Interpreting complex ecological patterns and processes across different scales using artificial intelligence

Abstract

Ecology, as a discipline, attempts to understand the patterns and drivers of those patterns in natural systems. Traditional field-based approaches often measure patterns at specific spatial or temporal scales within particular ecological hierarchies, such as transect-based population surveys or quadrats sampling for community composition. Such narrow focus limits the transferability of quantitative methods across habitats or hierarchies, hindering the identification of general ecological patterns at different scales. The lack of standardized, comparable methods which can be found in other disciplines (e.g. stoichiometric equations in chemistry) is a major hurdle to seeking ecological generalities and is partly due to the lack of tools to address scale mismatches among data properties and field observations. While advances in Artificial Intelligence (AI) have revolutionized data collection, analysis and interpretation of complex features, ecologists have yet to fully harness AI to tackle these scale challenges, especially those posed by complex ecological patterns and processes.

This thesis bridges the gap between field observations and the empirical understanding of ecological systems by applying AI models at different spatial scales and hierarchical levels to study ecological complexity. Firstly, a Python package HSC3D, was developed to quantify habitat structural complexity (HSC) at the community level. Built with machine learning algorithms and novel, scale-invariant metrics, this package provides more precise, scale-invariant representations of HSC than traditional approaches, which can be applied to a variety of habitats in different ecosystems. Secondly, at the population level, deep learning (DL) models were benchmarked to inform the best practices in quantifying distribution patterns of intertidal mussels over large spatial scales. By classifying and segmenting mussels at the pixel level, these DL models reduced both false positive samples and user workload, providing an efficient and accurate transferable approach for large scale population surveys. DL models, in combination with novel species annotation strategies, were applied to address a perennial challenge in the identification of populations of invasive mussels from natural populations. This methodology was also adapted for use on mobile phones, allowing users to conduct in-situ surveys and post-hoc analysis, and can be modified for broad accessibility, facilitating immediate application in ecological fieldwork. Lastly, extending to the individual level, explainable AI techniques were employed in the detection of subtidal sea cucumbers. By visualising morphological features important for model predictions, XAI verified that AI decisions align with biologically meaningful traits, thus empowering transparent, individual-based assessments that can also be generalised to other ecosystems.

These case studies, from community, population to individual ecological hierarchy levels, demonstrate the power of AI in bridging scale mismatches between microscopic interactions (i.e. organismal morphological features) and macroscopic ecological processes (i.e. habitat variations and population dynamics) that traditional field-based approaches struggle to capture. All the developed methods are open-source, offering ecologists readily adaptable tools for improved accuracy in ecological research at multiple scales across different ecological systems. Finally, this thesis cautions that the future development of ecological AI models must be balanced with empirical ecological theories, field observations and human expertise to ensure robust and meaningful insights across diverse systems and scales.