Soil visible and infrared (vis—IR; 400–20,000 nm) spectroscopy is an accurate, rapid, and cost-efficient method for characterising soil properties. Multivariate regression and machine learning can be used to ‘learn’ the complex relationship between soil properties and the spectra and used to make estimates on unknown samples and their uncertainty. Recent advances in soil spectroscopic modelling have exploited state-of-the-art deep convolutional neural networks for estimating soil properties, and transfer learning to generate accurate local estimates, e.g., for site-specific soil management.  However, some challenges remain, and research is ongoing on methods that (i) can provide a balance between good predictability and explainability and (ii) can localise spectroscopic models. This seminar will describe current developments to overcome those challenges using examples from the literature and our own research.

Biography: Dr. Zefang Shen’s PhD was on mechanical engineering and development of robotic exoskeletons for rehabilitation. His current research is the development and application of novel machine learning for optimal, transferable, and interpretable soil spectroscopic modelling.

REGISTRATION