Bird population census: AI has started to check the bird population!

According to statistics from the World Wildlife Fund, the population of representative species in the world decreased by 68% from 1970 to 2016, and biodiversity continued to decline.
To protect biodiversity, we need to accurately analyze local ecological conditions and formulate reasonable ecological protection policies. However, ecological data is too complex and statistical standards are difficult to unify, making large-scale ecological analysis difficult to carry out.

Recently, researchers from Cornell University used deep learning to analyze 9 million sets of bird data and obtained the distribution data of wood warblers in North America, opening a new chapter in ecological data analysis.

Author | Xuecai

Editor | Three Sheep, Iron Tower

This article was first published on HyperAI WeChat public platform~

According to statistics from the World Wildlife Fund (WWF), from 1970 to 2016, the average number of 4,392 representative species and 20,811 populations in the world decreased by 68% , and global biodiversity is declining.

Figure 1: Average population changes of 4,392 representative species and 20,811 populations worldwide from 1970 to 2016

Protecting biodiversity requires accurate large-scale analysis of species distribution in relevant areas. However, due to the large amount of data and the lack of a unified statistical method , researchers are currently unable to accurately count the biodiversity (species richness, population size, etc.) and biological composition data (the status of a species in the local ecosystem) of a specific area.

Traditional species richness statistics require superimposing distribution maps of different species for modeling and prediction, or directly predicting through macroecological models. Regardless of the method, the inference results will be affected by the accuracy of the model , and the former will also be affected by the accuracy of the map.

Moreover, this prediction method has poor temporal resolution and cannot accurately judge seasonal changes in species distribution, let alone study the connections between species, which is not conducive to the formulation of ecological protection policies.

Deep learning provides an effective means for large-scale spatiotemporal research on biodiversity. Researchers from Cornell University in the United States combined the Deep Reasoning Network (DRN) and the Deep Multivariate Probit Model (DMVP) to develop the DMVP-DRNets model, which analyzed the spatiotemporal distribution of warblers in North America from 9,206,241 sets of eBird data and inferred the relationship between warblers and the environment and other species . The relevant results have been published in "Ecology".

This result has been published in Ecology

Experimental procedures

Dataset: eBird with covariates

The researchers used eBird data between 170°-60° W, 20°-60° N from January 1, 2004 to February 2, 2019 as the dataset for this study. After excluding duplicate data, there were 9,206,241 sets of eBird data , each of which included time, date, location, and all bird species observed.

Figure 2: eBird data of a group of long-tailed tits

The researchers also introduced 72 covariates , including 5 observer-related covariates, such as activity status, number of observers, observation time, etc.; 3 time-related covariates, mainly used to bridge the deviations between different time zones; 64 variables related to topography, such as altitude, coastline, islands, etc.

Model framework: decoder + latent space

This study used DRN based on DMVP for data analysis and prediction. This model contains 3 layers of fully-connected network decoder to analyze the correlation of input features, and two structured latent spaces to represent the association between species and between species and environment.

Figure 3: Schematic diagram of DMVP-DRNets model results

Finally, the DMVP-DRNets model outputs three ecologically relevant results through an interpretable latent space:

1. Environmental-related characteristics : reflecting the connections and interactions between different environmental covariates;

2. Species-related characteristics : The relationship between different species is reflected through the residual correlation matrix;

3. Biodiversity-related characteristics : such as the abundance and distribution of a particular species.

Model Evaluation: Comparison with HLR-S

Before putting the DMVP-DRNets model into large-scale use, the researchers first compared it with the HLR-S model based on spatial Gaussian processes . HLR-S is one of the most commonly used models in ecology to study the joint distribution of multiple species.

First, the two models were trained using 10,000 sets of eBird data. The HLR-S model took more than 24 hours to train, while the DMVP-DRNets model took less than 1 minute.

Table 1: Performance comparison between DMVP-DRNets model and HLR-S model

Subsequently, eBird data of different scales were analyzed and the DMVP-DRNets model outperformed the HLR-S model in 11 evaluation criteria , lagging behind the HLR-S model only in species richness calibration loss.

Experimental Results

Distribution: Appalachian Mountains

After analyzing the eBird data, the DMVP-DRNets model outputs a monthly distribution map of North American warblers with a spatial resolution of 2.9 km2. The distribution of different species of warblers in North America is very dynamic, with different distribution hotspots every month. After superimposing the monthly distribution maps, the researchers found that the Appalachian Mountains are the area with the highest species diversity of warblers.

