3. Gross error detection 
The topic of reliability and blunder detection is not only of great scientific interest but 
also of highest practical importance in aerial triangulation. Any other progress in the field 
remains secondary against the successful operational solution of the automatic detection of 
gross observational errors which are almost always present in large sets of observational data. 
3.1. Following the concept of Baarda on reliability extensive investigations were carried out 
by a number of authors on the internal and external reliability of photogrammetric blocks. 
Conclusions were drawn and rules derived for project planning, see |2| for instance. The 
reliability of photogrammetric blocks compares very well with geodetic networks. Especially 
inside a block the detectability of gross observational errors is very good. At border areas 
either double overlap or double or triple tie points are required in order to ensure sufficient 
reliability. Unfortunately gross errors can only be poorly detected at ground control points. 
This means that in addition to asking good reliability from the geodetic survey the photo- 
grammetric identification of ground control points must be ensured by using double or triple 
points. 
In general terms the specifications for project planning are sufficiently known in order to 
allow algorithms for gross error detection to be effective and successful. It can only be 
urged that practice will accept such specifications. 
3.2. At about 1980 it had been recognized that gross error detection by algorithmic methods 
is possible and that it is of greatest practical importance. The theoretical insight into 
the limits of gross error detection (normally only errors larger tan 6 g or more can be 
identified) had been prepared by the pioneering work of W. Baarda who formulated the statistical 
"data-sneoping" test. On the other hand it was soon realized .that the separation of small gross 
errors from random errors (and systematic errors) by a statistical test constitutes only one 
part of the problem. In practice a computer program would also have to identify large and very 
large gross errors, and the algorithm should not break down in the presence of multiple gross 
errors. For such problems very little theoretical guidelines were available, except for the 
recommendation that robust estimators would be required. At that time the first computer programs 
had been developed attempting a solution for automatic blunder detection. 
Against that background the WG III/1 decided to go into experimental tests about the practical 
performance of available gross error detection algorithms. Also, by comparison, the performance 
of the conventional manual method was to be tested. 
3.3. For four different blocks sets of observation data were established by computer simulation 
and contaminated by random, systematic and gross errors. The participants obtained the data 
with some general information and were asked to clear the data from gross errors, to do the 
blockadjustment and to judge the results obtained. Each participant was free to use any method 
of error detection available to him. The results were subsequently analyzed with regard to the 
success rate of gross error detection, to the method applied, and to the number of adjustment 
runs required. 
The experimental test program was carried out in two phases. In phase I the data were particulary 
contaminated with quite a number of very large gross errors. The aim was to test the ability of 
methods to remove the large gross errors first before tackling the problem of separating small 
gross errors from all other random and systematic errors. This set up was deliberately chosen 
       
   
    
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
    
   
  
  
  
  
  
  
   
   
  
     
as 
ext 
The 
of 
mu! 
The 
|6 
th