UMass predicts bird migration using computer modeling and eBird dataset

This article was first published from the HyperAI Super Neural WeChat public account~

Content overview: Recently, a new predictive model BirdFlow was released in the journal "Methods in Ecology and Evolution" of the British Ecological Society, which can solve one of the most difficult challenges in biology: accurately predicting the trajectory of migratory birds. While the model is still being perfected, the researchers say it could be open to the public and put into use within a year. This article is an introduction and interpretation of this study.
Key words: BirdFlow, nature protection probability model

Bird migration is a fascinating natural phenomenon. It is understood that nearly one-fifth of the world's bird species migrate regularly for breeding and wintering. In ecology, the study of ecological laws such as bird migration routes is of great significance for protecting endangered bird species, maintaining ecological balance, and preventing the spread of epidemics.

Predicting bird migration has become more difficult in recent years due to factors such as global climate change and human activities. Recently, Miguel Fuentes, a graduate student at the University of Massachusetts Amherst, and Benjamin M. Van Doren of Cornell University published a new probability model BirdFlow in the journal "Methods in Ecology and Evolution", which uses computer Modeling and the eBird dataset to accurately predict the flight paths of migratory birds.

Paper address:

https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.14052

The researchers used the relative abundance estimates (abundance esitimates) generated by the eBird Status & Trends project to simulate bird movement, but there is also a problem. The past relative abundance information can only show the location range of birds on a weekly basis, and cannot track individual. Therefore, in this study, the researchers focused on solving this problem, and the key process is shown in the figure below:

• Data Preprocessing: Preprocessing relative abundance estimates to generate weekly population distributions;

• loss function: specify a loss function to score potential models using a proxy for weekly distributions and energy costs;

• Model Structure: select a model structure;

• Trained Model: optimize the loss function through a numerical process to select the best model parameters;

• Validation: Calculate the average log-likelihood and PIT values ​​of real birds to validate the trained model.

BirdFlow Modeling Overview

The researchers used ebird R to download relative abundance estimates from eBird Status & Trends for 11 bird species for which GPS or satellite tracking data was available.

eBird Status and Trends：

https://science.ebird.org/zh-CN/status-and-trends

As a next step, the researchers defined a loss function based on the weekly population distribution derived from eBird Status & Trends, the energy cost of movement of birds between different locations, and an entropy regularization term.

Before optimizing the loss function, it is necessary to specify a model structure. Here, the researchers proved that it is reasonable to limit the optimization process to search only on Markov chains. Therefore, they modeled the motion of the birds as a Markov model and performed optimizations, including using Markov chain parametrization and optimization algorithms.

After the above steps, the researchers obtained a trained model and verified it.

BirdFlow Verification Process

The verification process is divided into three parts: hyperparameter grid search, entropy calibration, and k-week forecasting. The specific process and test results are as follows.

Hyperparameter grid search

During the model validation phase, the researchers performed a hyperparameter grid search and used the search results to study two problems.

First, the researchers explored the effect of the entropy regularization term and distance exponent on model quality through an ablation study. The results of the ablation study are shown in the figure below. It can be seen that all BirdFlow models perform better than the baseline model that only includes the relative abundance of birds.

Second, the researchers explored the model's sensitivity to hyperparameter selection through two hyperparameter selection methods. The results of the experiment are shown in the figure below, for most birds, the model using the LOO parameter (selection of validation track data for other birds) performed as well as the model using the tuned parameter (use of validation track data for this bird). where performance is measured as the mean log-likelihood of the 1-week transition.

entropy correction

The figure below demonstrates the effect of entropy regularization on model calibration. Stochastic probability integral transform (PIT) histograms of the five versions of the American Woodcock model at different entropy weights showing how well the trained model predicts the woodcock's week-long east-west orientation.

It can be seen that the histograms are almost consistent, indicating that the calibration of the model is well behaved.

k-week forecast

Figures 5 and 6 show the model performance at different forecast times (in weeks). The researchers identified the best-performing model from the hyperparameter grid search and evaluated the performance of the best model relative to the baseline model from 1 to 17 weeks.

Figure 5(a) shows the results for each bird species. It can be seen that as time increases, the performance capabilities of the best models for each bird get closer and closer to the baseline model. Figure 5(b) shows the comparison of the gap between the woodcock tuned model, the LOO model and the baseline model. It can be seen that during the prediction time, the performance of the tuned model and the LOO model is better than that of the benchmark model.

After the above experiments, the researchers found that BirdFlow can use eBird's weekly relative abundance estimates to accurately infer the migration paths of individual birds, and the results showed that BirdFlow's prediction results were much better than the baseline model.

Based on this result, the researchers believe that in addition to studying the natural phenomenon of bird migration, the BirdFlow model may also be used to study other phenomena, such as the stopover behavior of birds and their responses to global changes.

However, despite the success of the BirdFlow model, some researchers in North America and Europe have questioned its use of the eBird database, arguing that bird watching is not a rigorous method of collecting data. In response, the BirdFlow researchers say the team is considering incorporating more data, such as satellite or GPS tracking of bird locations.

AI may become the god of nature protection

The emergence of the BirdFlow model means that humans have opened up a shortcut to machine learning in bird migration-related research. Although it is still in the early stage, and there is still a certain distance from landing applications such as nature protection, this research undoubtedly reveals an important trend that AI is being widely used in the field of nature protection.

PAWS, developed by researchers at Carnegie Mellon University, can generate a route for police to patrol for poachers; Merlin, developed by Cornell University, can identify species based on the songs and images of birds; and TrailGuard AI developed by Resolve It can protect wild animals by identifying images of suspected poachers and issuing alerts.

The importance of natural ecosystems to humans is self-evident, and the protection of ecosystems is imminent. As time goes by, what new role will AI play? Welcome everyone to diverge your thinking and discuss in the comment area.