Abstract:In response to the need for smart agriculture to accurately discriminate the degree of crop water demand, taking growing peppers as the experimental samples, different degrees of water stress treatments such as water immersion and drought to the leaves of peppers were applied to analyze the hyperspectral response characteristics of pepper leaves under different degrees of water stress. The samples were divided into four water stress groups, including severe drought, mild drought, mild water-soaked, and severe water-soaked, and one experimental control group, with a total of five data groups of 20 chili peppers in each group, and the chlorophyll fluorescence parameters and hyperspectral data of chili peppers’ leaves in each group were collected separately when the appearance of leaves in each group appeared to be obviously different. The effects of three different preprocessing methods, namely, multiplicative scatter correction (MSC), SG convolutional smoothing filter and standard normal variate transform (SNV), on the elimination of background information interference were compared. The SPA algorithm and CARS algorithm were used to extract the characteristic wavelengths sensitive to water stress. Support vector machine (SVM), BP neural network, radial basis function (RBF) and random forest (RF) modeling were established for predicting different levels of water stress. The results illustrated that SG-SPA-RFB was the optimal combination for predicting the degree of water stress with 99.02% accuracy in the training set and 94.00% accuracy in the test set. The research result can provide a convenient and reliable non-destructive method for determining the water stress status of pepper plants.
