Ensemble Hyperspectral Band Selection for Detecting Nitrogen Status in Grape Leaves

Abstract

The large data size and dimensionality of hyperspectral data demands complex processing and data analysis. Multispectral data do not suffer the same limitations, but are normally restricted to blue, green, red, red edge, and near infrared bands. This study aimed to identify the optimal set of spectral bands for nitrogen detection in grape leaves using ensemble feature selection on hyperspectral data from over 3,000 leaves from 150 'Flame Seedless' table grapevines. Six machine learning base rankers were included in the ensemble: random forest, LASSO, SelectKBest, ReliefF, SVM-RFE, and chaotic crow search algorithm (CCSA). The pipeline identified less than 0.45% of the bands as most informative about grape nitrogen status. The selected violet, yellow-orange, and shortwave infrared bands lie outside of the typical blue, green, red, red edge, and near infrared bands of commercial multispectral cameras, so the potential improvement in remote sensing of nitrogen in grapevines brought forth by a customized multispectral sensor centered at the selected bands is promising and worth further investigation. The proposed pipeline may also be used for application-specific multispectral sensor design in domains other than agriculture.

Assessing Grapevine Nutrient Status from Unmanned Aerial System (UAS) Hyperspectral Imagery

“…The ensemble feature ranking followed by MLR-RFE method and the standalone PLSR method each achieved comparable performances in terms of the LOOCV-RMSE, but both may be further improved by exploring alternatives. For example, in the case of ensemble ranking, alternative rankers might prove superior, or the recursive elimination of rankers might be used to optimize the regression results [26]. Additionally, other parent regression methods, such as SVR or RFR, may yield lower RMSE than MLR.…”

“…The ensemble feature selector used a combination of six feature rankers, similar to Omidi et al [26] and Moghimi et al [46]. Using a multimethod ensemble combines the advantages and disadvantages of a diverse set of methods to rank features.…”

“…The chlorophyll absorption-related wavelengths in the VNIR region were consistent with the most frequently selected bands found by Fu et al [23] in their review of the crop nitrogen status using hyperspectral remote sensing. Our selection of 606 nm fell right between 596 nm and 611 nm, a spectral region noted by Omidi et al [26] in their study of grapevine leaf handheld spectral data using an ensemble ranking method for the classification of nitrogen deficiency.…”

“…The major impediments to finding correlated spectral responses to vegetation traits using hyperspectral data are the redundancy in their high data dimensionality and the propensity of a limited sample (i.e., fewer than the number of potential predictor variables) to cause overfitting, preventing the development of robust generalized models [24]. This is why many studies use well-known or empirically derived vegetation indices consisting of just a small number of bands [25] while others combine a variety of machine learning feature selection methods to build models utilizing only the most target-relevant features [26,27]. Feilhauer et al [28] found that an ensemble of SVR, RFR, and PLSR machine learning methods achieved consistent wavelength selections between different datasets in good agreement with known spectral absorption features compared to the standard partial least squares regression technique for spectral band selection related to leaf biochemistry [29].…”

“…In grapevines, specifically, Friedel et al [25] found that the red-edge inflection point index (REIP) had the lowest RMSE (~0.17% nitrogen by dry mass in samples ranging from 2.18 to 2.58%) when comparing the vegetation indices derived from handheld spectral data. Conversely, Omidi et al [26] determined the optimal handheld spectral bands of the leaf samples centered at 596, 611, 401, 1619, 825, and 1386 nm to be the most important for the classification of high and low nitrogen content grapevines. Moghimi et al [38] modeled the nitrogen content for the same vines using UAS-based five-band multispectral imagery with a RMSE (hold-out test set) of 0.23% nitrogen in samples with contents ranging from 2.25 to 4.25% nitrogen.…”