Zhou Qiyu, Soldat Douglas J
Department of Soil Science, University of Wisconsin-Madison, Madison, WI, United States.
Front Plant Sci. 2022 May 19;13:863211. doi: 10.3389/fpls.2022.863211. eCollection 2022.
Nitrogen (N) is the most limiting nutrient for turfgrass growth. Few tools or soil tests exist to help managers guide N fertilizer decisions. Turf growth prediction models have the potential to be useful, but the lone turfgrass growth prediction model only takes into account temperature, limiting its accuracy. This study investigated the ability of a machine learning (ML)-based turf growth model using the random forest (RF) algorithm (ML-RF model) to improve creeping bentgrass () putting green management by estimating short-term clipping yield. This method was compared against three alternative N application strategies including (1) PACE Turf growth potential (GP) model, (2) an experience-based method for applying N fertilizer (experience-based method), and (3) the experience-based method guided by a vegetative index, normalized difference red edge (NDRE)-based method. The ML-RF model was built based on a set of variables including 7-day weather, evapotranspiration (ET), traffic intensity, soil moisture content, N fertilization rate, NDRE, and root zone type. The field experiment was conducted on two sand-based research greens in 2020 and 2021. The cumulative applied N fertilizer was 281 kg ha for the PACE Turf GP model, 190 kg ha for the experience-based method, 140 kg ha for the ML-RF model, and around 75 kg ha NDRE-based method. ML-RF model and NDRE-based method were able to provide customized N fertilization recommendations on different root zones. The methods resulted in different mean turfgrass qualities and NDRE. From highest to lowest, they were PACE Turf GP model, experience-based, ML-RF model, and NDRE-based method, and the first three methods produced turfgrass quality over 7 (on a scale from 1 to 9) and NDRE value over 0.30. N fertilization guided by the ML-RF model resulted in a moderate amount of fertilizer applied and acceptable turfgrass performance characteristics. This application strategy is based on the N cycle and has the potential to assist turfgrass managers in making N fertilization decisions for creeping bentgrass putting greens.
氮（N）是草坪草生长最具限制性的养分。几乎没有工具或土壤测试可帮助管理者指导氮肥决策。草坪生长预测模型有潜在的用途，但唯一的草坪草生长预测模型只考虑了温度，限制了其准确性。本研究调查了一种基于机器学习（ML）的草坪生长模型（使用随机森林（RF）算法，即ML-RF模型）通过估计短期剪草产量来改善匍匐翦股颖果岭管理的能力。该方法与三种替代氮肥施用策略进行了比较，包括（1）PACE草坪生长潜力（GP）模型，（2）基于经验的氮肥施用方法（基于经验的方法），以及（3）由植被指数归一化差值红边（NDRE）引导的基于经验的方法。ML-RF模型是基于一组变量构建的，包括7天天气、蒸散量（ET）、交通强度、土壤含水量、氮肥施用量、NDRE和根区类型。田间试验于2020年和2021年在两个沙基研究果岭上进行。PACE草坪GP模型的累计施氮量为281 kg/公顷，基于经验的方法为190 kg/公顷，ML-RF模型为140 kg/公顷，基于NDRE的方法约为75 kg/公顷。ML-RF模型和基于NDRE的方法能够针对不同根区提供定制的氮肥施用建议。这些方法导致了不同的平均草坪质量和NDRE。从高到低依次为PACE草坪GP模型、基于经验的方法、ML-RF模型和基于NDRE的方法，前三种方法产生的草坪质量超过7（1至9分制），NDRE值超过0.30。由ML-RF模型引导的氮肥施用导致施肥量适中且草坪性能特征可接受。这种施用策略基于氮循环，有潜力协助草坪管理者为匍匐翦股颖果岭做出氮肥决策。
