with a testing sample of about 25 % of the digitized input data 
for each class. The training pixels where neither used in the 
MRMR feature selection nor in the classifier training. The first 
lines refer to the feature sets in Table 2, the second part to 
feature sets of different sizes selected with the MRMR 
algorithm and the MIQ criterion. 
The results are quite convincing. The thermal data performed 
best of all feature sets from one sensor (with up to 96.3 % 
overall classification accuracy with the Multilayer Perceptron 
classifier). This is quite consequent, considering the thermal 
definition of the LCZ. The multispectral data also reached very 
good results (95.1% with Multilayer Perceptron), while the 
simple heights (74.0% with Multilayer Perceptron), the 
bandpass & directional filters (74.8 % with Random Forest) and 
the morphological profiles (83.2% with Random Forest) 
performed less well. 
The multisensor feature sets reached up to 97.496 (with 
Multilayer Perceptron and 100 features), but also small 
multisensoral feature sets reached good results (up to 92.5 % 
with 20, 93.7 % with 30, 94.9 % with 40 and 96.7 % with 50 
features). These results are comparable to other studies (c.f. 
Bechtel and Daneke, 2012). 
Regarding the different classifiers, Multilayer Perceptron 
showed the best results, but considering the much higher 
computing costs Random Forest seems a suitable and fast 
alternative (and can be assumed to be more robust and not 
overfitting). The Support Vector Machine classifiers performs 
worse than NN and RF for smaller feature sets but takes most 
benefit of further features (for all other classifiers the accuracy 
is higher for 100 selected features than for the full feature set). 
Naive Bayes performs less well than the highend classifiers and 
takes less advantage of further features (since the conditional 
independence assumption is violated with increasing 
redundancy in the feature set). 
Figure 2 shows the classification result for the full domain with 
the Multilayer Perceptron classifier and 100 selected 
multisensoral features. The visual evaluation reveals a very 
high accordance with the existing urban and natural structures. 
  
CLASSIFIER 
feature set NB SVM NN RFI RF2 RF3 
shs 063 064 1074. 072 071 0.72 
acp 0.73 068 070 071: 0.72 0.72 
ms& ndvi 074 094 095 090 093 0.93 
morph 067 078 080 080 0.82 0:83 
fft 0.60 0735 067 069 073 0:75 
tir 0.68 089 096 092 0.94 094 
MIQIO 083 0.85 091 091 09L 0.9] 
MIQ20 083 088 092 092. 092 092 
MIQ30 0.32 089 094 092093 093 
MIQ40 082 090 095 093 094 094 
MIQ50 083 091 097 093. 095 0.95 
MIQ60 0.82 092 096 094 0.93 095 
MIQ70 0.82 0.93 096 0.93. 0.95. 0.94 
MIQ80 082 093 09 095 095. 0.95 
MIQ90 082 094 097 094 095 0.96 
MIQ100 082 094. 097. 093-095 095 
all (484) 0.80 096 096 092 094 095 
Table 2. Classification results for different feature sets 
including multisensoral feature sets selected with the MRMR 
MIQ criterion. 
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 
industr 
modcore 
reghous 
  
Figure 2. LCZ classification with NN classifier and 100 
multisensoral features selected by MRMR (MIQ). Mixed 
classes overlay was digitized from map data. 
4.0 Empirical UHI model 
Further, the potential of the multitemporal und multisensoral 
parameter sets for spatiotemporal modelling of the mean UHI 
was tested with Linear Regression and Multilayer Perceptron 
models. The results are shown in Table 3. The models were 
again evaluated with a randomly chosen testing sample of about 
25% of all stations, which was not used in prior model 
calibration. 
As for the classification the multitemporal thermal (correlation 
R of 0.74/0.81 and mean absolute error of 0.17/0.19 K for 
Linear Regression and Neuronal Network) and the 
multitemporal spectral (R 0.75/0.77, MAE: 0.17/0.19 K) data 
performed better than the simple height statistics (R: 0.27/0.26, 
MAE: 0.24/0.26 K) and the annual cycle parameters (R: 
0.19/0.18, MAE: 0.24/0.24 K). This is, certainly partly a 
consequence of the different parameter set sizes. 
Although an error of only about 0.2 K in the predicted mean 
UHI seams promising at the first glance, the quality of the 
empirical models is not yet satisfying in respect to the low 
spatial variance in the dataset (even for the better models this 
corresponds to an relative error of about 60 %). This is not 
surprising regarding the large number of processes contributing 
to alterations in the urban atmosphere. Those effects are related 
to the urban structure, vegetation, and surface temperature in 
different ways and hence only very indirectly to the 
multispectral and thermal data. However, the multispectral data 
comprises different insulation and phenological conditions and 
this information can be utilised during the model calibration. 
Multisensoral data including thermal, spectral and height 
features performed better than sets from the same sensor. 
However, this was not the case for preselected features with 
high individual relevance (R > 0.25), which might indicate a 
certain redundancy. Further, the testing sample might not be 
completely independent, since close stations are likely to be 
correlated. 
    
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
   
  
   
   
   
  
   
   
   
   
  
   
   
   
   
   
    
    
   
  
  
    
  
     
   
   
   
   
  
   
   
   
  
   
   
  
  
   
  
    
   
  
  
    
Par 
‘shs 
acp 
tir 
nd 
R> 
all 
di 
Nev 
mul 
mod 
Bec] 
Lan 
flori 
The 
mod 
stan 
one 
sho 
with 
and 
of I 
spat 
The 
mul 
func 
urb: 
data 
the 
viey 
The 
clin 
Altl 
surf 
urb: 
data 
inve 
of | 
moc 
slig