Tasks list - Githubissues

[x] Implement optional reading of covariates from pickled files rather than reading from geotiffs
[x] flag targets that return nan for all covariates and ignore these targets in the training
[x] fix bug in transformset.py and features.py that was dropping nodata mask values (# 371b450)
[x] Add option to save raw covariates after intersecting the targets
[x] Convert any nans in the covariates to NoDataValue, in addition to the well defined NoDataValue pixels. This prevents many crashes as we can impute the nans just like NoDataValues.
[x] add nan count in geoinfo output
[ ] Pipeline/optimisation
[x] Support revrand v0.9
[x] Enable optimisation pipeline for revrand gaussian processes
[x] Mix resampling techniques
[x] Add bootstrapping on/off option during resampling
[x] Depth optimisation
[x] Kriging (Integration of PyKrige: https://github.com/bsmurphy/PyKrige)
[x] regression kriging: https://en.wikipedia.org/wiki/Regression-Kriging
[x] New score method to optimise based on ML fit score in the transformed targets domain
[x] Classification
[x] Classification - what constitutes the classes
[x] Bootstrap cubist
[x] Add csv output and plot of y_transformed real vs y_transformed pred, i.e., the regression that was performed before transforming targets back to original scale
[x] Fix cubist cross val: https://github.com/NICTA/uncover-ml/issues/129
[x] Add projection information to geoinfo output
[x] Add projection information in the output geotifs
[x] Optional generation of the variance image output and Q95-Q5
[ ] Add option to specify area of learn/predict (from large datafiles)
[x] assess intersected covariates by generating scatter plots for each
[ ] produce outputs for the optimised model
[ ] output scores in both transformed (ML score) and original scale.
[x] Cubist should not use committe model by default
[x] fix feature ranking

GeoscienceAustralia / uncover-ml

Tasks list #1