Deep Diving into deep learning and AI: Experiences and lessons learned

Around a decade ago Deep Learning emerged as the new frontier of computer vision, eclipsing the performance of previous methods for image classification, semantic segmentation and object detection/segmentation. At Manaaki Whenua Landcare Research we have successfully trialled deep learning for all three of these uses and applied it to many domains, including detecting and monitoring invasive predators (classification), mapping land cover, land use, wetlands and wilding conifers (semantic segmentation), and mapping individual trees (object segmentation). More recently we have “stretched” the use of deep learning still further by using it automate the generation of landscape-scale polygons for soil mapping and other uses, by teaching the model to perform this previously manual task.

While investigating deep learning’s use in these domains, we have also been developing inhouse tools, including user-friendly software for running and fine-tuning image detection models (CamTrapNZ), and for applying semantic and object segmentation to remote sensing imagery (DeepSeg). While these tools have been extremely valuable, they are still relatively immature. Further, the expertise required to use them effectively is largely contained within a small pool of specialised data science/remote sensing researchers. To overcome this, we are moving to a new stage where we are maturing the tools to make them more accessible to the wider research community, as well as developing online training material to enable uptake by non-data scientists. A key challenge for this effort is the wide range of abilities and expectations within the research community (our “customers”), requiring a range of solutions. Adding to this is the constant evolution of deep learning itself, meaning both educational material and tools quickly become obsolete.

More recently, Manaaki Whenua has embraced Generative AI as a highly promising addition to the toolset for a range of applications including information extraction, scenario visualisation and simulation. Unlike deep learning, which had obvious application to an existing set of problems and was therefore adopted in a fairly systematic manner, the use of GenAI is more general and speculative, and is therefore emerging organically throughout the organisation. Key to the successful adoption of GenAI is the sharing of knowledge and experiences across an extremely diverse range of projects.

In this presentation we will share our experiences over the past five years, as well as our near-term plans for maturing our tools, upskilling Manaaki Whenua’s researcher base and sharing project models, data, software and knowledge.

ABOUT THE AUTHOR

Brent Martin is a senior data scientist and machine learning specialist at Manaaki Whenua – Landcare Research. His 38-year career spans both academic research as a senior lecturer at Canterbury University, as well as software engineering and R&D roles in various commercial companies from local software house Jade to Google NY.

Brent’s research in AI and machine learning includes developing new ML classification algorithms, applying ML to real-world problems such as electricity demand forecasting and internet search engines, research and development in Intelligent Tutoring Systems, developing social network analysis and anomaly detection techniques for criminal investigation, and applying deep learning to environmental computer vision problems.

At Manaaki Whenua, Brent has applied deep learning to problems for a range of clients including government departments such as MfE and MBIE, and regional councils throughout New Zealand, as well as internal projects aimed at automating and upscaling some of the services Manaaki Whenua already provides. More recently this research has included applied generative AI. Brent holds a PhD in Computer Science (artificial intelligence) from the University of Canterbury, New Zealand.

-----

For more information about the eResearch NZ / eRangahau Aotearoa conference, visit:
https://eresearchnz.co.nz/