Predictive modeling of canola seed composition in western Canada: Integrating geospatial patterns with early-season insights

Abstract

Context

Western Canada is the largest global producer and exporter of canola (Brassica napus L.), a versatile crop used for its nutritious oil and high-protein meal. The industry faces challenges due to year-to-year variability in seed oil and protein concentrations, affecting processing logistics and financials.

Objective

This study aimed to identify key temporal and geographical patterns in canola seed composition in Western Canada and to develop a statistical model using weather variables associated to different phenological stages to predict seed composition.

Methods

Spring canola samples collected by the Canadian Grain Commission from 2009 to 2022 were analyzed for seed oil and protein concentrations. Weather data from NASA POWER assessed environmental conditions during different crop developmental stages. Temporal trends were modeled using mixed model analysis, and geospatial analysis characterized spatial patterns. Predictive modeling employed Random Forest algorithms using location, soil, and weather variables.

Results

From 2009–2022, seed oil concentration decreased by 0.17 % per year, while seed protein and meal protein concentrations increased by 0.19 % per year. Spatial analysis showed significant regional variations. Predictive modeling using Random Forest algorithms demonstrated improved accuracy at higher granularity levels, particularly with early reproductive stage weather data. Not improvement was observed by including weather information from later stages. The model's mean absolute error (MAE) for seed oil prediction at the Western Canada level was 0.41 %, compared to the benchmark of a null naïve model's MAE of 0.92 %. Similar improvements were observed for seed protein and meal protein predictions. Temperature, humidity, and rainfall were key factors influencing seed composition.

Conclusions

The study revealed significant spatial and temporal variation in canola seed composition across Western Canada, with early reproductive weather conditions playing a crucial role in determining seed oil and protein concentrations. The integration of geospatial patterns with early-season weather data enabled the development of a predictive model with a mean absolute error of 0.4 %, providing a valuable tool for forecasting seed composition.

Implications

The predictive model can aid industry stakeholders in logistical planning by offering early predictions of seed composition. This tool will allow processors and exporters to anticipate variations in seed oil and protein concentrations well before harvest, enabling informed decisions about sourcing, storage, and processing strategies. By forecasting seed composition, stakeholders can optimize operations to meet market demands and quality standards, potentially reducing costs and improving efficiency.