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

Abstract

17Statistical modeling is commonly used to relate the performance of potato 18 (Solanum tuberosum L.) to fertilizer requirements. Prescribing optimal nutrient doses is 19 challenging because of the involvement of many variables including weather, soils, land 20 management, genotypes, and severity of pests and diseases. Where sufficient data are 21 available, machine learning algorithms can be used to predict crop performance. The 22 objective of this study was to predict tuber yield and quality (size and specific gravity) as 23 impacted by nitrogen, phosphorus and potassium fertilization as well as weather, soils 24 and land management variables. We exploited a data set of 273 field experiments 25 conducted from 1979 to 2017 in Quebec (Canada). We developed, evaluated and 26 compared predictions from a hierarchical Mitscherlich model, k-nearest neighbors, 27 random forest, neuronal networks and Gaussian processes. Machine learning models 28 returned R 2 values of 0.49-0.59 for tuber marketable yield prediction, which were higher 29 than the Mitscherlich model R 2 (0.37). The models were more likely to predict medium-30 size tubers (R 2 = 0.60-0.69) and tuber specific gravity (R 2 = 0.58-0.67 ) than large-size 31 tubers (R 2 = 0.55-0.64) and marketable yield. Response surfaces from the Mitscherlich 32 model, neural networks and Gaussian processes returned smooth responses that agreed 33 more with actual evidence than discontinuous curves derived from k-nearest neighbors 34 and random forest models. When marginalized to obtain optimal dosages from dose-35 response surfaces given constant weather, soil and land management conditions, some 36 disagreements occurred between models. Due to their built-in ability to develop 37 recommendations within a probabilistic risk-assessment framework, Gaussian processes 38 stood out as the most promising algorithm to support decisions that minimize economic 39 or agronomic risks. 3 40 Keywords: 41 Precision fertilization, Mitscherlich model, k-nearest neighbors, random forest, neuronal 42 networks, Gaussian process, economic optimal dose, agronomic optimal dose, Solanum 43 tuberosum L. 4 44 2 Introduction 45 Modeling provides a quantitative understanding of how crop systems operate 46 [1]. Site-specific simulations of fertilizer requirements to obtain high local potato yield 47 and quality rely on models' ability to detect subtle variations in factors affecting plant 48 growth and environment and to learn from the past to make predictions [2]. Several crop 49 models have been developed with different degrees of sophistication, scale, and 50 representativeness [2]. Mechanistic models have been published for potato cropping 51 systems [3, 4]. Semi-mechanistic growth models could be used to downscale tuber yield 52 assessment from regional to field levels [5, 6]. Multilevel modeling can assist in 53 selecting a set of relevant parameters that impact tuber yield and fertilizer requirements, 54 but can hardly predict site-specific nutrient requirements [7]. 55 Several variables can im...