The Kansas Next-Generation Land. Use/Land Cover Mapping Initiative ... Used existing databases as reference or ground-truth to assess accuracy levels for ...
The Kansas Next-Generation Land Use/Land Cover Mapping Initiative Kansas Biological Survey Kansas Applied Remote Sensing Program April 2010 Dana Peterson
Project Summary
• NSF EPSCoR Project entitled, “Understanding and
Forecasting Ecological Change: Causes, Trajectories and Consequences of Environmental Change in the Central Plains”
• An update of a statewide LULC database for the Kansas GIS Poly Board – Previous map produced in 1990
Project Summary • Two Phases of LULC Mapping – Phase I: Produce a “general” (Modified Level I) LULC to match 1990’s LULC database – Phase II: Produce a Level II, Level III, and Level IV LULC • Crop type • Grassland type • Irrigation status
Project Summary
• A Hybrid-Hierarchical Classification Approach – Level I mapping using an unsupervised classification (ISODATA) – Level II and Level III mapping using the Level I map as a mask and a supervised classification approach (Decision Tree) – Level IV mapping (irrigation status) using the Level I map as a mask and an NDVI threshold
Phase I: Level I Land Cover Mapping Modified Level I LULC Mapping – Eleven Classes Mapped • Woodland, water, cropland, grassland, CRP, and urban commercial/industrial, urban residential, urban openland, urban water, urban woodland, and other
– Multitemporal Landsat TM (30-meter resolution) • Imagery from existing 2004-05 KSID database and 12 additional scenes purchased by EPSCoR
– Comparable to 1990 LULC map (classification approach, classes & MMU) • Allows change detection
Phase I Land Cover Classification Methodology Image Classification Grass/Crop
Using MMUs Using CLUs
Two Stage Generalization
Urban Water Woodland
Multitemporal Landsat TM data
Image Acquisition & Processing
Mosaic Maps
Accuracy Assessment
Two Stage Map Generalization
Pre-generalized
Stage I: Using MMUs
Stage 2: Zonal Majority in CLUs
Phase I Accuracy Assessment • Stratified random sampling design – Sample size proportionate to the area mapped for each land cover class – More than 31,000 sample sites used – Sample unit = field polygon
• Used existing databases as reference or ground-truth to assess accuracy levels for cropland, grassland, and woodland – Kansas GAP database – USDA database
• Used aerial imagery (NAIP) interpretation techniques to assess accuracy levels for water and urban • Target overall accuracy: > 85%; Achieved accuracy:
90.72%
Phase I: User and Producer’s Accuracy Levels
LULC Class
LULC Code
User Accuracy (%)
Producer Accuracy (%)
Urban Commercial Industrial
11
61.05
74.36
Urban Residential
12
48.35
77.19
Urban Openland
13
78.43
64.17
Cropland
20
90.92
93.37
Grassland
30
91.23
88.58
CRP
31
NA
NA
Woodland (rural and urban)
14 & 40
95.77
80.68
Water (rural and urban)
15 & 50
95.81
92.93
60
NA
NA
Other
2005 Level I Land Use/Land Cover Map
Phase I: Mapping Difficulties
• Misclassifying grassland as cropland
Attributed CLU on ungeneralized map
Unattributed CLU on generalized map
1990 KLCP Map Revisions
•
Converted file format from vector to raster
•
Projected map to match 2005 KLCP map projection
•
Changed coding scheme to match 2005 KLCP coding scheme
•
Removed road network that was not mapped in 2005
•
Generalized 1990s map using methods developed for 2005 map
•
Recalculated accuracy levels using updated methods
•
Revisions facilitates comparing the 1990 KLCP map to the 2005 KLCP map on a more equal basis
– USDA CLU boundaries for generalization of grassland and cropland
Land Use/Land Cover Change
Phase II: Mapping Grassland Types • Cool-season & warm-season grassland identification – 2004/2005 Multitemporal 30-meter Landsat Thematic Mapper data – KS GAP and USDA databases for classification training and validation
Comparison of MODIS and Thematic Mapper imagery
Landsat TM 30m resolution
MODIS 250m resolution
Phase II: Mapping Grasslands Pilot 2005 MODIS NDVI
Overall = 52.2%
2005 Landsat TM
Overall = 75.6%
Phase II: Mapping Grasslands
