Abstract: In order to obtain water information from plant leaves rapidly and accurately, a new method for determining moisture content in soybean leaves was developed on the basis of the terahertz spectral imaging system. The terahertz images of selected 96 soybean leaves with different moisture contents were captured by the terahertz time-domain spectroscopy system (THz-TDS), and the moisture contents of leaves were measured by the electronic scale. Principal component analysis (PCA) was conducted on the 0.2-1.6 THz terahertz images, and three effective bands 0.557, 1.098, 1.163 THz were determined. Adaptive threshold segmentation was adopted to divide the leaf image into the vein image and the mesophyll image. Then gray features of the gained images of leaf, vein and mesophyll were computed,which were classified into three groups: the group of leaf (G1), the group of vein (G2) and the group of mesophyll (G3). At the same time, algorithms of multiple linear regression (MLR), back propagation (BP) neural network (BP-ANN) and least squares support vector machine (LS-SVM) were used to establish 9 prediction models of moisture content in soybean leaves, and the 3 groups were used as input. It turned out that the LS-SVM model based on G3 had the best prediction results among all models, as the determination coefficient of the calibration set and the prediction set reached 0.967 8 and 0.963 2, respectively, and the root mean square errors were 0.057 8 and 0.046 5, respectively. The experiment results showed that the proposed method was accurate and offered an effective means to measure the moisture content of crop leaves.

Key words: terahertz, leaves, moisture content, principal component analysis, image processing, data modeling