Site-specific machine learning predictive fertilization models for potato crops in Eastern Canada

Area of Science:

• Agricultural Science
• Data Science
• Agronomy

Background:

• Optimizing potato (Solanum tuberosum L.) fertilizer requirements is complex due to numerous interacting variables.
• Statistical modeling is common, but predicting crop performance with high accuracy requires advanced methods like machine learning when sufficient data is available.

Purpose of the Study:

• To identify the optimal model for predicting nitrogen, phosphorus, and potassium requirements for high potato tuber yield and quality.
• To evaluate models based on their ability to account for weather, soil, and land management impacts.

Main Methods:

• Compared a hierarchical Mitscherlich model with machine learning algorithms: k-nearest neighbors, random forest, neural networks, and Gaussian processes.
• Utilized a dataset of 273 field experiments in Quebec (Canada) from 1979 to 2017.
• Evaluated model performance using R2 values for tuber yield and quality traits (size, specific gravity).

Main Results:

• Machine learning models achieved higher R2 values (0.49-0.59) for marketable yield prediction than the Mitscherlich model (0.37).
• Models predicted medium-size tubers (R2 = 0.60-0.69) and specific gravity (R2 = 0.58-0.67) better than large-size tubers (R2 = 0.55-0.64).
• Gaussian processes, neural networks, and the Mitscherlich model provided smoother, more evidence-aligned response surfaces than k-nearest neighbors and random forest.

Conclusions:

• Machine learning algorithms, particularly Gaussian processes, show significant promise for optimizing potato fertilization strategies.
• Gaussian processes offer a robust framework for probabilistic risk assessment, aiding decisions to minimize economic and agronomic risks.
• Accurate prediction of potato yield and quality requires sophisticated modeling that accounts for diverse environmental and management factors.