3. DATA AND METHODS 
Google Earth was used to examine the occurrence of gullies in 
randomly selected sites, and also as a tool for mapping gully 
extents. This data was then imported into a GIS and examined 
against several raster and vector data layers representing various 
landscape variables such as slope (derived from SRTM DEM), 
Foliage Projective Cover (FPC) (Armston et al., 2009), land 
clearing (Muir et al., 2011), 1:100,000 vector drainage line, 
geology, soils, and regional ecosystems (DSEWPC, 2011). 
Imagery on Google Earth enabled the identification of gullies; 
however, this was only possible over about a third of the 
Burdekin where Google Earth has Quickbird imagery coverage 
(<1 m pixel resolution). In the remaining 70% SPOT imagery 
(-2.5m pixel resolution) was found not to have sufficient 
resolution to identify gullies with high certainty (Figure 2). 
  
  
Random training site 
Random validation site 
  
  
| [7] Quickbird imagery A 
Hj] SPOT imagery 0 km 50 
I——À | 
  
Figure 2. Google Earth imagery coverage and training and 
validation sites. 
Five hundred and eighty two training sites (circle, 28.3 ha each) 
were randomly generated for areas in the Burdekin where 
Quickbird or GeoEye imagery were available on Google Earth 
(Figure 2). Sites were visually inspected on the imagery for the 
presence or absence of gullies. 104 training sites had gullies 
present, 433 training sites had no evidence of gullies or gullies 
were not visible. Forty five training sites were removed from 
the data set as determination of presence or absence was 
uncertain due to vegetation or similarity to other erosion 
features. 
Landscape variables were recorded at each gullied site to create 
a set of conditional classes representing the probability of 
gullies. Further assumptions were made about presence or 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
absence of gullies based on information provided by local 
experts and further visual inspection of imagery and field 
observations. These conditions were then applied across the 
Burdekin to create a binary map of no-gully areas and gully 
sensitive areas. 
Validation of the results was undertaken by field observation, 
expert local knowledge and the examination of a further 456 
validation sites (0.25 ha squares) on Google Earth (Figure 2). 
The validation sites were smaller than the training sites in order 
to prevent gully overestimation. This is believed to have been 
the case with the training sites as these were classified as 'gully' 
even if gullies composed only a small portion of their area. In 
line with this assumption, gullies were identified on only 12 
validation sites whereas 444 validation sites had no gullies or 
gullies were not visible. 
Knowledge of no-gully areas allowed targeted mapping on areas 
where gullies are more likely to occur. The mapping (Figure 3) 
was conducted in the form of manual digitizing on Quickbird 
imagery on Google Earth using the polygon tool and then 
importing vectors into ArcGIS. Overall, more than 5100 gullies 
were digitized over an area of more than 3500 km?. These data 
were used to validate and further refine the abilities to locate 
gullies within the gully sensitive area. 
  
White polygons indicate gullies. All other areas classified as 
having no gullies. Background imagery from Google Earth. 
The 5100 mapped gullies were then used as training data to 
examine a group of additional potential explanatory topographic 
variables derived from the SRTM 1" DEM (~30 m spatial 
resolution). The efficiency of the explanatory variables in 
predicting gully occurrence was assessed by calculating the 
Area Under the Curve (AUC). The AUC provides an effective 
measure of how much of the response variables distribution (in 
this case gully presence) is explained by a particular 
explanatory variable. Most variables offered a higher than 
random, yet still modest gully prediction ability (AUC 60-70%, 
compared with AUC of 50% for a random prediction). A further 
test using a multivariate logistic regression model combining 
several layers did not show significantly improved results. The 
variable that showed highest correlation with gully occurrence 
was elevation above drainage line, which had an AUC of 80% 
   
   
   
  
   
    
    
   
   
    
   
  
  
  
   
   
    
   
   
  
  
   
   
    
  
  
     
   
   
    
(Fig 
prot 
area 
Cumulative mapped gullies (%) 
pres 
nto 
of s 
hig] 
met 
obs 
hig] 
assi 
the 
no- 
Wit] 
pre 
for 
con 
det: 
Init 
froi 
Or 
ext 
on 
less