Search

Spectroscopy is an efficient way to estimate the chlorophyll content of crop canopy in the visible and near-infrared region to indicate the photosynthesis capacity and growth status. Although hyperspectral data of crop canopy provide a lot of information, a critical issue is that the subsequent redundant and interference wavelengths could reduce the accuracy and robustness of the estimated results in the field. Therefore, a cascade method of interval-wavelength screening was proposed to reduce the data dimension and improve the crop estimation accuracy and robustness for the chlorophyll content diagnosis of maize crops. In the experiments, spectral reflectance data of maize canopy within 325–1075 nm were obtained, and chlorophyll contents were measured under the conditions of six nitrogen application rates and three growth periods following the early flaring, flaring, and tasseling stages. First, the backward interval partial least squares (BiPLS) method was used to optimize the spectrum interval. Then, competitive adaptive reweighted sampling (CARS) and genetic algorithm (GA) were used for secondary wavelength screening, and the partial least squares regression (Bi-CARS-PLSR and Bi-GA-PLSR) model was established with the chlorophyll content. Compared with the full-spectrum modeling (Full-PLSR) results and the GA and CARS screening wavelength modeling (Full-CARS-PLSR and Full-GA-PLSR) results in the full-spectrum range, the number of wavelengths in the Full-PLSR model was 751. The accuracy of the Full-PLSR model with the coefficient of determination was 0.88 for the calibration set (Rc2) and 0.66 for the verification set (Rv2). The root mean square error (RMSE) was 2.05 mg/L. The wavelength numbers of Full-CARS-PLSR and Full-GA-PLSR were 125 and 99, and their Rc2 and Rv2 were 0.88 and 0.67, respectively. The RMSE values of Full-CARS-PLSR and Full-GA-PLSR were 2.04 and 1.89 mg/L, respectively. By using the proposed cascade method, the wavelength numbers of the Bi-CARS-PLSR model and Bi-GA-PLSR model after interval optimization were 60 and 49, respectively. Their accuracy values for the calibration set (Rc2) were both 0.87, while those for the verification set (Rv2) were 0.7 and 0.78, respectively. The RMSE values of both models were 1.97 and 1.86 mg/L, respectively. The results showed that the proposed cascaded interval-wavelength screening method could eliminate redundant and collinearity variables and improve model performances. At the same time, the Bi-GA-PLSR modeling result was used for field prediction to establish a chlorophyll distribution map, which had a high consistency with the real chlorophyll distribution, and showed the potential to detect chlorophyll content of maize in the field.

To achieve a non-destructive detection of chlorophyll content in field crops based on the reflectance characteristics of chlorophyll in the visible and near-infrared spectrum (400 nm–1000 nm), a crop chlorophyll detector based on an interference filter optical sensor was designed. The hardware part of this detector mainly comprises a microcontroller unit, a sensor module, an input/output module, and a power module. The software is written in Python language and includes main functions, acquisition sub-functions, data processing sub-functions, and data storage sub-functions. Calibration and test experiments were carried out to evaluate the performance of the sensor. Results show that the sensor has a good responsivity of light intensity changes, so as to measure the reflected radiation from crops with the absorption of chlorophyll content. Field verification experiments of corn crops were also carried out, and chlorophyll content detecting models were built by using four combinations of characteristic wavelengths, including 3 peak bands, 9 bands selected via the stepwise regression analysis method, 8 bands selected via the Monte Carlo uninformed variable elimination method, and all 18 bands. Among them, the stepwise regression method obtained the best modeling results. The model showed better performance after calibration than before the calibration with RC2 of 0.72, RV2 of 0.61, RMSEc of 2.35 mg/L, and RMSEv of 2.43 mg/L. The crop chlorophyll detector based on the interference filter optical sensor was used for filed estimation of chlorophyll content which showed a potential for the analysis of crop growth differences.

Leaf chlorophyll content (LCC) is one of nutritional parameters which could be estimated by Hyperspectral image (HSI) technology by combining spatial and spectral information. The objective of this study was to propose a novel segmentation algorithm to remove the influences of the veins in maize leaves so as to improve the accuracy of LCC modeling. Firstly, an enhanced green peak feature (EGPF) was built by local extremum points at 451, 552 and 648 nm, and then the image of EGPF was automatically segmented by OTSU method to get region of interest (ROI) of all leaf regions (ROI-ALR) without background and main stem. Secondly, an enhanced red valley feature (ERVF) was proposed to enlarge the edge difference between veins and mesophylls. ROI of only mesophyll regions (ROI-OMR) were extracted by removing of leaf veins resorted to edge segmentation. For the two spectral reflectance datasets from ROI-ALR and ROI-OMR, multiplicative scattering correction (MSC) was implemented. Correlation analysis (CA) and Random-Frog (RF) coupled with partial least square regression (PLSR) were applied to select characteristic wavelengths and establish LCC estimating models. Compared ROI-ALR- CA-PLSR with ROI-OMR-CA-PLSR, Rv2 increased from 0.43 to 0.52 and RMSE decreased from 4.11 to 3.61; Compared ROI-ALR- RF -PLSR with ROI-OMR- RF –PLSR, Rv2 increased from 0.83 to 0.86 and RMSE decreased from 2.14 to 1.86. The proposed HSI segmentation method for veins removal provides an effective strategy to improve LCC diagnosis of maize leaves.