Target season: prebreeding


<< Back to BirdFlow model collection

Quick summary

Model quality assessment (please read limitations document)

criterion assessment
Correlation with Status and Trends acceptable
Model calibration good
Simulated routes: straightness good
Simulated routes: number of stopovers acceptable

Simulated route map (n = 10)

Route characteristics

routes straightness n_stopovers
simulated (n = 100) 0.90 3.4
real tracking data 0.93 2.5

Technical details

Internal model filename

buwtea_2021_100km_obs1.0_ent0.0020202_dist0.0080808_pow0.5

Grid search setup

Parameters for grid expansion

  • obs_prop: 0.95, 0.975, 0.99, 0.999, 0.9999

    • proportion of all weights allocated to the observation weight.

    • generally, we want the model to closely follow eBird Status and Trends distributions

  • de_ratio: 2, 4, 6, 8, 10

    • specifying this ratio (dist_weight / ent_weight) is useful because this proportion strongly affects simulated routes

  • dist_pow: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9

    • this hyperparameter controls the relative cost of many short movements vs. few long movements

Grid size

  • 5 x 5 x 9 = 225

Hyperparameters calculated from de_ratio and obs_prop

  • dist_weight

  • ent_weight

Hyperparameters passed to model fitting process on GPU

  • dist_weight

  • ent_weight

  • dist_pow

Grid search space

Parameters for grid expansion (left); hyperparameters passed to model fitting process (right)

Hyperparameters

de_ratio obs_prop dist_pow dist_weight ent_weight obs_weight
4 0.99 0.5 0.0080808 0.0020202 1

Model selection

  • model selection target season: prebreeding

  • desirability criteria: real_tracking

    • this desirability method co-optimizes the following

      • overall model calibration score (which itself is a co-optimization of row and column PIT scores and PIT 95% CI)
      • “end traverse correlation” with eBird Status and Trends data (start with modeled distribution at season start, forward-predict distribution to season end, then compare with S&T distribution at season end)
      • average number of stopovers in simulated tracks targets average number of stopovers in real tracks
      • average straightness of simulated tracks targets average straightness of real tracks

Correlation and calibration

end_traverse_cor pit_row pit_col pit_in_95
0.973 0.0955 0.0438 0.881

Desirability

Overall desirability rank within grid search: 1

desirability sub-scores for this model
d_pit_row d_pit_col d_pit_in_95 pit_d etc_d str_d nso_d overall_des
0.71 0.925 0.867 0.829 0.971 0.965 0.951 0.927

Desirability sub-scores for this model vs. all models in this grid search


<< Back to BirdFlow model collection