Generate Random Training Points
The RBFLN neural network object requires two sets of training points:
The two sets of points (locations) are combined as training data.
This second type of data set is often not readily available so this function offers one way of approximating such a data set. The idea is to generate a set of points in parts of the study area where there is a very low probability of the object occurring.
One possibility for applying the function is to generate a probability map (response theme) using the weights of evidence or logistic regression methods and then generate a set of random points in areas of very low probability. Another could be to simply use some data whose low values would indicate that such a "Deposits" site would never occur. Then to generate the "Non-Deposits" training sites, the following can be done:
In this example, the
threshold should be lower than the prior probability. The
prior probability can be found in the third column of the
last row of the weights table associated with the
response theme, or can be reported by clicking the button
in the deposit training point theme properties dialog.
This dialog is displayed by clicking the 'Properties'
button in the 'Analysis Parameters' dialog.
|
General Application
The 'Generate Random Training Points' function can use any integer grid theme in the active view. The minimum and maximum values in the grid theme are reported as guidance and the default value is the midpoint between these two. An evidence theme, whose low values indicated areas where 'deposits' are highly unlikely to occur, can be used to define a 'non-deposits' training set. Another strategy might be to create several such sets of 'non-deposits' and evaluate the differences between the resulting models.
Next | Section Contents | Home |