CARS-ABC-SVR model for predicting leaf moisture of leaf-used lettuce based on hyperspectral

Abstract: In order to realize more reasonably irrigation management during the growth of leaf -used lettuce, a new method for accurately, rapidly and effectively detecting leaf-used lettuce moisture based on hyperspectral technology was investigated in this study. Leaf-used lettuces of 5 different water stress levels were adopted as experimental objects. In the first group, sufficient water irrigation was maintained during the growth period of leaf-used lettuces, and the amount of water irrigated in the second, third, fourth and fifth groups decreased in turn according to the gradient. Firstly, hyperspectral images of leaf-used lettuce samples were acquired by using the hyperspectral image acquisition system, then the water contents of all leaves were measured by the drying method and the dry-basis moisture content was calculated according to formula. Secondly, the hyperspectral data was extracted from the images by selecting the region of interest (ROI) in the ENVI software. Thirdly, a method for data pretreatment, Savitzky-Golay (SG) combined with the standard normalized variable (SNV), was applied for smoothing and denoising of the original hyperspectral data. Fourthly, the competitive adaptive reweighted sampling (CARS) algorithm was used to extract the characteristic wavelengths ranged from 965 nm to 1666 nm of leaf-used lettuce samples, simultaneously the effect of CARS algorithm was compared with that of the stepwise regression (SR) analysis and the successive projections algorithm (SPA) in order to determine the optimal method for characteristic wavelength selection. Finally, the support vector regression (SVR) machine was respectively carried out to establish the relationship models between full spectral data, three kinds of characteristic spectral data and dry-basis moisture content of leaf-used lettuce samples. And the performances of all the models were evaluated by the index of determination coefficient for calibration set (Rc2), root mean square error for calibration set (RMSEC), determination coefficient for prediction set (RP2) and root mean square error for prediction set (RMSEP). The results showed that CARS-SVR model performed better than the other model with full-SVR, SR-SVR or SPA-SVR, selecting the optimal wavelength combination (973, 993, 997, 1 050, 1 140, 1 181, 1 184, 1 188, 1 191, 1 198, 1 237, 1 240, 1 243, 1 259, 1 263, 1 285, 1 310, 1 336, 1 348, 1 354, 1 376, 1 389, 1 392, 1 395, 1 408, 1 414, 1 601, 1 662 nm), and achieving the highest accuracy with Rc2 = 0.917 2, RMSEC = 2.33%, RP2 = 0.859 9 and RMSEP = 3.95%. Whereas, the prediction accuracy of CARS-SVR model were not achieved the desired effect. For improving the prediction accuracy of SVR model, the artificial bee colony (ABC) algorithm was further introduced to intelligently optimize the parameters (c and g) in the SVR model to find the optimum, then the model on the basis of CARS characteristic data was reconstructed. Consequently, the optimised model, CARS-ABC-SVR, achieved the Rc2 of 0.942 7, RMSEC of 1.60%, RP2 of 0.921 4 and RMSEP of 2.95%, which was indeed improved significantly and proved that the method of selecting characteristic wavelengths by CARS algorithm combined with optimizing the parameters in SVR model by ABC algorithm can extremely raise the performance of prediction model for the moisture content of leaves. Hence, the method of hyperspectral technology combined with the CARS-ABC-SVR model is feasible for detecting the moisture content of leaf-used lettuces, also hopefully providing a new method and thought for water detection of other crops.