Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements

Shaohui Zhang; Poul Erik Lærke; Mathias Neumann Andersen; Esben Øster Mortensen; Kiril Manevski

doi:10.1016/j.compag.2025.110159

Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements

Shaohui Zhang^*, Poul Erik Lærke, Mathias Neumann Andersen, Esben Øster Mortensen, Kiril Manevski

^*Corresponding author af dette arbejde

Institut for Agroøkologi - Klima og Vand

Publikation: Bidrag til tidsskrift/Konferencebidrag i tidsskrift /Bidrag til avis › Tidsskriftartikel › Forskning › peer review

1 Citationer (Scopus)

Abstract

This paper presents a modelling ensemble to estimate nitrogen requirements (NR) for perennial ryegrass and mixed grasslands from remotely sensed canopy reflectance. We designed a pipeline consisting of two layers, one estimating plant N concentration (PNC) using machine learning regression (MLR), and another estimating biomass production using an improved Carnegie-Ames-Stanford Approach (CASA) involving environmental constraints. The MLR in the first later was selected from screening of random forest, multitask gauss process, partial least square and support vector regression to predict the field measured PNC. The estimated PNC and biomass were located in the critical nitrogen dilution curve developed from the field data to precisely diagnose NR between two consecutive cuts. For comparison, the pipeline was conducted with canopy reflectance data obtained at field and airborne scale. Plant N uptake (PNU), nitrogen nutritional index (NNI) and NR estimated from the field data were also estimated directly by the four MLRs and compared with the pipeline estimates to reveal the added value of the modelling ensemble pipeline involving biomass estimation. Most MLRs estimated better PNC with vegetation indices compared to band reflectance given as input. All MLRs estimated better the N variables of ryegrass than of the mixed grasslands at both field and airborne scale, with partial least square regression having the lowest RMSE for NR of 12–15 kg N ha⁻¹. In nearly all cases, partial least square regression performed stable with lower RMSE compared to the other MLRs. When this regressor was integrated in the dual-layer pipeline for the two grassland systems, the RMSE decreased compared to using only MLRs for PNU, NNI and NR from, respectively, 20–30 to 19 kg N ha⁻¹, 0.15–0.20 to 0.13 and 13–21 to 8 kg N ha⁻¹ at field scale. At airborne scale the RMSE of NNI and NR decreased from, respectively, 0.15–0.28 to 0.15 and 12–18 to 11 kg N ha⁻¹ with no difference for the PNU. The PNC was estimated better by MLRs at airborne than at field cale, while the dual-layer pipeline performed better at field scale. Overall, the study shows that canopy reflectance data and vegetation indices have relatively consistent correlation to N status across different stages for different MLRs, either as stand-alone or integrated in productivity models such as CASA to accurately estimate NR of complex agronomic systems such as perennial single and mixed grasslands. Considering the CASA modelling of the grasses can be considerably improved in relation to belowground biomass considerations, the study thus reveals a significant potential of the modelling ensemble to estimate NR of different grasslands with high accuracy and thus significantly optimize N fertilizer use. The study also provides new knowledge of remote quantitative diagnosis of grass N status with high daily resolution.

Originalsprog	Engelsk
Artikelnummer	110159
Tidsskrift	Computers and Electronics in Agriculture
Vol/bind	233
ISSN	0168-1699
DOI	https://doi.org/10.1016/j.compag.2025.110159
Status	Udgivet - jun. 2025

