Model development and automated data collation for global biodiversity indicators
File(s)
Author(s)
Cornford, Richard Edward
Type
Thesis
Abstract
Biodiversity databases and models are vital to monitoring ecological change and identifying effective actions to improve the state of the natural world. Current models and databases must continue to develop so that policy can be informed by up-to-date, globally representative evidence.
In this thesis, I first analyse data in the Living Planet Database (LPD), assessing how delayed impacts of land use and climate combine with socioeconomic drivers to influence abundance trends of mammals and birds worldwide. Model selection identifies delayed responses to both environmental change drivers, with synergistic declines where past warming and recent land conversion combine. I also find contrasting effects of management/protected areas (positive) and exploitation (negative) on population growth. Urgent action addressing the multiple drivers of biodiversity loss is therefore needed to safeguard nature. To evaluate the reliability of global projections from the selected models I use spatially blocked cross- validation. Although the models have strong in-sample performance, accurately predicting abundance change in new locations is a challenge, highlighting the need for more, more-representative, data on population abundance.
Collating such biodiversity data from the literature is typically a slow, manual process. I assess the potential for using text-mining to increase the rate at which published ecological data is integrated into biodiversity databases. I show that machine-learning classifiers trained using the abstracts of articles contributing to the LPD and the PREDICTS database can distinguish relevant from irrelevant articles with over 90% accuracy. Using such classifiers speeds up the identification of recently-published relevant data compared to a current manual protocol...
In this thesis, I first analyse data in the Living Planet Database (LPD), assessing how delayed impacts of land use and climate combine with socioeconomic drivers to influence abundance trends of mammals and birds worldwide. Model selection identifies delayed responses to both environmental change drivers, with synergistic declines where past warming and recent land conversion combine. I also find contrasting effects of management/protected areas (positive) and exploitation (negative) on population growth. Urgent action addressing the multiple drivers of biodiversity loss is therefore needed to safeguard nature. To evaluate the reliability of global projections from the selected models I use spatially blocked cross- validation. Although the models have strong in-sample performance, accurately predicting abundance change in new locations is a challenge, highlighting the need for more, more-representative, data on population abundance.
Collating such biodiversity data from the literature is typically a slow, manual process. I assess the potential for using text-mining to increase the rate at which published ecological data is integrated into biodiversity databases. I show that machine-learning classifiers trained using the abstracts of articles contributing to the LPD and the PREDICTS database can distinguish relevant from irrelevant articles with over 90% accuracy. Using such classifiers speeds up the identification of recently-published relevant data compared to a current manual protocol...
Version
Open Access
Date Issued
2022-07-31
Date Awarded
01/02/2023
License URL
Advisor
Purvis, Andy
Freeman, Robin
Sponsor
Natural Environment Research Council (Great Britain)
Grant Number
NE/R012229/1
Publisher Department
Life Sciences
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)