, Remote Sens. Environ., in press.
39