Functions to generate probabilistic estimates of annual accumulation from ice-penetrating radar without the need for manual layer selection or correction
3
stars
0
forks
source link
Improved optimization of logistic regression parameters #30
After extensive experimentation, this PR represents improvements to the methods and final results of logistic regression parameter optimization. It includes optimizations based on residuals in age-depth scales rather than direct layer tracing (better represents final objectives of PAIPR) and a more probabilistic approach by estimating parameters at each trace in the validation echograms. We use an assumed value of the k parameter (such that a distance-brightness value of 0 represents < 1% chance of representing an annual layer) in order to avoid issues with multiple local minima making a final scalar estimate of parameters challenging to produce.
durbank
commented
5 years ago
After extensive experimentation, this PR represents improvements to the methods and final results of logistic regression parameter optimization. It includes optimizations based on residuals in age-depth scales rather than direct layer tracing (better represents final objectives of PAIPR) and a more probabilistic approach by estimating parameters at each trace in the validation echograms. We use an assumed value of the k parameter (such that a distance-brightness value of 0 represents < 1% chance of representing an annual layer) in order to avoid issues with multiple local minima making a final scalar estimate of parameters challenging to produce.