Comparing the PIF approach to a pixel-based approach for birds in Alberta

This talk is an earlier draft of the research paper:

Sólymos, P., Toms, J. D., Matsuoka, S. M., Cumming, S. G., Barker, N. K. S., Thogmartin, W. E., Stralberg, D., Crosby, A. D., Dénes, F. V., Haché, S., Mahon, C. L., Schmiegelow, F. K. A., and Bayne, E. M., 2020. Lessons learned from comparing spatially explicit models and the Partners in Flight approach to estimate population sizes of boreal birds in Alberta, Canada. Condor, https://doi.org/10.1093/condor/duaa007

Transcript

1. At the end of the road
   Comparing the PIF approach to a
   pixel-based approach for birds in Alberta
   Péter Sólymos
   & Boreal Avian Modelling Project Team
   PIF Science – July 12, 2018
   1
   

   View Slide

2. Measure what you manage
   PIF population size estimator is sample based
   The approach involves several key assumptions:
   • about detectability
   • and sample representation
   2
   

   View Slide

3. Adjustment factors
   Detection distance:
   • MDD based on a field study & expert assessment
   • Found to be most influential on estimates [1]
   • Pop size can be 5x higher when using EDR [2]
   • MDDs were revised in 2013
   [1] Thogmartin 2010 Ecol Mod 221:173
   [2] Matsuoka et al. 2012 Auk 129:268
   Distance from observer
   P(detection | availability)
   MDD
   1
   0
   3
   

   View Slide

4. Adjustment factors
   Detection distance:
   • MDD based on a field study & expert assessment
   • Found to be most influential on estimates [1]
   • Pop size can be 5x higher when using EDR [2]
   • MDDs were revised in 2013
   [1] Thogmartin 2010 Ecol Mod 221:173
   [2] Matsuoka et al. 2012 Auk 129:268
   Distance from observer
   P(detection | availability)
   EDR MDD
   1
   0
   4
   

   View Slide

5. Adjustment factors
   Time-of-day adjustment:
   • Empirically derived
   • Found to be less influential on estimates
   • Pop size can be 1.2x higher when using P(avail) [1]
   5
   Time-of-day (stop #)
   Mean count / max
   Tadj = Max / Mean
   0
   1
   0
   

   View Slide

6. Adjustment factors
   Time-of-day adjustment:
   • Empirically derived
   • Found to be less influential on estimates
   • Pop size can be 1.2x higher when using P(avail) [1]
   [1] Sólymos et al. in press Condor Time-of-day (stop #)
   Mean count / max
   Tadj = Max / Mean 1/P(avail)
   P(availability | presence)
   1
   0
   1
   0
   P max = 0.66 < 1
   6
   

   View Slide

7. Adjustment factors
   Pair adjustment:
   • Originally defaulted to 2
   • Species specific values used since 2013
   7
   

   View Slide

8. Adjustment factors
   Pair adjustment:
   • Originally defaulted to 2
   • Species specific values used since 2013
   These adjustment factors are not fully effective,
   but they do go some way towards addressing
   potential bias in the population estimates
   8
   

   View Slide

9. Other potential sources of bias
   • Habitat representation
   • Roadside count
   Assumptions were necessary to extrapolate
   from BBS counts to unsampled areas
   Comparison to Atlas data indicated that assumptions
   were reasonable.
   9
   

   View Slide

10. BCR 6
    “any land-cover-based
    roadside bias in the bird
    data of the BBS is likely
    minimal” [1]
    10
    [1] Veech et al. 2017 Condor 119:607
    

    View Slide

11. BCR 6
    “any land-cover-based
    roadside bias in the bird
    data of the BBS is likely
    minimal” [1]
    “anthropogenic disturbance
    is overrepresented in
    southern strata, and forest
    fires are underrepresented in
    almost all strata.” [2]
    [1] Veech et al. 2017 Condor 119:607
    [2] Van Wilgenburg et al. 2015 ACE 10:5 11
    

