Machine Learning Analysis of Seasonal Bird Communities Along Urbanization Gradients

Abstract

Introduction: Urbanization is among the strongest factors of land cover alteration and affects bird communities in terms of vegetation cover, food supply, noise environment and nest-building substrates from natural or agricultural areas towards cities. The study of changes of species composition and diversity parameters at such gradients and over seasons usually relied on field sampling and identification of calls manually. Methodology: In the present paper we discuss theoretically the use of machine learning algorithms for analysis of seasonal data about bird communities recorded along urbanization gradients. Supervised classifiers, unsupervised clustering techniques and deep learning models for acoustic data are considered, along with a proposed conceptual framework of analysis based on the combination of remote-sensing derived gradient variables, field survey data, features building and modelling of classification into diversity outcomes. Results: From the literature review, it is clear that the ensembles and deep learning classifiers usually perform better than simple classifiers, given high richness and high background noises; that seasonality affects the performance of the classification system; and that including habitat gradient covariates in addition to acoustic features results in an effective separation between the urban-tolerant and urban-avoiding species. Conclusion: Machine learning provides a scalable concept for studying bird community seasonal dynamics in fragmented and urbanized environments, taking into consideration data imbalance, seasonality, and covariates.