Figure 4: Distribution map of warblers in North America

a: Distribution of maximum species richness of wood warblers across North America

b: The main distribution area of ​​the wood warbler in North America

At the same time, the researchers also found hot spots for the distribution of warblers at different migration stages. During the pre-breeding migration period, warblers were mainly distributed near the Appalachian Mountains in Ohio, West Virginia and Pennsylvania. After breeding, the northern Appalachian Mountains were the most distributed area for warblers.

Figure 5: Distribution of wood warblers during the pre-breeding migration period (a) and the post-breeding migration period (b)

Warblers - Environment: Water, Land and Seasonal Preferences

Furthermore, the researchers used the DMVP-DRNets model to analyze the interactions between warblers and the environment in the northeastern United States.

First, the researchers were able to roughly distinguish the preferences of different warblers for aquatic and terrestrial environments. Then, they found that different warbler species had different preferences for environments during the breeding season. Blue-winged yellow warblers, northern yellow warblers, and yellow-throated warblers, which prefer aquatic environments, roosted closer during the breeding season, while pine warblers would roost closer to other species associated with pine forests, such as brown-headed nuthatches and red-headed woodpeckers.

The distribution of different warblers changes with the seasons . Most warblers roost in groups during the post-breeding migration period, while palm warblers migrate later in the fall. Pine warblers and yellow-rumped whitethroats roost year-round in the northeastern United States.

Figure 6: Correlations between warblers during the breeding season and the environment and other species

Figure 7: Correlations between warblers and the environment and other species during post-breeding migration

Interspecies associations: competition and cooperation

Warblers show different relationships with other species during the breeding, non-breeding and migratory seasons.

During the breeding season, warblers mainly defend their own habitats, with weak associations with other species. There is even a negative association between species that have similar habitats and are more aggressive, such as the Black-naped and Orange-tailed Warblers.

During migration, most warblers showed strong positive correlations with each other and with other species in the forest, which is consistent with observations that warblers form mixed migration groups with other species such as red-eyed green cuckoos and black-crowned chickadees.

During this period, the warblers had poor relationships with predators such as the great-winged buzzard, barred hawk, chicken hawk, and red-shouldered buzzard, and the negative correlation coefficient between the two was high.

Figure 8: Correlation coefficients between warblers and other species during the breeding period (a) and post-breeding migration period (b)

The above results show that the DMVP-DRNets model can make accurate judgments on the distribution of warblers in different periods, and can infer the connection between warblers and the environment and other species, providing a basis for formulating ecological policies.

AI "Bird Population Census"

In addition to data analysis, data collection is also an important part of ecological research. Unlike plants, birds are highly alert and move quickly, and some species are small, making it difficult to observe them accurately.

Traditional methods rely on telephoto cameras, high-power telescopes and stationary cameras to observe birds from a distance. Although this method avoids disturbing the birds, it requires a lot of manpower and material resources, and requires observers to have considerable knowledge of ecology and taxonomy.

Through deep neural networks, AI can perform efficient image and sound recognition, providing a new method for bird observation . Deploy audio and video recording equipment in the main activity areas of birds. The equipment can upload the recorded data to the server, and then analyze the data through AI to extract the information in the audio and video, and finally obtain the distribution of birds in this area. This method has been widely used by the State Forestry and Grassland Administration in parks, wetlands and ecological reserves.

Figure 9: Bird smart monitoring system deployed in the Yellow River Delta

At the same time, this skill of AI can also reduce the workload of scientific researchers. AI can eliminate the interference of background and noise, focus on the characteristics of the image, and quickly solve problems that ecologists find difficult to judge. For example, in the photo below, if you don’t have any bird knowledge, it is difficult to quickly judge the number of chicks from the complex feathers.

Figure 10: Photo of a nest of chicks. Can you tell how many chicks there are in the picture?

AI is being widely used in bird activity monitoring and bird distribution analysis, building a complete system for bird research from the bottom up to achieve a "bird census" in a specific area. I believe that with the help of AI, we can have a more thorough understanding of the ecosystem, formulate ecological policies that are more in line with local conditions, gradually restore the earth's biodiversity, and protect our home planet.