Hyperspectral reflectance and machine learning to monitor legume biomass and nitrogen accumulation

Authors: Flynn, Kyle; BAATH, GURJINDER - Texas Agrilife Research; Lee, Trey; Gowda, Prasanna; Northup, Brian

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/11/2023
Publication Date: 6/18/2023
Citation: Flynn, K.C., Baath, G., Lee, T.O., Gowda, P.H., Northup, B.K. 2023. Hyperspectral reflectance and machine learning to monitor legume biomass and nitrogen accumulation. Computers and Electronics in Agriculture. 211. Article 107991. https://doi.org/10.1016/j.compag.2023.107991.
DOI: https://doi.org/10.1016/j.compag.2023.107991

Interpretive Summary: Hyperspectral remote sensing provides opportunity for a nondestructive tool for estimating biochemical or biophysical characteristics of agricultural crops. Importantly, among rotational legume production, timely identification of the extent of nitrogen accumulation (N accum.) by the legumes is vital for optimization of N fertilizer application and simultaneous reduction of environmental impacts (i.e. excess runoff). We developed hyperspectral data-based models to predict nitrogen concentration (N conc.), biomass, and N accum. for three legumes (soybean, tepary bean, mothbean) through application and testing of various methods of machine learning. An exploration of potential applications of a future hyperspectral satellite (i.e. CHIME) was also conducted. Results suggest hyperspectral remote sensing has great promise across both the in-situ and convolved satellite (i.e. CHIME) bands. These findings enable the modeling of biophysical and biochemical characteristics of legumes using combinations of hyperspectral remote sensing and machine learning methodologies.

Technical Abstract: Hyperspectral remote sensing provides opportunity for a nondestructive tool for estimating biochemical or biophysical characteristics of agricultural crops. Importantly, among rotational legume production, timely identification of the extent of nitrogen accumulation (N accum.) by the legumes is vital for optimization of N fertilizer application and simultaneous reduction of environmental impacts (i.e. excess runoff). We developed in-situ hyperspectral data-based models to predict nitrogen concentration (N conc.), biomass, and N accum. for three legumes (soybean, tepary bean, mothbean) through application and testing of four methods of machine learning (k-Nearest Neighbors [KNN], partial least squares regression [PLS], support vector machine [SVM], and random forest [RF]). An exploration of potential applications of a future hyperspectral satellite (i.e. CHIME) was also conducted by convolving hyperspectral wavelengths to fit hyperspectral bands of the satellite to conduct a similar analysis to that of the in-situ hyperspectral data. Other analytics such as legume type incorporation and direct versus derived N accum. from hyperspectral data were explored. Results suggest hyperspectral remote sensing has great promise across both the in-situ and convolved satellite (i.e. CHIME) bands. Moreover, models based on SVM and RF machine learning algorithms had the greatest outcomes across the tested algorithms, with SVM being less computationally expensive than RF. Moreover, legume type was not significantly important to model development, and N accum. is best modeled directly rather than being derived independently from the modeling of N concentrations and biomass. These findings enable the modeling of biophysical and biochemical characteristics of legumes using combinations of hyperspectral remote sensing and machine learning methodologies.