Estimating Above-Ground Biomass of Potato Using Random Forest and Optimized Hyperspectral Indices

Abstract

Spectral indices rarely show consistency in estimating crop traits across growth stages; thus, it is critical to simultaneously evaluate a group of spectral variables and select the most informative spectral indices for retrieving crop traits. The objective of this study was to explore the optimal spectral predictors for above-ground biomass (AGB) by applying Random Forest (RF) on three types of spectral predictors: the full spectrum, published spectral indices (Pub-SIs), and optimized spectral indices (Opt-SIs). Canopy hyperspectral reflectance of potato plants, treated with seven nitrogen (N) rates, was obtained during the tuber formation and tuber bulking from 2015 to 2016. Twelve Pub-SIs were selected, and their spectral bands were optimized using band optimization algorithms. Results showed that the Opt-SIs were the best input variables of RF models. Compared to the best empirical model based on Opt-SIs, the Opt-SIs based RF model improved the prediction of AGB, with R2 increased by 6%, 10%, and 16% at the tuber formation, tuber bulking, and for across the two growth stages, respectively. The Opt-SIs can significantly reduce the number of input variables. The optimized Blue nitrogen index (Opt-BNI) and Modified red-edge normalized difference vegetation index (Opt-mND705) combined with an RF model showed the best performance in estimating potato AGB at the tuber formation stage (R2 = 0.88). In the tuber bulking stage, only using optimized Nitrogen planar domain index (Opt-NPDI) as the input variable of the RF model produced satisfactory accuracy in training and testing datasets, with the R2, RMSE, and RE being 0.92, 208.6 kg/ha, and 10.3%, respectively. The Opt-BNI and Double-peak nitrogen index (Opt-NDDA) coupling with an RF model explained 86% of the variations in potato AGB, with the lowest RMSE (262.9 kg/ha) and RE (14.8%) across two growth stages. This study shows that combining the Opt-SIs and RF can greatly enhance the prediction accuracy for crop AGB while significantly reduces collinearity and redundancies of spectral data.

Estimating the Leaf Water Status and Grain Yield of Wheat under Different Irrigation Regimes Using Optimized Two- and Three-Band Hyperspectral Indices and Multivariate Regression Models

“…This may explain why the coupling of NWIs-3b with the three models showed a better performance in the estimation of the three plant traits in both calibration and validation datasets as compared with the other types of SRIs. Similarly, Yang et al [41] reported that the type of SRIs significantly influenced the performance of RF models for estimating the AGB of potato crops across different growth stages.…”

“…This is because the combined few wavelengths in specific formulas increase the sensitivity of SRIs to different vegetation physical and biochemical proprieties rather than the target traits [70,77,83,84]. Therefore, recent studies have demonstrated that the SRIs coupled with the development of data-driven models, such as RF or multivariate regression models such as PLSR and MIR, can improve the accurate estimation of plant traits as compared to the usage of individual SRIs [41,48,50,52,84,85]. However, the numbers and types of input variables in these models can significantly influence their performance in estimating the plant traits.…”

“…However, the numbers and types of input variables in these models can significantly influence their performance in estimating the plant traits. For example, Wang et al [51], Niu et al [52], and Yang et al [41] reported that the optimized SRIs were more suitable as input variables in the RF, PLSR, and MLR models than the full-spectrum bands and some published SRIs based on two-band wavelengths, such as the NDVI, modified soil-adjusted vegetation index, and the simple ratio vegetation index, in the estimation of the aboveground biomass (AGB) of potato and maize crops at a single growth stage or across different growth stages. In this study, the four groups of SRIs individually or combined were used as input variables in the RF, PLSR, and MLR models for the estimation of the three plant traits.…”

“…Interestingly, the photochemical reflectance index (PRI) and normalized difference vegetation index (NDVI) showed also a significant relationship with different plant water status indicators, although the former index includes two wavelengths from the VIS region (531 and 570 nm) and the latter index includes one wavelength from the VIS region (680 nm) and one from the NIR region (900 nm) [37][38][39]. However, the relationships between the different SRIs and plant traits are often inconsistent due to the impact of several factors, such as crop types, years, environmental conditions, and growth stages, on the algorithms of the SRIs and corresponding sensitive band combinations [11,40,41]. Therefore, extracting the sensitive wavelengths from the full spectrum range and formulating them in suitable SRI algorithms is a necessary step to adopt the SRIs as an efficient and simple approach to estimate plant traits.…”

Combining Hyperspectral Reflectance Indices and Multivariate Analysis to Estimate Different Units of Chlorophyll Content of Spring Wheat under Salinity Conditions

“…In general, the performance of SRIs in the estimation of various plant traits is often inconsistent when measured across genotypes, years, sites, and phenological stages, as almost all of them include only two to three wavelengths [ 40 , 41 , 42 , 43 ]. Additionally, because the Chl content and LAI have similar influences on the spectral reflectance of the canopy, particularly at the wavelengths of 550 and 750 nm, these limited wavelengths influence the performance of SRIs when they are used individually to uncouple the combined effect of both traits [ 33 ].…”