    View Slide

12. Land cover Availability % Roadside %
    Deciduous forest <80 yr 13 23
    Deciduous forest ≥80 yr 10 3
    Mixedwood forest <80 yr 3 3
    Mixedwood forest ≥80 yr 3 2
    Coniferous forest <80 yr 3 1
    Coniferous forest ≥80 yr 4 2
    Pine dominated forest <80 yr 6 5
    Pine dominated forest ≥80 yr 4 3
    Black spruce forest <80 yr 9 3
    Black spruce forest ≥80 yr 11 2
    Larch dominated wetland <80 yr 2 0
    Larch dominated wetland ≥80 yr 2 0
    Grasses and herbs 0 1
    Shrubs 2 1
    Swamp 4 1
    Shrubby wetlands 8 1
    Non-treed wetlands 2 0
    Cultivation 11 45
    Urban and industrial 1 3
    BCR 6
    “at the northern limit of BBS coverage where roadless areas predominate and
    roads typically sample a geographically biased portion of the landscape” [1]
    [1] Rosenberg and Blancher 2005
    12
    Roadside surveys do not
    sample land covers in
    proportion to their
    availabilities in N Alberta
    

    View Slide

13. BAM Team
    13
    Steering Committee
    Erin Bayne (U of A), Steve Cumming (U. Laval), Samantha Song (ECCC), Fiona Schmiegelow (U of A)
    Project Staff, Students, and Contributing Scientists
    Nicole Barker (Coordinating Scientist), Trish Fontaine (Spatial Database Manager), Peter Solymos (Statistical Ecologist), Diana
    Stralberg (Ecologist), Lionel Leston (PDF), Alberto Suarez Esteban (PDF), Tara Stehelin (PhD Student), Alana Westwood (Contributing
    Scientist), Samuel Haché (Contributing Scientist, CWS), Lisa Mahon (Contributing Scientist, CWS), Steven Van Wilgenburg
    (Contributing Scientist, CWS), Judith Toms (Contributing Scientist, CWS), Steve Matsuoka (Contributing Scientist, USGS Alaska).
    Technical Committee
    • Marcel Darveau, Ducks Unlimited Canada & Université Laval
    • André Desrochers, Université Laval
    • Pierre Drapeau, Université de Québec à Montréal
    • Charles Francis, Environment and Climate Change Canada
    • Colleen Handel, United States Geological Survey
    • Keith Hobson, Environment and Climate Change Canada
    & University of Western Ontario
    • Craig Machtans, Environment and Climate Change Canada
    • Julienne Morissette, Ducks Unlimited Canada
    • Gerald Niemi, University of Minnesota - Duluth
    • Rob Rempel, Ontario Ministry of Natural Resources
    • Stuart Slattery, Ducks Unlimited Canada
    • Phil Taylor, Acadia University/BSC
    • Lisa Venier, Canadian Forest Service
    • Pierre Vernier, University of Alberta
    • Marc-André Villard, Université de Moncton
    

    View Slide

14. Funding partners
    14
    Grants
    Environment and Climate Change Canada (Canadian Wildlife Service, Migratory Birds Program)
    Joint Oil Sands Monitoring
    Institutional/Infrastructure Support
    University of Alberta
    Université Laval
    Past Funders
    Alberta Biodiversity Monitoring Institute, Alberta Conservation Association, Alberta Innovates Technology Futures, Alberta Land-use
    Framework (Gov. Alberta), Alberta Research Council Inc., Alberta Pacific Forest Industries Inc., Boreal Ecosystems Assessment for
    Conservation Networks (BEACONs), Canada Foundation for Innovation, Canada Research Chairs, Canadian Boreal Initiative, Canada
    Foundation for Innovation, Climate Change and Emissions Management Corporation, Ducks Unlimited Canada, EC Habitat
    Stewardship Funds, Environmental Studies Research Fund, Fonds Québécois de la recherche sur la nature et les technologies,
    Forest Products Association of Canada, Geomatics for Informed Decisions (GEOIDE), Killam Trusts, Ministère des Ressources
    naturelles et de la Faune (MRNF) Quebec, National Fish & Wildlife Foundation, NSERC, Sustainable Forest Management Network, US
    Landscape Conservation Cooperatives, USFWS Neotropical Migratory Bird Conservation Act Grants Program, Vanier Canada Graduate
    Scholarships
    

