A Cost-Effective and Portable Optical Sensor System to Estimate Leaf Nitrogen and Water Contents in Crops

Abstract

Non-invasive determination of leaf nitrogen (N) and water contents is essential for ensuring the healthy growth of the plants. However, most of the existing methods to measure them are expensive. In this paper, a low-cost, portable multispectral sensor system is proposed to determine N and water contents in the leaves, non-invasively. Four different species of plants—canola, corn, soybean, and wheat—are used as test plants to investigate the utility of the proposed device. The sensor system comprises two multispectral sensors, visible (VIS) and near-infrared (NIR), detecting reflectance at 12 wavelengths (six from each sensor). Two separate experiments were performed in a controlled greenhouse environment, including N and water experiments. Spectral data were collected from 307 leaves (121 for N and 186 for water experiment), and the rational quadratic Gaussian process regression (GPR) algorithm was applied to correlate the reflectance data with actual N and water content. By performing five-fold cross-validation, the N estimation showed a coefficient of determination ( R 2 ) of 63.91% for canola, 80.05% for corn, 82.29% for soybean, and 63.21% for wheat. For water content estimation, canola showed an R 2 of 18.02%, corn showed an R 2 of 68.41%, soybean showed an R 2 of 46.38%, and wheat showed an R 2 of 64.58%. The result reveals that the proposed low-cost sensor with an appropriate regression model can be used to determine N content. However, further investigation is needed to improve the water estimation results using the proposed device.

Development of low‐cost leaf spectroscopy based on featured wavelength extraction by fix‐bandwidth Gaussian mixture model

“…Motivated by the Gaussian mixture model (GMM) for the fitting spectrum of UV-light [13], the noisy and multi-peaks spectrum of input sunlight in the leaf spectroscopy is described mathematically by GMM, and the standard parameters of the GMM are assumed to be equal because of the optical-electrical sensor arrays fabricated with same band width [12]. That is, the spectrum I(𝜆; Θ) is,…”

“…Motivated by the Gaussian mixture model (GMM) for the fitting spectrum of UV‐light [13], the noisy and multi‐peaks spectrum of input sunlight in the leaf spectroscopy is described mathematically by GMM, and the standard parameters of the GMM are assumed to be equal because of the optical‐electrical sensor arrays fabricated with same band width [12]. That is, the spectrum$${\rm{\;I}}( {\lambda ;\Theta } )$$ is, $$$$\begin{equation}I\left( {\lambda ;\Theta } \right) = \mathop \sum \limits_{k = 1}^N {{{\pi}}_k}{G_k}\left( {\lambda ;{\mu _k},\sigma _0^2} \right),\;\lambda \in \left( {360,960} \right)\end{equation}$$$$where λ denotes the wavelength with a unit of nm, and $$\Theta = \{ {\sigma _0^2;{{{\pi}}_k},{\mu _k}} \}_{k = 1}^N $$ denotes the parameter set of N components, and $${G_k} ( {\lambda ;{\mu _k},\sigma _0^2} ) = {\rm{exp}} ( { - {{( {\lambda - {\mu _k}} )}^2}/2\sigma _0^2} )$$ is the k th Gaussian function with a mean of $${\mu _k}$$ and a variance of $$\sigma _0^2$$, and $${{{\pi}}_k}$$ is the amplitude.…”

“…Currently, there is an attempt to develop a low‐cost sensor node for estimating plant variables by detecting optical properties at certain wavelengths using silicon (Si) optical sensors rather than FBG sensors. For example, the plant stress is evaluated at four peak wavelengths of 850, 630, 535 and 465 nm [11], and another example is water content estimation in leaves by the 12‐wavelength multispectral sensors [12]. However, there is no report on the optimal approximation for leaf spectrum measured by the high‐resolution FBG sensors, still fewer applications in predication and validation of vegetation based on modelling.…”

What Is the Effect of Quantitative Inversion of Photosynthetic Pigment Content in Populus euphratica Oliv. Individual Tree Canopy Based on Multispectral UAV Images?

“…GPR, as a new machine learning algorithm, has a good adaptability to deal with small sample, high-dimensional nonlinear and random data, and it has the benefits of hyperparameter adaptive projection, output distribution probability derivation and strong generalization ability. In this study, the GPR model to estimate PPC in PeITC was realized by appropriately modifying the code disclosed by Habibullah et al [74].…”