Emails: qiaomin.chen@uq.edu.au or scott.chapman@uq.edu.au

Improving crop productivity with a low environmental impact is highly needed to fulfill a sustainable food demand in face of increasing population and irreversible climate changes. Dynamic biomass prediction is critical for crop growth monitoring and subsequent crop production management. Although methods for dynamic biomass prediction have been progressing from descriptive model to mechanism model and process-based model, these models are unsuitable for large-scale and wide-range applications due to parameterization issues.

Developments in remote sensing and machine learning technologies lead to data-driven models that overcome difficulties related to parameterization, but extending applications of these models in new environments are usually restricted to training data collected from limited environments. To address these challenges, here we present a biophysics-informed machine learning framework model that incorporates biophysical information into training data and learning process.

The resulting framework model provides plausible and reliable prediction of wheat biomass dynamics from sowing to harvest at different spatial scales using daily weather data and limited spectral data at corresponding spatial resolution during crop growth season.

Our proposed model provides plausible prediction of dynamic biomass from near weekly frequency of spectral observations combined with daily weather data on independent simulation datasets, with seasonal relative mean absolute error (RMAE) within 20% and of ~7% on average for more than 250 thousand scenarios. While validating on experimental datasets, it offers a prediction accuracy with seasonal RMAE of 20-26% for trial-level biomass and of 12-19% for plot-level biomass.

This model offers an effective and efficient manner for multiscale dynamic biomass prediction, which not only bridges agricultural remote sensing and plant high-throughput phenotyping but also demonstrates the potentials of combining crop modelling and machine learning for crop growth monitoring.