    View Slide

15. Data partners
    15
    Individuals
    K. Aitken, A. Ajmi, B. Andres, J. Ball, E. Bayne, P
    . Belagus, S. Bennett, R. Berger, M. Betts, J. Bielech, A. Bismanis, R. Brown, M. Cadman, D.
    Collister, M. Cranny, S. Cumming, L. Darling, M. Darveau, C. De La Mare, A. Desrochers, T. Diamond, M. Donnelly, C. Downs, P
    . Drapeau, C.
    Duane, B. Dube, D. Dye, R. Eccles, P
    . Farrington, R. Fernandes, M. Flamme, D. Fortin, K. Foster, M. Gill, T. Gotthardt, N. Guldager, R. Hall, C.
    Handel, S. Hannon, B. Harrison, C. Harwood, J. Herbers, K. Hobson, M.-A. Hudson, L. Imbeau, P
    . Johnstone, V. Keenan, K. Koch, M. Laker, S.
    Lapointe, R. Latifovic, R. Lauzon, M. Leblanc, L. Ledrew, J. Lemaitre, D. Lepage, B. MacCallum, P
    . MacDonell, C. Machtans, C. McIntyre, M.
    McGovern, D. McKenney, S. Mason, L. Morgantini, L. Morton, G. Niemi, T. Nudds, P
    . Papadol, M. Phinney, D. Phoenix, D. Pinaud, D. Player, D.
    Price, R. Rempel, A. Rosaasen, S. Running, R. Russell, C. Savignac, J. Schieck, F
    . Schmiegelow, D. Shaw, P
    . Sinclair, A. Smith, S. Song, C.
    Spytz, D. Swanson, S. Swanson, P
    . Taylor, S. Van Wilgenburg, P
    . Vernier, M.-A. Villard, D. Whitaker, T. Wild, J. Witiw, S. Wyshynski, M.
    Yaremko, as well as the hundred of volunteers collecting Breeding Bird Survey (BBS) data.
    Institutions
    Acadia University; Alaska Bird Observatory; Alaska Natural Heritage Program; Alberta Biodiversity Monitoring Institute; Alberta Pacific
    Forest Industries Inc.; AMEC Earth & Environmental; AREVA Resources Canada Inc.; Avian Knowledge Network; AXYS Environmental
    Consulting Ltd.; Bighorn Wildlife Technologies Ltd.; Bird Studies Canada; Breeding Bird Survey (coordinated in Canada by Environment
    Canada); BC Breeding Bird Atlas; Canadian Natural Resources Ltd.; Canfor Corporation; Daishowa Marubeni International Ltd; Canada
    Centre for Remote Sensing and Canadian Forest Service, Natural Resources Canada; Canadian Wildlife Service and Science & Technology
    Branch, Environment Canada; Global Land Cover Facility; Golder Associates Ltd.; Government of British Columbia; Government of Yukon;
    Hinton Wood Products; Hydro-Québec Équipement; Kluane Ecosystem Monitoring Project; Komex International Ltd.; Louisiana Pacific
    Canada Ltd.; Manitoba Breeding Bird Atlas; Manitoba Hydro; Manitoba Model Forest Inc.; Manning Diversified Forest Products Ltd.;
    Maritimes Breeding Bird Atlas; Matrix Solutions Inc. Environment & Engineering; MEG Energy Corp.; Mirkwood Ecological Consultants Ltd.;
    NatureCounts; Nature Serve; Numerical Terradynamic Simulation Group; Ontario Breeding Bird Atlas; Ontario Ministry of Natural Resources;
    OPTI Canada Inc.; PanCanadian Petroleum Limited; Parks Canada (Mountain National Parks Avian Monitoring Database); Petro Canada;
    Principal Wildlife Resource Consulting; Regroupement QuébecOiseaux; Rio Alto Resources International Inc.; Saskatchewan Environment;
    Shell Canada Ltd.; Suncor Energy Inc.; Tembec Industries Inc.; Tolko Industries Ltd.; U.S. Army; U.S. Fish and Wildlife Service; U.S.
    Geological Survey, Alaska Science Center; U.S. National Park Service; Université de Moncton; Université du Québec à Montréal; Université
    du Québec en Abitibi-Témiscamingue; Université Laval; University of Alaska, Fairbanks; University of Alberta; University of British Columbia;
    University of Guelph; University of New Brunswick; University of Northern British Columbia; URSUS Ecosystem Management Ltd.; West
    Fraser Timber Co. Ltd.; Weyerhaeuser Company Ltd.; Wildlife Resource Consulting Services MB Inc.
    

