libMontyML.py

read_volcano_def():

• reads volcano_def.csv into a DataFrame
• strips left space from column headings
• returns DataFrame

build_master_event_catalog():

• reads and concatenates all reawav_MVOE_YYYYMM.csv files, sets index to filetime and sorts.
• for mainclass 'R' and 'D', set subclass=mainclass, then drop mainclass
• initialize columns for each subclass in subclasses_for_ML (100% for original subclass, 0% elsewhere)
• initialize weight to 3, and checked, split, delete and ignore to False
• return DataFrame and save to CSV file

parse_STATION0HYP():

• selects lines of length 25, reads station name and coordinates
• creates a DataFrame with columns name, lat, lon, elev

get_weighted_fingerprints():

• calls _select_best_events() to build a dict of the best N events of each class
• loops over subclasses_in_ML, selecting that element of the dict

• if at least 30 events, initialize a fingerprint DataFrame, with columns mean, std, 25%, 50%, 75%
• computes these stats with a weighted statistical model using statsmodels.stats.weightstats.DescrStatsW() with events weighted using weight*subclass_percentage

• returns fingerprints dict of DataFrames

save_fingerprints() just saves fingerprints dict to CSV files.

qc_event():

• calls _select_next_event(), which examines checked events, and picks the least common label from subclasses_for_ML. Then it extracts all matching unchecked events, and ranks them based on detection_quality, snr, and quality, in that order. Also multiples by weight, but that should always be 3 for unchecked events. Sorts in descending order and subsets the DataFrame to the first row and returns it.
• loops over all rows (but I think there should only be 1):

• reads the picklefile, applies 0.5-25.0 Hz, 4 pole filter, and calls deconvolve_instrument_response(). This loops over each trace and, if the picklefile is not already in m/s units, it calls load_mvo_inventory() on the CAL directory which tries to return the corresponding Inventory, and then tr.remove_response() is called and then units updated.
• computes amplitude and energy for each tr, and stores these in dfenergy DataFrame. Could just call an existing method here.
• oddly a suggested_weight is then computed, but it appears to be log10 of energy/2 of the first tr if that tr is in Counts, otherwise None. So why bother correcting the data? And the focus on st[0] seems arbitrary. *> plot event
• add station locations to tr.stats in st by calling add_station_locations() and then call plot_amplitude_locations()
• print info including row, tracedf (id, medianF, bw_min, peakF, bw_max, band_ratio (1-6, 6-11), kurtosis)
• guess label based on fingerprints and row by calling _guess_subclass()
• show Seisan subclasses (from VOLCANO.DEF)
• parse input event percentages and weight, updating DataFrame accordingly (or q/s/I/d to quit, split, ignore or delete)

• returns the DataFrame

remove_marked_events(): subsets the catalog DataFrame to eliminate rows with delete, ignore or split flags set to True. returns new DataFrame

to_AAA():

• subsets catalog DataFrame on checked events
• eliminates labels with less than 10 events, those remaining are included_subclasses, and merged into dfAAA
• columns renamed: twin->length, new_subclass->class
• subsetted on columns class, year, month, day, hour, minute, second, length, path
• WAV-files optionally copied to /new/path/class/WAVfile
• dfAAA written to CSV

report_checked_events():

• subsets df on checked events
• reports:

• number of checked events
• number of classified events after those flagged to ignore, delete or split removed
• number of events by label in subclasses_for_ML
• total number of events matching subclasses_for_ML
• total number of reclassified events (different label than original)
• total number of events that were correctly classified
• error rate

Others: