Adaptive Multi-Source Transfer Learning Approaches for Environmental Challenges

点击翻译按钮获取中文摘要

Improvements in environmental analytical sensing technologies has led to a dramatic increase in the quantity and quality of data generation. This has resulted in an increased complexity of patterns that the data contain. This situation thus demands advanced machine learning (ML) approaches beyond traditional statistical and physical models, in order to understand how the Earth's climate and ecosystem have been changing and how they are being impacted by human behaviours. Many environmental problems have begun actively seeking input from the ML community, due to the powerful data fitting abilities of ML algorithms in various scenarios without human intervention. For example, the Plymouth Marine Laboratory has recently been developing data-driven approaches to automate coastal observation and marine management. It effectively lowers the cost of environmental observation and records past and future change in the ocean climate at an unprecedented scale. This is only the beginning of witnessing the success of AI/ML to help people understand nature and tackle environmental challenges. Many more are still laborious and lack of accurate modelling approaches. One key obstacle of having ML contribute to environmental problems is the inconsistent data quality and quantity across regions. Many problems suffer the difficulty that, the data from the region of interest is insufficient for building an accurate learner. However, relevant data can be available from other regions, although there may exist distribution differences, feature mismatches, etc. This project is thus motivated to study and develop transfer learning (TL) approaches for such environmental problems, which can transfer the useful knowledge from various regions (i.e. multi-source data domains) to build an accurate predictive model for the region of interest (i.e. the target domain).To successfully transfer knowledge from related data domains to the target domain, two specific learning challenges need to be addressed: class imbalance and concept drift. The data distribution can be very skewed in some natural events, such as flooding, earthquakes and heatwaves. This is called class imbalance and leads to poor generalization of a learner on the minority events. Environmental data is often collected over time, so that distribution changes in data may happen at some point. This is called concept drift and can deteriorate the learning performance significantly.This project aims to tackle these two fundamental learning challenges by developing advanced TL approaches. They will be used to train accurate models for two concrete environmental problems - early ice jam prediction and multi-plant wastewater inflow prediction, through close collaboration with the partner. Pioneering work will be conducted through four carefully designed work packages (WPs), each of which aims at one proposed objective. - WP1: TL for class imbalanced data.- WP2: TL for time drifting data.- WP3: Early ice jam prediction using TL.- WP4: Wastewater inflow prediction using TL. The above will lead to innovative solutions that add values to the current EPSRC's world-class impact targets with demonstrable case studies. In the meantime, they will not be limited to these two applications. They have the potential to benefit a wide range of environmental problems, such as climate pattern discovery and flood risk estimation, and even other fields, such as agricultural planning, transportation and manufacturing. This project will recruit one PDRA. Some key activities include two-way research visits, regular team meetings, research workshops and dissemination activities.