Optimized feature selection and machine learning for non-destructive estimation of soil volumetric water content in Chinese cabbage using hyperspectral imaging

Abstract

Soil volumetric water content (SVWC) is a critical factor in plant health, influencing water uptake, nutrient transport, and overall physiological performance. Adverse environmental conditions like drought and high temperatures challenge crop growth and reduce yields. Accurate monitoring of SVWC is essential for optimizing growing conditions, preventing water stress, and promoting sustainable agriculture. This study explores a non-destructive method for predicting SVWC in Chinese cabbage seedlings using short-wave infrared (SWIR, 894–2504 nm) hyperspectral imaging coupled with machine learning. Daily hyperspectral images and corresponding SVWC measurements were collected over three days following irrigation cessation, resulting in a dataset of 2700 spectra. Gaussian process regression (GPR) and support vector regression (SVR) models were applied, with Lasso and Ridge regression used for feature selection. The models were evaluated using all spectral bands (E164) and 30 selected bands (L30 and R30). The GPR model with Lasso-selected bands and smoothing preprocessing achieved the highest accuracy (R² = 0.87, RMSE = 1.33). The SVR model with smoothing preprocessing and the entire spectral range demonstrated R² = 0.82 and RMSE = 1.52. Multivariate regression models using 14 shared bands selected by Lasso and Ridge regression yielded moderate performance (R² = 0.67, RMSE = 2.07). These findings highlight the potential of hyperspectral imaging combined with machine learning for non-destructive SVWC prediction, enabling early crop detection of water stress.