A Comparison of Pond Inventories Using Satellite and Airborne SensorsA Comparison of Pond Inventories Using Satellite and Airborne Sensors

Artificial ponds exist throughout the Kansas landscape, far outnumbering natural water bodies, and they play a substantial role in modifying the environment. For example, they trap sediment, thereby affecting biogeochemical cycles, and they also provide habitat diversity and may provide a partial counterbalance to lost wetlands. For a number of reasons, including their small size, their location primarily on private property, and variations in their numbers and locations over time, small artificial ponds are often underrepresented on the digital map products and databases normally used for hydrologic analyses. To address the issue of the underestimation of ponds, images from three different satellite and airborne sensors were used to see how accurately they could locate and inventory ponds in a study area Jefferson county. Landsat Enhanced Thematic Mapper (ETM+) 30m multispectral imagery, Terra ASTER 15m multispectral imagery, and 1m multispectral imagery from an airborne digital camera were used to create maps of water impoundments. For each study area, we computed the number of water bodies, their size classes, and the total water surface area. Based on our assumption that the maps derived from the 1m airborne digital imagery would provide the most detailed and accurate estimate of the actual number of ponds in the study areas, we used them as the basis for comparison with the maps derived from Landsat and ASTER imagery. Since it is generally impractical (due to cost and time considerations) to manually map small ponds from detailed imagery, our objective was to determine by how much the number of ponds in the Kansas landscape is underestimated using satellite imagery. In addition to comparing results of the digital airborne camera inventory to maps from the two satellite sensors, we also compared them to two inventories of water bodies that were previously created. The most recent is the Kansas Surface Water Database (KSWD), which was derived from 2000 and 2001Landsat ETM+ imagery at a minimum mapping unit of 1.5 acres and became available for use in 2003. The second inventory of water bodies is the Surface Waters Information Management System (SWIMS). This database was created using the Environmental Protection Agency’s (EPA) River Reach Files (RF3). The RF3 files were developed from 1:500,000-scale NOAA aeronautical charts and 1:100,000-scale digital line graphs developed by USGS.

Abstract

ASTERThe ASTER image (August 6, 2001) was processed using an unsupervised classification

procedure in ERDAS Imagine. Using the ISODATA clustering algorithm, 100 spectral clusters were defined. The clusters that represented water were then combined into a ‘Water’ class and the remaining classes were combined into a class called ‘Non-Water.’ The result was a raster data set with two classes: water and non-water, that was then brought into ArcMAP and converted to a polygon shapefile. Using the Editor extension, all polygons were visually confirmed to represent actual water bodies. If a polygon did not represent a water body (typically edge polygons), it was deleted. The result was a vector-format estimate of the water bodies. The reason for converting from raster to vector format was to be able to calculate the surface area of each polygon. To facilitate extracting surface area, a tool was developed using ArcObjects to extract each polygon area from the “shape” field within the shapefile.

 Landsat Enhanced Thematic Mapper (ETM+)

The ETM+ image (July 21, 2001) was processed in the same manner as the ASTER image, first using an unsupervised classification procedure in ERDAS Imagine. Using the ISODATA clustering algorithm, 100 spectral clusters were defined. The clusters that represented water were then combined into a ‘Water’ class and the remaining classes were combined into a class called ‘Non-Water.’ The result was a raster data set with two classes: water and non-water, that was then brought into ArcMAP and converted to a polygon shapefile. Using the Editor extension, all polygons were visually confirmed to represent actual water bodies. If a polygon did not represent a water body (typically edge polygons), it was deleted. The result was a vector-format estimate of the water bodies.

DuncanTech Digital Aerial ImageryForty-four scenes from three different dates (12 April 2003, 9 May 2003, and 9 June 2003)

were mosaicked together using ERDAS Imagine. All water bodies were then digitized into a vector layer using standard heads-up digitizing procedures. The resulting vector layer was then saved as a polygon shapefile, which was then brought into ArcMap for calculation of the number of water bodies and their surface areas. In addition a polygon layer was created that represented the extent of all the 44 DuncanTech images. This layer constituted the extent of the study sites within the study area and was used to clip all other map layers.

 Kansas Surface Water Database (KSWD)

The KSWD was clipped to the extent of the 44 Duncan Tech images. It was converted from a raster layer to a polygon shapefile. The number of ponds and their surface area were then calculated.

 Surface Water Information Management System (SWIMS)

This dataset was downloaded from DASC in shapefile format. The polygons were clipped to the extent of the 44 DuncanTech scenes and the resulting shapefile was added to ArcMap, where the number of ponds and surface area were calculated.

 

Study AreaThe Jefferson County landscape, with an annual precipitation of 35 inches per year, is dotted with small water bodies containing only a few acre/feet of water to large water bodies such as Perry Lake and is typical of the northeast Kansas landscape. The primary factor in choosing this area is the availability of rectified DuncanTech imagery with coverage from the other four data sources. The study area is covered by 44 DuncanTech images which have been mosaicked together. This imagery overlaps rectified imagery from the ASTER sensor as well as Landsat ETM+. In addition the KSWD and SWIMS databases also have full coverage.

DOQQ for study area overlaid with 44 mosaicked DuncanTech scenes.

Jefferson County

Kansas

ObjectiveThe overall objective was to determine how accurately each imagery source could locate and inventory ponds in Jefferson county . The three sources of digital visible infrared imagery (Landsat ETM+, Terra ASTER, and the Duncan Tech aerial camera) were compared. Two existing water databases (KSWD and SWIMS) were compared as well. The objectives can be summarized:

1. What is the minimum spatial resolution of digital imagery that can accurately distinguish small water bodies in Kansas?

