Deciphering Head Rice Yield: Interpretable Machine Learning Models for Rice Milling Quality Predictions in Australia

Head Rice Yield (HRY) holds significant weight in determining financial returns for rice growers and processors as it measures the proportion of intact grains that remain after milling. However, determining HRY pre-storage is challenging due to the high moisture content of rice during harvest time. This can hinder efficient post-harvest management and grower feedback. This challenge is further intensified by the variability of HRY, influenced by a confluence of genetic, agronomic, and environmental factors. Advances in empirical modelling and interpretable machine learning techniques present new avenues for addressing these issues, facilitating the creation of models that not only accurately predict HRY but also shed light on the factors affecting crop outcomes.
The accuracy of empirical models depends on dataset construction and pre-processing. A comparison of two methods for integrating in-season climate data—using estimated phenology versus defined time intervals—assessed the impact of various aggregation stages on HRY model accuracy. The findings were conducted with a ten-season dataset of the primary Australian medium grain variety and indicated superior model accuracy utilising time-based aggregation. Model precision improved with increased aggregation stages for both methods, with the most effective configuration emerging as eight two-week (14-day) intervals working back from the date of harvest.
The subsequent study enriched the initial time-based aggregation dataset by incorporating satellite-derived vegetative indices and soil features. Model training involved a four-season subset with field spatial data available. Models combining climate data, vegetative indices, and soil characteristics outperformed those based solely on climate or vegetative indices. Within the combined model, harvest grain moisture emerged as the most important factor, overshadowing the collective influence of climate and vegetative indices-based features.
The initial studies explored the utility of HRY prediction models for grain elevator application, targeted at improved storage management. The third study expanded this work by developing in-season HRY forecasts using data available at various pre-harvest intervals. Results demonstrated that forecast precision improved closer to the harvest date, with combined climate and vegetative indices models outperforming existing industry methods up to eight weeks before harvest. This improvement suggests potential enhancements in planned post-harvest logistics and budgets for milled rice sales.
The final analysis assessed model accuracy and pivotal factors impacting HRY across Short, Medium, and Long-grain rice varieties, employing interpretable machine learning on an extensive twelve-season climate-based dataset. Model accuracy remained uniform across varieties, yet influential factors varied, with harvest moisture emerging as crucial for all types but notably more for short and medium grains. Short grains showed increased sensitivity to late-season weather, whereas medium and long grains were more affected by conditions during earlier growth phases. This underscores the value of interpretable machine learning in understanding the differential impact of agronomic conditions on rice varieties, guiding the development of variety-specific strategies to enhance HRY.
This research presents critical advancements in grain quality forecasting and agronomic analysis through adaptable model training methods, leveraging climate and satellite data for enhanced agricultural decision-making. By developing specialised HRY models, the study offers a strategic approach for improving post-harvest processes in the Australian rice industry. Interpretable machine learning techniques facilitate improved communication of model insights, informing agricultural strategies based on empirical data. Furthermore, encouraging improved data collection and sharing highlights the potential of agricultural datasets and interpretable machine learning in advancing agricultural methodologies and addressing food security.