Skip to content

FlukeAndFeather/stickleback

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

87 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

stickleback

A machine learning pipeline for detecting fine-scale behavioral events in bio-logging data.

Installation

Install with pip.

pip install stickleback

Key concepts

  • Behavioral events are brief behaviors that can be represented as a point in time, e.g. feeding or social interactions.
  • High-resolution bio-logging data (e.g. from accelerometers and magnetometers) are multi-variate time series. Traditional classifiers struggle with time series data.
  • stickleback takes a time series classification approach to detect behavioral events in longitudinal bio-logging data.

Quick start

Load sample data

The included sensor data contains the depth, pitch, roll, and speed of six blue whales at 10 Hz, and the event data contains the times of lunge-feeding behaviors.

import pandas as pd
import sktime.classification.interval_based
import sktime.classification.compose
from stickleback.stickleback import Stickleback
import stickleback.data
import stickleback.util
import stickleback.visualize

# Load sample data
sensors, events = stickleback.data.load_lunges()

# Split into test and train (3 deployments each)
def split_dict(d, ks):
    dict1 = {k: v for k, v in d.items() if k in ks}
    dict2 = {k: v for k, v in d.items() if k not in ks}
    return dict1, dict2

test_deployids = list(sensors.keys())[0:2]
sensors_test, sensors_train = split_dict(sensors, test_deployids)
events_test, events_train = split_dict(events, test_deployids)
sensors[test_deployids[0]]
<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }
.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}
</style>
depth pitch roll speed
datetime
2018-09-05 11:55:52.400 14.911083 -0.059933 -0.012899 4.274450
2018-09-05 11:55:52.500 14.910864 -0.067072 -0.010815 4.044154
2018-09-05 11:55:52.600 14.915853 -0.075173 -0.008335 3.820568
2018-09-05 11:55:52.700 14.923190 -0.085225 -0.005727 3.602702
2018-09-05 11:55:52.800 14.928955 -0.096173 -0.002803 3.432342
... ... ... ... ...
2018-09-05 13:55:51.900 22.552306 -0.010861 0.005441 2.246061
2018-09-05 13:55:52.000 22.571625 -0.010534 0.004674 2.257525
2018-09-05 13:55:52.100 22.588129 -0.010081 0.003841 2.267966
2018-09-05 13:55:52.200 22.603341 -0.009627 0.003042 2.272327
2018-09-05 13:55:52.300 22.619537 -0.009355 0.002164 2.277328

72000 rows × 4 columns

Visualize sensor and event data

plot_sensors_events() produces an interactive figure for exploring bio-logger data.

# Choose one deployment to visualize
deployid = list(sensors.keys())[0]
stickleback.visualize.plot_sensors_events(deployid, sensors, events)

Animated loop of interactively exploring data with plot_sensors_events()

Define model

Initialize a Stickleback model using Supervised Time Series Forests and a 5 s window.

# Supervised Time Series Forests ensembled across the columns of `sensors`
cols = sensors[list(sensors.keys())[0]].columns
tsc = sktime.classification.interval_based.SupervisedTimeSeriesForest(n_estimators=2,
                                                                      random_state=4321)
stsf = sktime.classification.compose.ColumnEnsembleClassifier(
    estimators = [('STSF_{}'.format(col),
                   tsc,
                   [i])
                  for i, col in enumerate(cols)]
)

sb = Stickleback(
    local_clf=stsf,
    win_size=50,
    tol=pd.Timedelta("5s"),
    nth=10,
    n_folds=4,
    seed=1234
)

Fit model

Fit the Stickleback object to the training data.

sb.fit(sensors_train, events_train)

Test model

Use the fitted Stickleback model to predict occurence of lunge-feeding events in the test dataset.

predictions = sb.predict(sensors_test)

Assess results

Use the temporal tolerance (in this example, 5 s) to assess model predictions. Visualize with an outcome table and an interactive visualization. In the figure, blue = true positive, hollow red = false negative, and solid red = false positive.

outcomes = sb.assess(predictions, events_test)
stickleback.visualize.outcome_table(outcomes, sensors_test)
<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }
.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}
</style>
F1 TP FP FN Duration (hours)
deployid
bw180905-49 1.000000 44 0 0 1.999972
bw180905-53 0.943396 25 2 1 1.999972
deployid = list(events_test.keys())[0]
stickleback.visualize.plot_predictions(deployid, 
                                       sensors_test, 
                                       predictions, 
                                       outcomes)

Animated loop of interactively exploring predictions with plot_predictions()

About

Automated behavioral event detection in bio-logging data.

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages