Abstract:

In order to quickly detect the soil water content and achieve timely monitoring of tomato plant growth, the average spectral reflectance of 304 soil samples was extracted using hyperspectral imaging technology. The original spectra underwent preprocessing and optimization by removing outliers, dividing the sample set, applying three preprocessing methods, successive projections algorithm (SPA), uninformation variable elimination (UVE), iterative retained information variable (IRIV), genetic partial-least-squares algorithm (GAPLS) to extract the feature wavelengths. Following this, a partial-least-squares regression (PLSR) model was established based on the identified feature wavelengths. Utilizing these preferred feature wavelengths, multiple models were then constructed, including the PLSR model, multiple linear regression (MLR) model, principal component regression (PCR) model, and convolutional neural network (CNN) model. The results showed that: the preferred moving average smoothing (MAS) preprocessing of soil water content was applied, and the quantitative prediction model of soil water content, which was established using the characteristic wavelength extracted by the IRIV method, proved to have the best effect (R_c=0.7167, RMSE_c=0.0193; R_p=0.6631, RMSE_P=0.0272). Additionally, the model of soil water content based on IRIV-CNN also demonstrated good performance (R_c=0.7655, RMSE_c=0.0172). This study holds great practical significance and usefulness for the development of water utilization efficiency in the facility vegetable industry, as well as the scientific management of water in tomato crops. It currently provides technical support for online monitoring of tomato plant growth.

Key words: tomato, soil moisture content, hyperspectral imaging, detection, feature wavelength selection, tomato plant growth monitoring