‘Batteries’ in Machine Learning: A First Experimental Assessment of Inference for Siberian Crane Breeding Grounds in the Russian High Arctic Based on ‘Shaving’ 74 Predictors

Abstract

The Siberian crane (Leucogeranus leucogeranus,) remains an elusive but highly regarded species of global conservation concern. Breeding regions occur in the Russian high arctic, and two subpopulations are known. Here we present for the first time a machine learning-based summer habitat analysis using nesting data for the eastern population in the breeding grounds employing predictive modeling with 74 GIS predictors. There is a typical desire for parsimony to help increase interpretability of models, but findings generally show that it would not result in greatest improvement to the model and inference. ‘Batteries’ are a new concept in machine learning allowing to test a set of experiments that help to test on predictors and model selection. Here we show 28 of those ‘batteries’ and compared multiple approaches to model runs from iteratively dropping the least or most important predictor (‘variable shaving’) to allow all predictors to contribute. It was found that the generic ‘kitchen sink’ model with TreeNet (an optimized boosting algorithm from Salford Systems Ltd) performs best. However, while the use of ‘batteries’ remain widely underused in wildlife conservation management, ‘shaving’ was of great use to learn about the structure, role and impacts of predictors and their spatial performance supporting non-parsimonious work. Of great interest is the finding that a bundle of low-ranked predictors performs almost equal to, or better than, the so-called top predictors. This is called ‘Predictor swapping’. This is the best and most detailed habitat study and prediction for the Siberian crane in summer, thus far. It is to be used for conservation management and as a generic template for any species while data availability and the environmental crisis are on the rise, specifically for the high Arctic.

Appendices

Appendix 1: Details of 74 GIS Environmental layers Used in the Model Prediction (+ 3 Additional Internal Columns)

#	Name and abbreviation of GIS layer	Source	Comment
1–12	Monthly mean temperature tmen_1–12	Worldclim.org	These are standard layers used for GIS modeling
13–24	Monthly minimum temperature tmin_1–12	Worldclim.org	(see above)
25–36	Monthly maximum temperature tmax 1–12	Worldclim.org	(see above)
37–48	Monthly precipitation prec_1–12	Worldclim.org	(see above)
49–67	Bioclim bio_1–19	worldclim.org/bioclim	(see above)
68	Altitude	Worldclim.org	(see above)
69	Aspect	Worldclim.org	(see above)
70	Slope	Worldclim.org	(see above)
71	Landcover Landcv	Herrick et al. (2013)	Several of global landcover layers exist
72	Human infrastructure index Hii	Herrick et al. (2013)	Human footprint. Several human footprint layers
73	Distance to waterbody/lake Dislke	Mi unpublished	While essential for cranes, this layer is unlikely to be very accurate due to the huge and ephemeral wetlands worldwide
74	Distance to coastline Discsln	Mi unpublished	Relies on the coastline map resolution
75	x coordinate	ArcGIS	Not often used in most GIS model work but important for geo-referencing
76	y coordinate	ArcGIS	Not often used in most GIS model work but important for geo-referencing
77	Row index FID	ArcGIS	Not often used in most GIS model work but important for row identification

Appendix 2

1.1 List of Top 20 Predictors, as identified by TreeNet ranking

Predictor	Relative Importance
Bio12	100.0
Bio14	71.2
Bio17	44.2
TMEN9	40.1
Prec12	37.6
Distance to lake	35.1
TMAX12	29.8
Altitude	27.3
Slope	25.9
Tmin1	23.8
Bio1	23.0
Bio19	20.4
Tmen2	19.2
Bio3	18.9
Tmax3	17.9
Bio6	16.3
Tmen7	15.9
Prec6	14.3
Prec7	13.9
Tmin6	12.9

Appendix 3

1.1 Prediction Model Details for the Best Performing Model (the ‘Kitchen sink model’ with 74 predictors)

Siberian crane with a battery run on TreeNet (SPM7) balanced

The kitchensink model, all 74 environmental predictors

Frequency of Prediction Relative Index of Ocurrence (RIO 0-1) for known presence (1)

Appendix 4

(For Prediction map 1 for the ‘Kitchen sink model ’ see Fig. 8.4 in the text; for map legends please see this figure; same for all other appendix maps)

(For Prediction map 2 for the ‘TMax12 model’ see Fig. 8.5 in the text)

1.1 Prediction Map 3 for the ‘BIO14 model’

1.2 Prediction Map 4 for the ‘TMax12BIO14 model’

1.3 Prediction Map 5 for the ‘Top5 model’

1.4 Prediction Map 6 for the ‘Top10 model’

1.5 Prediction Map 7 for the ‘Top29 model’

1.6 Prediction Map 8 for the ‘Top35 model’

1.7 Prediction Map 9 for the ‘Bottom 44 model’

1.8 Prediction map 10 for the ‘Leaving out top 3 interacting predictors model’