Adgang til dokumentet

10.1016/j.compag.2025.110159

Andre filer og links

Link to publication in Scopus

Citationsformater

@article{35f6352ff28846af8d5cdaa431877302,

title = "Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements",

abstract = "This paper presents a modelling ensemble to estimate nitrogen requirements (NR) for perennial ryegrass and mixed grasslands from remotely sensed canopy reflectance. We designed a pipeline consisting of two layers, one estimating plant N concentration (PNC) using machine learning regression (MLR), and another estimating biomass production using an improved Carnegie-Ames-Stanford Approach (CASA) involving environmental constraints. The MLR in the first later was selected from screening of random forest, multitask gauss process, partial least square and support vector regression to predict the field measured PNC. The estimated PNC and biomass were located in the critical nitrogen dilution curve developed from the field data to precisely diagnose NR between two consecutive cuts. For comparison, the pipeline was conducted with canopy reflectance data obtained at field and airborne scale. Plant N uptake (PNU), nitrogen nutritional index (NNI) and NR estimated from the field data were also estimated directly by the four MLRs and compared with the pipeline estimates to reveal the added value of the modelling ensemble pipeline involving biomass estimation. Most MLRs estimated better PNC with vegetation indices compared to band reflectance given as input. All MLRs estimated better the N variables of ryegrass than of the mixed grasslands at both field and airborne scale, with partial least square regression having the lowest RMSE for NR of 12–15 kg N ha−1. In nearly all cases, partial least square regression performed stable with lower RMSE compared to the other MLRs. When this regressor was integrated in the dual-layer pipeline for the two grassland systems, the RMSE decreased compared to using only MLRs for PNU, NNI and NR from, respectively, 20–30 to 19 kg N ha−1, 0.15–0.20 to 0.13 and 13–21 to 8 kg N ha−1 at field scale. At airborne scale the RMSE of NNI and NR decreased from, respectively, 0.15–0.28 to 0.15 and 12–18 to 11 kg N ha−1 with no difference for the PNU. The PNC was estimated better by MLRs at airborne than at field cale, while the dual-layer pipeline performed better at field scale. Overall, the study shows that canopy reflectance data and vegetation indices have relatively consistent correlation to N status across different stages for different MLRs, either as stand-alone or integrated in productivity models such as CASA to accurately estimate NR of complex agronomic systems such as perennial single and mixed grasslands. Considering the CASA modelling of the grasses can be considerably improved in relation to belowground biomass considerations, the study thus reveals a significant potential of the modelling ensemble to estimate NR of different grasslands with high accuracy and thus significantly optimize N fertilizer use. The study also provides new knowledge of remote quantitative diagnosis of grass N status with high daily resolution.",

keywords = "Carnegie-Ames-Stanford Approach, Double-layer model, Mixed grassland, Nitrogen sensing, Vegetation index",

author = "Shaohui Zhang and L{\ae}rke, {Poul Erik} and Andersen, {Mathias Neumann} and Mortensen, {Esben {\O}ster} and Kiril Manevski",

note = "Publisher Copyright: {\textcopyright} 2025 Elsevier B.V.",

year = "2025",

month = jun,

doi = "10.1016/j.compag.2025.110159",

language = "English",

volume = "233",

journal = "Computers and Electronics in Agriculture",

issn = "0168-1699",

publisher = "Elsevier B.V.",

}

Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements. / Zhang, Shaohui ; Lærke, Poul Erik ; Andersen, Mathias Neumann et al.
I: Computers and Electronics in Agriculture, Bind 233, 110159, 06.2025.

Publikation: Bidrag til tidsskrift/Konferencebidrag i tidsskrift /Bidrag til avis › Tidsskriftartikel › Forskning › peer review

TY - JOUR

T1 - Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements

AU - Zhang, Shaohui

AU - Lærke, Poul Erik

AU - Andersen, Mathias Neumann

AU - Mortensen, Esben Øster

AU - Manevski, Kiril

PY - 2025/6

Y1 - 2025/6

N2 - This paper presents a modelling ensemble to estimate nitrogen requirements (NR) for perennial ryegrass and mixed grasslands from remotely sensed canopy reflectance. We designed a pipeline consisting of two layers, one estimating plant N concentration (PNC) using machine learning regression (MLR), and another estimating biomass production using an improved Carnegie-Ames-Stanford Approach (CASA) involving environmental constraints. The MLR in the first later was selected from screening of random forest, multitask gauss process, partial least square and support vector regression to predict the field measured PNC. The estimated PNC and biomass were located in the critical nitrogen dilution curve developed from the field data to precisely diagnose NR between two consecutive cuts. For comparison, the pipeline was conducted with canopy reflectance data obtained at field and airborne scale. Plant N uptake (PNU), nitrogen nutritional index (NNI) and NR estimated from the field data were also estimated directly by the four MLRs and compared with the pipeline estimates to reveal the added value of the modelling ensemble pipeline involving biomass estimation. Most MLRs estimated better PNC with vegetation indices compared to band reflectance given as input. All MLRs estimated better the N variables of ryegrass than of the mixed grasslands at both field and airborne scale, with partial least square regression having the lowest RMSE for NR of 12–15 kg N ha−1. In nearly all cases, partial least square regression performed stable with lower RMSE compared to the other MLRs. When this regressor was integrated in the dual-layer pipeline for the two grassland systems, the RMSE decreased compared to using only MLRs for PNU, NNI and NR from, respectively, 20–30 to 19 kg N ha−1, 0.15–0.20 to 0.13 and 13–21 to 8 kg N ha−1 at field scale. At airborne scale the RMSE of NNI and NR decreased from, respectively, 0.15–0.28 to 0.15 and 12–18 to 11 kg N ha−1 with no difference for the PNU. The PNC was estimated better by MLRs at airborne than at field cale, while the dual-layer pipeline performed better at field scale. Overall, the study shows that canopy reflectance data and vegetation indices have relatively consistent correlation to N status across different stages for different MLRs, either as stand-alone or integrated in productivity models such as CASA to accurately estimate NR of complex agronomic systems such as perennial single and mixed grasslands. Considering the CASA modelling of the grasses can be considerably improved in relation to belowground biomass considerations, the study thus reveals a significant potential of the modelling ensemble to estimate NR of different grasslands with high accuracy and thus significantly optimize N fertilizer use. The study also provides new knowledge of remote quantitative diagnosis of grass N status with high daily resolution.

