Our Research Methods
In the year 2025, TAIAO developed new methods in data science and artificial intelligence through three main research aims - Machine Learning for Data Streams and Time Series, Machine Learning for Weak signals, Anomalies and Extreme Events, and Deep Learning. The team have achieved so much in this space, finding new ways to advance environmental research through AI.
1. Machine Learning for Data Streams and Time Series
This year has seen significant progress in data stream learning and continual learning, with advances spanning regression methods, decision tree algorithms, concept drift, real-world electricity data, and AutoML. Our contributions also extended to the organisation of tutorials in leading international venues and strengthening the visibility of data stream research. Below are some examples of our work.
Regression and AutoML in Data Streams: We advanced streaming regression by introducing adaptive methods that boost predictive accuracy as data evolves. One line of work developed an efficient adaptive filtering approach for handling continuous, changing streams, while another created a dynamic AutoML framework that automatically selects the best regression models in real time as conditions shift.
Anomaly Detection and Novel Algorithms: We strengthened streaming anomaly detection by adapting Isolation Forest to work effectively on continuous, evolving data, enabling robust identification of unusual patterns in dynamic environments (J Liu, F Liu, Cassales, Pfahringer, Bifet; PAKDD, 2025). In parallel, we advanced decision‑tree learning by introducing methods that leverage model plasticity, improving how incremental trees adapt to changing data over time ((Heyden, Gomes, Fouché, Pfahringer, Böhm; ECML PKDD, 2024).
Concept Drift and Real‑World Data: We deepened understanding of concept drift by analysing how recurring drifts emerge and behave in practical settings. This included a systematic study of recurrent drift patterns in real‑world streams, as well as an applied investigation of evolving electricity‑pricing data in New Zealand, demonstrating the value of stream‑learning methods for critical infrastructure.
Continual Learning: We advanced continual learning research through both new benchmarks and architectural insights. A new dataset for distinguishing real and synthetic image‑classification tasks enabled more realistic evaluation of continual learning behaviour, while an analysis of Kolmogorov–Arnold Networks revealed how these emerging architectures forget differently from traditional MLPs, offering fresh perspectives on catastrophic forgetting.
Research Milestones: Yibin Sun completed his PhD thesis on improving ensembles and prediction intervals for data‑stream learning. Lea Casse is developing quantum machine learning models for river‑level prediction and parametric flood insurance using TAIAO datasets, with the project having reached the final phase of the Global Industry Challenge.
2. Machine Learning for Weak signals, Anomalies and Extreme Events
We have also made some significant progress in our research of Machine Learning for Weak signals, Anomalies and Extreme event. Below is some of our work which is divided into 4 different categories including Anomaly Detection, Climate Forecasting, Air Quality Monitoring, and other Anomaly Detection directions.
Anomaly Detection: Anomaly detection in live trajectory data is important but traditional anomaly detection methods struggle with dynamic and evolving trajectory patterns that require the ability to adapt with increased traffic, geopolitical events, and global warming. TAIAO has leveraged a continual learning method that enables that model to learn from new data continuously and recognise specific behaviors dependent on position and recent movements. The method involves the implementation of an adapter-based framework known as Continual Learning for AIS Anomalies (CLAISA).
Climate Forecasting: Climate events exhibit intricate, multivariate dynamics driven by region-wise interactions which are key to the global food supply. However, accurately forecasting climate events has been a major challenge. TAIAO has proposed a Hierarchical Graph Neutral Networks (Hierarchical GNN) that will enhance the immersion of key global physics onto projections via global trend retention.
Air Quality Monitoring: Air‑quality monitoring with low‑cost sensors often requires predicting time‑varying features at entirely unseen locations using only surrounding context and no exogenous priors. This allows estimates in regions without direct observations, reducing deployment costs and informing optimal sensor placement to better capture extreme conditions.
Other Anomaly Detection Directions: Antarctic sea‑ice behaviour has shifted dramatically over recent decades, with long‑standing increases giving way to a sharp decline after 2014, underscoring the need for better models that can explain and predict these anomalies. We developed new graph‑neural‑network approaches that identify key drivers and forecast sea‑ice anomalies weeks ahead.
3. Deep Learning
Our deep learning has been used for a range of applications including the flood prediction tool, and the Forest flows project. Deep Learning attempts to not just predict, but also explain the behaviour of the trained models. Below are some of the deep learning work that our TAIAO team is leading or supporting.
The ongoing work on continual learning is spear-headed by Dr. Yaqian Zhang who is developing an alternative to cross‑entropy loss was developed using a simple reward function, supported by theoretical results and empirical evidence of its effectiveness.
New PhD student Rafia Malik, who began in March 2025, is exploring Vision‑Language Models for remote‑sensing change detection, focusing on predicting New Zealand shoreline evolution by combining high‑resolution aerial imagery with more readily available satellite data to achieve aerial‑level performance from low‑cost inputs.
The Aotearoa species‑classifier apps have been updated using newer, larger iNaturalist datasets, with source code, trained models, and a new journal publication accompanying the release.
Adaptation work for night‑time imagery from Maungatautari Sanctuary Mountain has delivered strong performance gains for this specialised setting, though this component is still awaiting publication.
The University of Waikato
University of Canterbury
The University of Auckland
Victoria University of Wellington
MetService
Beca
