Modelling And Instrumentation Of Soil And Plant Selected Properties Based On Visible-near-infrared Spectroscopy

Wavelength optimization for calibration models using visible and near infrared (Vis-NIR) spectroscopy is essential for time- and cost-effective analysis of soil properties. This study aims at comparing several wavelength reduction algorithms and rates in terms of model prediction accuracy for total nitrogen (TN), total carbon (TC), organic carbon (OC) and inorganic carbon (IC) in soil. We exploited the uninformative variables elimination (UVE), successive projections algorism (SPA) and two uniform-interval wavelength reduction approaches with successive wavelength reduction rates (WRRs) of 2,5,10,20,50,100,200,500 and 1000. Spectra of soil samples were recorded by a LabSpec2500 spectrophotometer (ASD, USA) in diffuse reflectance mode from 400 to 2499 nm at 1 nm interval and transformed into absorbance spectra using Log(1/R). The absorbance spectra were divided into calibration (n=90) and prediction (n=30) sets. The calibration sets were subjected to a partial least squares regression (PLSR) with leave-one-out cross validation. Validation results prove that wavelength variables can be significantly reduced without pronounced deterioration of model prediction accuracy. The PLSR model for full wavelengths spectra produced excellent prediction accuracy with coefficient of determination (R2) of 0.92,0.90 and 0.89, residual prediction deviation (RPD) of 3.75, 3.22 and 3.11 for TN, TC and OC, respectively, while model prediction for IC was unsatisfactory with R2 of 0.51 and RPD of 1.58. The PLSR models calibrated with visible range (400-759nm) or Vis-SWNIR(400-999nm) still achieved similar performance with R2 of 0.88-0.91 and RPD of 3.16-3.53 for TN, R2 of 0.84-0.87 and RPD of 2.60-2.86 for TC, R2 of 0.86-0.88 and RPD of 2.71-2.96 for OC, R2 of 0.41-0.43 and RPD of 1.45-1.52 for IC, respectively. UVE algorithms reduced the original 2100 wavelengths, respectively, to 71,129,66 and 15 for TN, TC, OC and IC, with which the calibrated PLSR models performed as well as those for full wavelengths spectra but with less number of latent varivales for model calibration. The UVE-SPA produced much more parsimonious models calibrated only with 3 or 4 wavelengths. These models achieved competitive prediction performance, compared to those for full 2100 wavelengths, with R2 of 0.92,0.91,0.90 and 0.44 and RPD of 3.75,3.30,3.14 and 1.40 for soil TN, TC, OC and IC, respectively. Subjected to WRRs from 2 to 100, the calibrated models responded insensitively to the reduced number of wavelengths. For example, the models calibrated for the spectra with 100 nm spacing interval (21 wavelengths left) performed as almost well as those for raw spectra with 1 nm spacing interval. Although these findings might only be valid at farm scale, it is recommended to examine the proposed wavelength reduction algorithms for soils originated from larger geographical areas. The result is also supportive to the instrumentation for soil properties measurement. Non-destructive in situ measurement of tomato fruits is essential to determine the growing stages and to assist automatic picking of fruits. This study evaluates the applicability of visible and near infrared (VIS-NIR) spectroscopy for in situ determination of growing stages and harvest time of tomato fruits of three cultivars. A mobile, fibre-type, AgroSpec VIS-NIR spectrophotometer (Tec5 Co., Germany) with a spectral range of 350-2200 nm was used to measure tomato spectra in reflection mode. Tomato plants were cultivated in the Silsoe Horticultural Center, Bedfordshire, UK during summer growing season in 2010. A new growing stage index (GS) defined as the ratio of current growing age in days to the on-vine duration before harvest in days was proposed. After dividing spectra into a calibration set (70%) and an independent prediction set (30%), spectra in the calibration set were subjected to a partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models to relate GS with spectra of tomato fruits. Separate models were developed for each tomato cultivar and compared with a general model using combined spectra of all three cultivars. Result shows that PLSR based on the new GS is successful and robust in predicting the growing stages and harvest time of tomato fruits. Validation of calibration models on the independent prediction set indicates that successful prediction of GS can be achieved using the three models developed separately for each cultivar with coefficient of determination (R2) of 0.91-0.92, root-mean-square error of prediction (RMSEP) of 0.081-0.097 and residual prediction deviation (RPD) of 3.29-3.70. General calibration using the combined spectra produces good prediction performance, although less accurate than three individual cultivar models. The analysis of regression coefficient plot resulting from PLSR analysis indicates consistent assignment of important wavelengths for individual cultivar spectra and combined spectra. It is concluded that the VIS-NIR-PLSR based on GS index can be adopted successfully for in situ determination of growing stages and harvest time of on-vine tomato fruits, which allows for automatic picking of fruits by a horticultural robot.Handheld SPAD meter is often used to measure chlorophyll content of plant and nitrogen level for some species. For plant production automation, however, it loses its popularity due to its point-by-point checking. In the study, we examined optic-fiber reflection spectroscopy used to measure chlorophyll content of some plant leaves for their SPAD prediction. Totally 120 leaves randomly picked from Hua-jia-chi Campus, Zhejiang University. Of them,70 samples were chosen as calibration set and the remaining as independent prediction set. Each sample was water-cleaned and air-dried. To locate each measurement point precisely when using SPAD meter and spectrometer, a circle with a diameter of 10 mm was drawn on each leaf. By comparing the spectral curves of various leave, it was found that the spectral band between 650-750nm is significant for SPAD modeling since this spectral band for leave with same SPAD reading was close to each other. Besides, we discovered that LED's narrow spectral range used by SPAD meter should be concerned because optic-fiber spectrometer has much more wide spectral range. Based on this feature, we designed an adjustment factor of light to linearly rebuild spectrometer's reflective intensity so that it reached zero outside the band of 650-750nm. Moreover, leaf thickness may make an impact on SPAD prediction since the light of SPAD meter goes through the leaf while the reflection spectrometer does not. First, an equation for SPAD prediction was built with uncertain parameters. Then, a standard genetic algorithm (GA) was designed with Visual Basic 6.0 for parameter optimization. As a result, the optimal reflection band was narrowed within 683.24-733.91nm. The result shows that leave thickness strongly affects the precision of SPAD prediction. Without the information of leaf thickness, the coefficients of determation (R2) between measured and predicted SPAD values were 0.23 and 0.57 for calibration set and prediction set respectively. However, by integrating the values of leaf thickness, better correlation was achieved with R2 of 0.87 and 0.92 for calibration set and prediction set, respectively.In order to design compact Vis-NIR spectrometers for specific application, the author proposed a new electronic system integrating OEM-module for spectral sampling and microchip-based data processing system. The main board, functional programs and data processing software were implemented. The new instrument was used to predict total nitrogen in soil and leaf SPAD value. The result shows that the error between reference and predicted values was no more than 10%, which allows for designing general multiparamter instrument for fast measurement of plant and soil properties.