PhD Studentship: Improving Artificial Intelligence through Neuroscience

About the Project

Project details

The forestry industry in New Zealand, with NZ$5.89 billion in annual export revenue, is under pressure due to cyclones and tropical storms. Sales are arranged ahead of harvest using estimates of wood production constrained by UAV and Airborne laser scanning (LiDAR) surveys. Due to weather conditions, it is unsafe for several days to months after a cyclone or tropical storm to use these technologies and/or visit the affected area to evaluate storm-related tree damage. Therefore, to support sales planning and the safety of foresters working in the field, there is a need to quickly quantify the damage to forest plantations after a cyclone or a tropical storm. There is unrealised potential in using multi-modal computer vision methods that synthesis multi-source Earth Observation data for this application. 

A key challenge lies in making adaptable approaches as it is unknown which type of data will become available first. Pre-storm LiDAR data are available from open Topography. Examples of post-storm relevant spaceborne missions include, BIOMASS L-band Synthetic Aperture Radar (SAR), NISAR SAR, and GEDI LiDAR, that can capture structural information about the forests. Additionally, missions like Landsat optical, Sentinel-1 SAR and Sentinel-2 optical enable time-series observations, important as spectral signatures changes after a storm disturbance. Commercial datasets (e.g. Umbra Lab’s 25m high-resolution SAR) will be explored if access can be secured but are not essential to feasibility.

Ideas with significant potential value for this application include: Investigate how multimodal satellite Earth Observation and machine learning can be used to quantify cyclone and storm damage in plantation forests. The core focus could be on integrating pre-storm LiDAR with post-storm satellite SAR, LiDAR and/or optical imagery to enable rapid, safe, and scalable assessments of damage. 

Candidate methods for temporal modelling and anomalies detection, which are likely to occur at affected areas, could include deep learning (e.g. Long Short Term Memory - LSTM), statistical baselines (e.g. Autoregressive Integrated Moving Average - ARIMA, Kalman filters) and transformers (e.g., spatio-temporal visual transformers - ST-ViT). Fine-tuning foundation models like Cambridge TESSERA and IBM-NASA Prithvi could map the extent of damage to wood stocks. Road maps can be queried (e.g., from PostGIS) and an agent-based reinforcement learning system can developed to identify and optimise safe journeys to reach affected forest plantations. 

The specific research objectives will be co-developed with the doctoral student in alignment with Interpine Group Ltd’s priorities. Method developed will be compared systematically to assess trade-offs in accuracy, interpretability, and computational efficiency. Uncertainty quantification shall be incorporated to strengthen robustness. The development of an interactive visualisation application shall support decision making and planning. 

The technical insights from this project have direct applicability to UK commercial forestry and Net Zero ambitions due to increasing tree planting programs coupled with increased extreme weather events.

Please direct project specific enquiries to: Enquiries shall be adressed to Milto Miltiadou (m.miltiadou@exeter.ac.uk) Please ensure you read the entry requirements for the potential programme you are applying for. To Apply for this project please click on the following link - https://www.exeter.ac.uk/study/funding/award/?id=5732

Advert information