Co-occurrence Pattern Learning Species Distribution Model (SDM) Quantifies Annual Reduction of American Coccinellids

Abstract

Aim: To predict annual distribution patterns and reduction rates of insufficiently observed species by using co-occurrence pattern learning and devising filling-in strategy to overcome structural and temporal inconsistencies in multi-source noisy data.

Idea: Although more than 10% of insects will face extinction in the coming decades, studies on their reduction rates that will form the basis for conservation strategies are still limited. This limitation is first due to the dominance of unstructured records available for invertebrates, secondly, to the inconsistencies among them, and thirdly, to the insufficiencies of them. While compelling to gather data across multiple sources, the small amount of data precludes deep filtering to handle structural and temporal inconsistencies among sources for time-series comparison. This is the first study to estimate annual reductions with machine learning from multi-sourced, presence-only, and small data, by overcoming its inconsistencies and insufficiencies. This study proposes and validates the following two novel strategies. (1) Co-occurrence pattern learning: By grouping low-quality, unreliable individual occurrence records into patterns, I validate that structural and temporal inconsistencies can be overcome without deep filtering. (2) Filling-in strategy: I propose a procedure for estimating population trends by filling in the prediction into the deficiencies of the collected yearly data to be evenly compared.

Location: 51 states of the USA and 6 provinces of Canada

Taxa: four ladybugs native to North America

Methods: In chapter 2, seven performance scores were used to evaluate the predictions on presence versus absence in the following three situations: (1) learning unstructured data to predict structured data or low-efficiecy data to high-efficiency data; (2) learning data before a particular year to predict after that year and vice versa; (3) learning 70% of multi-source data to predict the rest. During both the evaluation and generalization phases, a comparison was made between the performance of the co-occurence pattern using models and the environmental information using models, as well as with the commonly accepted benchmark.

In chapter 3, reduction rates and extinction status were estimated by ML's predicting the occupancy of species annually at all coordinates where species have appeared since 2007. In addition to that, the newly suggested approach's methodological reliability was verified, in comparison with pre-established methods. Furthermore, the reliability of the newly proposed method was validated by examining discrepancies in estimations under the following scenarios: variances in data extraction for pseudo-absence data points, variances in variable selection techniques, and the stochastic incorporation of missing or false information within the presence data.

Results: 1) The COP models' performance surpassed acceptable criteria for all validation steps and all species. They also ouperformed over the ENV models. 2) Reduction rates were 36.4% for H. parenthesis (2007–2021; VU), 29.7% for A. bipunctata (2010–2019; NT), 23.7% for C. novemnotata (2009–2018; NT), and 14% for C. trasversoguettata (2007–2018; LC). Additionally, the newly proposed approach was confirmed to possess strong methodological validity when compared to pre-established methods. In terms of reliability tests, the range of estimations from the new method did not misrepresent IUCN conservation status to a significant extent.

Conclusion: The combination of using co-occurrence patterns as variables and filling-in strategy enabled SDM to predict species' finer time scale distribution patterns and reduction rates by overcoming structural and temporal inconsistencies in multi-source data integrating considerable citizen science data. In North America, four native ladybug species have been declining steadily. This study suggests that ML developed with COP can integrate multiple-source data without filtering, allowing for the acquisition of more data, and that COP-based SDMs may be advantageous for predictions at finer temporal scales (and thus more precise than commonly used SDMs developed with environmental variables usually spanning over decades). This can aid in tackling the challenge in global conservation initiatives posed by rare and invertebrate taxa, which frequently face restricted data availability and are often underrepresented in conservation lists.