• Spatial resolution appears to be more critical than temporal resolution – Regional patterns were similar between MODIS & TM maps – Due to the size (small) and shape (irregular) of CLUs used for generalization, overall the higher spatial resolution of TM data performed better – Often the MODIS IFOV was too coarse to capture small fragmented grassland types in the region (especially coolseason) – Small sample size for training and validation using MODIS data • MODIS: – Cool-season: n = 90 – Warm-season: n = 202 • Landsat TM: – Cool-season: n = 1,731 – Warm-season: n = 800
Phase II: Mapping Grasslands Using a Decision Tree Classifier Extract Grassland Pixels
Mosaic & Generalization Supervised Classification •80/20% data split for model training & validation • Independent sample used for formal accuracy assessment
Process Training Data
Reassign Grassland Pixels to Subclasses
Level II Grassland Map
Phase II: Level II Mapping Crop Types Five Cropland Classes Mapped:
– Spring Crops (Small Grains) Summer Crops (Row Crops), Alfalfa, Fallow, and Double-crop
Data Used: – 2005 MODIS NDVI Time-Series – USDA database for classification training & validation
Alfalfa Fallow Summer Crops Winter Wheat
Dec. 19
Dec. 3
Nov. 17
Nov. 1
Oct. 16
Sept. 30
Sept. 14
August 29
August 13
July 28
July 12
June 26
June 10
May 25
May 9
April 23
April 7
March 22
March 6
Feb. 18
Feb. 2
Jan. 17
Jan. 1
NDVI
General Crop Types
Average multi-temporal NDVI profiles for Kansas in 2001
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Phase II: Mapping Crop Types Using a Decision Tree Classifier Extract Cropland Pixels
Process Training Data
Wardlow & Egbert, 2008
Supervised Classification •80/20% data split for model training & validation • Independent sample used for formal accuracy assessment
Resample & Generalization
Reassign Crop Pixels to Crop Subclasses
Phase II: Stratifying the Classification USDA Crop Reporting Districts
Level II Cropland Map
Phase II: Level III Cropland Mapping Three Additional Crop Classes Mapped: – Mapped Summer Row Crop into subclasses of Corn, Soybean and Sorghum
Data Used: – 2005 MODIS NDVI Time-Series – USDA database for classification training & validation
Level III Crop Type Map
Phase II: Mapping Irrigation Status (Level IV)
• Irrigation status: Unsupervised Approach – 2005 Multitemporal MODIS NDVI data (23 composites) – USDA CLU database to summarize NDVI and for classification validation – USDA County Level statistics as model restraint or threshold
Phase II: Mapping Irrigation Status Irrigated lands have a higher peak NDVI than nonirrigated lands 0.9
0.8
0.7
0.5
0.4
0.3
0.2
0.1
Corn (Irrigated)
Corn (Non-Irrigated)
Winter Wheat (Irrigated)
Dec. 19
Dec. 3
Nov. 17
Nov. 1
Oct. 16
Sept. 30
Sept. 14
August 29
August 13
July 28
July 12
June 26
June 10
May 25
May 9
April 23
April 7
March 22
March 6
Feb. 18
Feb. 2
Jan. 17
0
Jan. 1
NDVI
0.6
Winter Wheat (Non-Irrigated)
Phase II: Mapping Irrigation Status (Level IV) Extract peak NDVI Resample to 30-meter
Extract Cropland & Calc Zonal AVG of peak NDVI
NDVI >= Threshold
=
Irrigated
Identify peak NDVI thresholds
USDA County Irrigation Statistics Adapted from Brown et al.,2007
Phase II: Mapping Irrigation Status (Level IV)
County Code
USDA Reported Wheat Acreage
NDVI Value 229
NDVI Value 228
NDVI Value 227
NDVI Value 226
NDVI Value 225
NDVI Value 224
NDVI Value 225
20151
14,400.0
939
2,505
3,175
4,209
5,621
7,508
8,016
20153
3,700.0
12
12
12
18
668
670
952
20155
6,500.0
1,127
1,729
2,800
3,957
5,230
5,807
6,953
20157
22.5
11
15
100
159
186
244
339
20159
3,945.8
1,375
1,739
2,079
2,742
4,041
4,557
5,289
Irrigation Status Map
Accuracy Assessment
• Level IV Map: – Used the same random, stratified sampling approach
Mapping Level
Overall Accuracy
– Weighted by area mapped
Level II Map
86.2%
– 16,000+ sample sites
Level III Map
82.5%
– Aggregated Error Matrix to derive Level III and Level II accuracy levels
Level IV Map
76.5%
Data Availability/Distribution • KARS www.kars.ku.edu or DASC websites www.kansasgis.org (.zip of raster data file) • KARS GeoNetwork Data Portal (downloadable .zip or .kmz) • KARS REST Services (Web Services) • Web Mapping Application
Data Availability