    View Slide

16. Hutto et al. 1995
    Matsuoka et al. 2011
    57% positive
    20% neutral
    23% negative
    16
    

    View Slide

17. Roadside count assumption
    To what extent it is:
    • numerical response,
    • habitat use / behaviour, or
    • detectability?
    Yip et al. 2017 Condor 119:73
    17
    

    View Slide

18. Ball et al. 2016 ACE 11:10
    “Pixel” based
    approaches
    CAWA
    Stralberg et al. 2015 Ecol Appl 21:1112
    Spatial proximity
    Land use and disturbance
    Climate
    18
    Thogmartin et al. 2004 Ecol Appl 14:1766
    

    View Slide

19. My goal today
    Compare sample based (PIF approach) and
    model based (“pixel” approach)
    population size estimates in BCR 6 Alberta
    Understand how and why
    population size estimates differ
    Suggest future improvements to
    continental scale population size estimation
    1
    2
    3
    19
    

    View Slide

20. N
    PIF
    E[Y
    1
    ] P T
    adj
    Area
    1
    πMDD2
    =
    Roadside count
    Pair adjustment
    Time adjustment
    Detection distance
    Area of region
    20
    

    View Slide

21. N
    PIF
    N
    PIX
    E[Y
    1
    ]
    E[Y
    0
    ]
    P T
    adj
    Area
    1
    πMDD2
    =
    Roadside count
    Pair adjustment
    Time adjustment
    Detection distance
    = P 1/p
    3min
    1
    πEDR2
    Area
    Area of region
    21
    

    View Slide

22. N
    PIF
    N
    PIX
    E[Y
    1
    ]
    E[Y
    0
    ]
    P T
    adj
    Area
    1
    πMDD2
    =
    P 1/p
    3min
    1
    πEDR2
    Area
    22
    

    View Slide

23. N
    PIF
    N
    PIX
    E[Y
    1
    ]
    E[Y
    0
    ]
    P T
    adj
    Area
    1
    πMDD2
    =
    P 1/p
    3min
    1
    πEDR2
    Area
    23
    We estimate # of singing individuals Area of region is the same (BCR 6 Alberta)
    

    View Slide

24. N
    PIF
    N
    PIX
    E[Y
    1
    ]
    E[Y
    0
    ]
    T
    adj
    1
    πMDD2
    =
    1/p
    3min
    1
    πEDR2
    24
    

    View Slide

25. N
    PIF
    N
    PIX
    E[Y
    1
    ]
    E[Y
    0
    ]
    T
    adj
    1
    πMDD2
    =
    1/p
    3min
    1
    πEDR2
    Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    25
    

    View Slide

26. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    26
    Σ D
    k
    w
    k
    Σ D
    k
    a
    k
    N
    PIF
    ~ N
    true
    H
    adj
    H
    adj
    =
    N
    PIX
    N
    PIF
    =
    1
    H
    adj
    …
    

    View Slide

27. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    27
    Absolute abundance
    increased >10x
    for some species
    Others became
    less abundant
    without pair adjustment
    

    View Slide

28. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    28
    Ranking has changed for
    some sensitive species,
    but not for others
    Absolute changes do not always
    translate into change in rank
    without pair adjustment
    

    View Slide

29. 29
    Canonical correspondence analysis using abundance in land cover types
    

    View Slide

30. 30
    Canonical correspondence analysis using abundance in land cover types
    Open, developed
    Treed wetlands
    Upland forests
    Mature – Young
    Coniferous – Deciduous
    

    View Slide

31. 31
    Canonical correspondence analysis using abundance in land cover types
    Species associated with
    open habitats tend to have higher
    PIF population estimates
    

    View Slide

32. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    N
    PIX
    > N
    PIF
    for ¾ of the species
    OBS = Observed log ratio of PIX/PIF estimates
    EXP = Expected log ratio = R + T + A + H
    EXP is a little bit higher, but not too much
    32
    OBS EXP
    

    View Slide

33. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    The roadside count component
    is almost symmetric around 0,
    large variation across species
    33
    

    View Slide

34. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    The time component
    is positive for most
    species, variation is
    relatively small
    34
    

    View Slide

35. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    Area component is positive
    for all species, higher than
    time adjustment, smaller
    variation across species
    35
    

    View Slide

36. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    36
    Habitat component
    is centered at 0,
    moderate variation
    across species
    

