526
September 2020
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
SECTOR
INSIGHT:
.
edu
have several stages of regression estimation using
different spatial resolution sensors (Koeln and
Kollasch, 2000). For example, a product at 1 m
could be used to adjust 5 m SPOT imagery as a
rst stage. The corrected SPOT could then be used
to adjust 30 m Landsat, and it in turn to correct
MODIS 250 m imagery. This is often referred
to as nested sampling or Nested Area Frame
Sampling and is a module in at least one of the
standard image processing systems.
One study using regression estimation was to
locate open water in North Dakota suitable for
waterfowl. The original Landsat estimates of the
area of ponds were only about 70% of the actual
open water compared to validation data. The
regression estimates using aerial photography
samples increased those estimates to within 8% of
the actual extent.
Regression estimation was performed over seven land
covers with an emphasis on forests in Brazil using aerial
photographs and Landsat images. The regression estimation
procedure was able to provide accurate results in a time-
effcient and cost-effective manner with better results
than from Landsat independently. Landsat imagery (30
m) in conjunction with AVHRR (1 km) images accurately
estimated Canadian burned forest areas to calculate carbon
storage (Fraser, 2004).
A study to ascertain the most accurate method to monitor
biomass burning in Central Africa determined the best
approach would utilize data from both fine and coarse spatial
resolution sensors and a regression estimator strategy. Im-
proved walrus counts were obtained using this strategy for a
rookery in the North Pacific Ocean (Barber et al., 1991).
Regression estimation has also been employed for mapping
agriculture. One study used field data with Landsat to
accurately determine the amount of winter rice area in Ban-
gladesh (Haack and Rafter, 2010). Figure 1 illustrates the
relationship between the two data sets. In Tanzania, crop
statistics were accurately determined via this approach and
similarly aerial photography and Landsat correctly estimat-
ed the amount of wheat in a region in Brazil
Summary
There is a record of successful applications of the regression
estimation strategy to improve the spatial statistics of a
variety of surface features. Unfortunately, this method does
not seem to be widely used, understood or even included
in current remote sensing textbooks. Given the increased
availability of fine spatial resolution satellite-based remote
sensing data at minimal or no cost, this column encourages
wider introduction of this very effective technique in univer-
sity curriculum.
References
Barber, D.; Richard, P.; Hochheim, K.; Orr, J., 1991. Cali-
bration of Aerial Thermal Infrared Imagery for Walrus
Population Assessment. Arctic (44):58-65.
Fraser, R.; Hall, R.; Landry, R.; Lynham, T.; Raymond,
D.; Lee, B.; Li, Z., 2004. Validation and Calibration of
Canada-wide Coarse Resolution Satellite Burned-area
Maps. Photogrammetric Engineering & Remote Sensing
(70):451–460.
Gallego, F., 2004. Remote Sensing and Land Cover Es-
timation. International Journal of Remote Sensing 25
(15):3019–3047.
Haack, B.; Rafter A., 2010. Regression Estimation Tech-
niques with Remote Sensing: A Review and Case Study.
Geocarto International 25(1): 71-82.
Koeln, G.; Kollasch, R., 2000. Crop Area Assessments Using
Low, Moderate and High Resolution Imagery: a Geotools
Approach. Rockville, MD: Earth Satellite Corporation.
Nelson, R., 1989. Regression and Ratio Estimators to Inte-
grate AVHRR and MSS Data. Remote Sensing of Environ-
ment (30):201–216.
Author
Dr. Barry Haack is a Professor of Geographic and Carto-
graphic Sciences at George Mason University in Fairfax,
Virginia, USA and is an ASPRS Fellow.
Figure 1. Scatter plot of the relationship between the Landsat identi ed areas of winter rice and
reported estimates by agricultural agents for 10 administrative areas in central Bangladesh. The
correlation between the sets of data was 0.98. The average ratio of Landsat to eld estimates of
boro area was determined to be 0.52 (Haack and Rafter, 2010)
Hectares Reported
Landsat Digitally Identi ed Hectares
564
452
340
230
113
0
325
261
196
130
65
0