2. How well does the classification of the digital imagery compare with the existing KSWD and SWIMS databases?

3. To provide a recommendation on what digital imagery source would be the most cost effective to use without a significant loss of accuracy.

Data Processing

Resolution Resolution DifferencesDifferences

ETM+

ASTER

Duncan Tech

15 Meter Resolution (ASTER)

30 Meter Resolution (Landsat ETM+)

1 Meter Resolution (Duncan Tech)

As expected, the number of ponds identified by each of the three multispectral sensors (ETM+, ASTER, and DuncanTech) varied directly with spatial resolution, with the greatest number of ponds being identified by the sensor with the highest spatial resolution (DuncanTech digital aerial camera). In particular, it is noteworthy that imagery from Landsat’s ETM+ sensor, which is the most widely available low-cost multispectral imagery source, successfully mapped only 60% of the actual ponds in the study sample. Based on the results, it appears likely that multispectral imagery with spatial resolution on the order of 4 meters (such as imagery from the Ikonos and Quickbird satellites) would permit mapping small ponds with sufficient accuracy without incurring the storage and processing overhead entailed in using 1-meter imagery.

 An interesting, and somewhat unexpected result was that the total

estimated surface area actually increased with poorer (i.e., coarser) spatial resolution. This is undoubtedly attributable to the large relative size of the coarser pixels and the tendency of the image processing methodology to identify mixed water pixels as belonging to the water class.

 As expected, the two surface water databases (KSWD and

SWIMS) grossly underestimated the number of water bodies, although, to be fair, neither database was designed to be an inclusive map of all water bodies. It does underscore, however, the potential danger of using databases for purposes for which they were not designed – in this case the identification and mapping of small, but environmentally important, farm ponds.

Table: Estimates of Number of Water Bodies and Total Surface Area

Table: Commission and Omission Errors in Satellite Imagery-Derived Estimates

Explanation:ASTER erroneously identified 6 non-existent ponds, but failed to identify 20 ponds that were mapped using the DuncanTech imagery•ETM+ erroneously identified 1 non-existent pond, but failed to identify 40 ponds that were mapped using the DuncanTech imagery

Results

Dataset Number WaterBodies

% of ActualNumber

Total Sfc. Area (sq.km.)

% of ActualArea

Duncan Tech

97 100% 179.9 100%

Terra ASTER

83 86% 202.0 112%

ETM+ 58 60% 231.4 128%

KSWD 3 3% 26.1 15%

SWIMS 1 1% 23.6 13%

Sensor CommissionError

OmissionError

Terra ASTER 6 20

ETM+ 1 40

Duncan Tech (DT) Digital Aerial Imagery: The 1-meter Duncan Tech digital aerial camera acquires imagery with three spectral bands:

44 scenes from three dates: 12 April 2003, 9 May 2003, and 9 June 2003

ASTER: ASTER is a multi-band sensor on board NASA’s Terra satellite. For this study, only the three 15-meter spectral bands were used:

Image date: 6 August 2001

Landsat Enhanced Thematic Mapper (ETM+): The Landsat 7 ETM+ imagery used for this project was a six-band dataset with 30-meter spatial resolution. (The thermal band was removed for this analysis because of its 60 m spatial resolution):

ImageDate: 21 July 2001

Ponds resolved by sensor type

Band 1 Blue/Green 0.45 - 0.52 μm

Band 2 Red 0.63 - 0.69 μm

Band 3 NIR 0.76 - 0.90 μm

Band 1 Green 0.52 - 0.60 μm

Band 2 Red 0.63 - 0.69 μm

Band 3 NIR 0.76 - 0.86 μm

Band 1 Blue/Green 0.45 - 0.52 μm

Band 2 Green 0.52 - 0.60 μm

Band 3 Red 0.63 - 0.69 μm

Band 4 NIR 0.76 - 0.90 μm

Band 5 Mid IR 1.55 - 1.75 μm

Band 6 Mid IR 2.08 - 2.35 μm

Histogram AnalysisHistogram Analysis

Size Distribution of Water Bodies, by Sensor (Y-axis = # of ponds; X-axis = surface area, in square meters):The histogram analysis graphically illustrates that even though the ASTER and ETM+ sensors detect a lower total number of ponds the total surface area is greater than the Duncan Tech. This is because of the relatively coarse spatial resolution (15 meters and 30 meters, respectively) of those sensors.

DuncanTech (1 m)

AREA

15

25

0

14

25

0

13

25

0

12

25

0

112

50

10

25

0

92

50

82

50

72

50

62

50

52

50

42

50

32

50

22

50

12

50

25

0

40

30

20

10

0

Std. Dev = 2634.15

Mean = 1855

N = 97.00

ASTER (15 m)

AREA

15250

14250

13250

12250

11250

10250

9250

8250

7250

6250

5250

4250

3250

2250

1250

250

40

30

20

10

0

Std. Dev = 3024.32

Mean = 2434

N = 83.00

ETM+ (30 m)

AREA

15250

14250

13250

12250

11250

10250

9250

8250

7250

6250

5250

4250

3250

2250

1250

250

40

30

20

10

0

Std. Dev = 3743.00

Mean = 3990

N = 58.00

S.L. Egbert, B.N. Mosiman, and P. Taylor. 2003. Creating a Pond Inventory in Kansas Using Satellite and Airborne Sensors, Water and the Future of Kansas 21st Annual Conference. Lawrence, Kansas. March 11, 2004.