    View Slide

37. Roadside count
    Habitat adjustment
    Time adjustment
    Area sampled
    E[Y
    0
    ]
    E[Y
    1
    ]
    1
    p
    3min
    T
    adj
    MDD2
    EDR2
    =
    1
    H
    adj
    N
    PIX
    N
    PIF
    OBS = a + b
    1
    R + b
    2
    T + b
    3
    A + b
    4
    H + error
    Ideally, a = 0, b = 1, error = 0, so that OBS = R + T + A + H
    37
    

    View Slide

38. OBS = a + b
    1
    R + b
    2
    T + b
    3
    A + b
    4
    H + error
    Ideally, a = 0, b = 1, error = 0, so that OBS = R + T + A + H
    • Road count & Habitat related biases dominate
    • EDR vs. MDD was thought to be most influential – now only 3rd
    • Habitat adjustment values are underestimated
    • There is still something left to be explained (~10%)
    • Interactions were not important
    38
    Model terms Estimate SE P-value % variance
    a –0.141 0.170P(a = 0) = 0.411
    b
    1
    (roadside count) 0.951 0.054P(b1 = 1) = 0.364 58.5
    b
    2
    (time adjustment) 0.895 0.150P(b2 = 1) = 0.368 3.8
    b
    3
    (area adjustment) 1.003 0.124P(b3 = 1) = 0.981 9.2
    b
    4
    (habitat adjustment) 1.208 0.092P(b4 = 1) = 0.024 18.6
    error σ2 = 0.256 9.9
    

    View Slide

39. OBS = a + b
    1
    R + b
    2
    T + b
    3
    A + b
    4
    H + error
    Ideally, a = 0, b = 1, error = 0, so that OBS = R + T + A + H
    • Road count & Habitat related biases dominate
    • EDR vs. MDD was thought to be most influential – now only 3rd
    • Habitat adjustment values are underestimated
    • There is still something left to be explained (~10%)
    • Interactions were not important
    39
    Model terms Estimate SE P-value % variance
    a –0.141 0.170P(a = 0) = 0.411
    b
    1
    (roadside count) 0.951 0.054P(b1 = 1) = 0.364 58.5
    b
    2
    (time adjustment) 0.895 0.150P(b2 = 1) = 0.368 3.8
    b
    3
    (area adjustment) 1.003 0.124P(b3 = 1) = 0.981 9.2
    b
    4
    (habitat adjustment) 1.208 0.092P(b4 = 1) = 0.024 18.6
    error σ2 = 0.256 9.9
    

    View Slide

40. OBS = a + b
    1
    R + b
    2
    T + b
    3
    A + b
    4
    H + error
    Ideally, a = 0, b = 1, error = 0, so that OBS = R + T + A + H
    Model terms Estimate SE P-value % variance
    a –0.141 0.170P(a = 0) = 0.411
    b
    1
    (roadside count) 0.951 0.054P(b1 = 1) = 0.364 58.5
    b
    2
    (time adjustment) 0.895 0.150P(b2 = 1) = 0.368 3.8
    b
    3
    (area adjustment) 1.003 0.124P(b3 = 1) = 0.981 9.2
    b
    4
    (habitat adjustment) 1.208 0.092P(b4 = 1) = 0.024 18.6
    error σ2 = 0.256 9.9
    • Road count & Habitat related biases dominate
    • EDR vs. MDD was thought to be most influential – now only 3rd
    • Habitat adjustment values are underestimated
    • There is still something left to be explained (~10%)
    • Interactions were not important
    40
    

    View Slide

41. Summary
    Spatially explicit predictions lead to altered ranking
    Forest species were under-, generalists were over-
    estimated by PIF
    Differences were driven by the habitat & roadside
    count assumptions (70%)
    The usual suspects (time & area) contributed to 13%
    41
    1
    2
    3
    4
    

    View Slide

42. Adjusted adjustments
    BBS data, same methodology, some adjustments:
    1. Use EDR instead of MDD where applicable
    (or provide a ‘lower bound’ in tables so that
    estimates can be recalculated if needed)
    2. Consider adjusting time component
    (set max as 0.66 instead of 1)
    42
    Time adjustment
    Area sampled
    

    View Slide

43. Different methodology
    Other data sources, revised methodology, with
    adjustments:
    1. Stratify based on broad land cover
    (open/forest)
    2. Incorporate off-road data
    Need to better understand & attribute mechanisms
    contributing to roadside count bias
    43
    Roadside count
    Habitat adjustment
    

    View Slide