AB - This paper presents a modelling ensemble to estimate nitrogen requirements (NR) for perennial ryegrass and mixed grasslands from remotely sensed canopy reflectance. We designed a pipeline consisting of two layers, one estimating plant N concentration (PNC) using machine learning regression (MLR), and another estimating biomass production using an improved Carnegie-Ames-Stanford Approach (CASA) involving environmental constraints. The MLR in the first later was selected from screening of random forest, multitask gauss process, partial least square and support vector regression to predict the field measured PNC. The estimated PNC and biomass were located in the critical nitrogen dilution curve developed from the field data to precisely diagnose NR between two consecutive cuts. For comparison, the pipeline was conducted with canopy reflectance data obtained at field and airborne scale. Plant N uptake (PNU), nitrogen nutritional index (NNI) and NR estimated from the field data were also estimated directly by the four MLRs and compared with the pipeline estimates to reveal the added value of the modelling ensemble pipeline involving biomass estimation. Most MLRs estimated better PNC with vegetation indices compared to band reflectance given as input. All MLRs estimated better the N variables of ryegrass than of the mixed grasslands at both field and airborne scale, with partial least square regression having the lowest RMSE for NR of 12–15 kg N ha−1. In nearly all cases, partial least square regression performed stable with lower RMSE compared to the other MLRs. When this regressor was integrated in the dual-layer pipeline for the two grassland systems, the RMSE decreased compared to using only MLRs for PNU, NNI and NR from, respectively, 20–30 to 19 kg N ha−1, 0.15–0.20 to 0.13 and 13–21 to 8 kg N ha−1 at field scale. At airborne scale the RMSE of NNI and NR decreased from, respectively, 0.15–0.28 to 0.15 and 12–18 to 11 kg N ha−1 with no difference for the PNU. The PNC was estimated better by MLRs at airborne than at field cale, while the dual-layer pipeline performed better at field scale. Overall, the study shows that canopy reflectance data and vegetation indices have relatively consistent correlation to N status across different stages for different MLRs, either as stand-alone or integrated in productivity models such as CASA to accurately estimate NR of complex agronomic systems such as perennial single and mixed grasslands. Considering the CASA modelling of the grasses can be considerably improved in relation to belowground biomass considerations, the study thus reveals a significant potential of the modelling ensemble to estimate NR of different grasslands with high accuracy and thus significantly optimize N fertilizer use. The study also provides new knowledge of remote quantitative diagnosis of grass N status with high daily resolution.

KW - Carnegie-Ames-Stanford Approach

KW - Double-layer model

KW - Mixed grassland

KW - Nitrogen sensing

KW - Vegetation index

UR - http://www.scopus.com/inward/record.url?scp=85218866147&partnerID=8YFLogxK

U2 - 10.1016/j.compag.2025.110159

DO - 10.1016/j.compag.2025.110159

M3 - Journal article

AN - SCOPUS:85218866147

SN - 0168-1699

VL - 233

JO - Computers and Electronics in Agriculture

JF - Computers and Electronics in Agriculture

M1 - 110159

ER -

Combining biomass production model with machine learning regression of critical nitrogen concentration for estimating grassland nitrogen requirements

Abstract

Adgang til dokumentet

Andre filer og links

Fingeraftryk

Citationsformater