Inversion and validation of soil water-holding capacity in a wild fruit forest, using hyperspectral technology combined with machine learning

Soil water retention is a critical aspect of water conservation. To quantitatively assess the Soil Water-Holding Capacity (SWHC), this study focused on a typical wild fruit forest in Xinjiang, China. The spectral characteristics of the forest canopy were employed as a bridge to enhance the sensitivity between the SWHC and various vegetation indices using mathematical statistical methods. This study integrated hyperspectral technology with machine learning algorithms to model complex nonlinear relationships and to select the optimal SWHC model. The spatial distribution of SWHC in the wild fruit forests of Emin County was retrieved using Sentinel-2 imagery. The results revealed a significant negative correlation between SWHC and the smoothed leaf spectral reflectance, with the best correlation coefficient was r = - 0.59. The use of third-order derivatives and logarithmic second-order derivatives further enhanced this correlation, yielding optimal coefficients of r = - 0.78 and r = - 0.72, respectively. Moreover, uncertainty analysis demonstrated that the SWHC estimation model constructed using the Random Forest (RF) algorithm exhibited the highest stability, with a coefficient of determination R² = 0.73, RMSE = 0.158, and RPD = 1.90. The spatial inversion results indicated that SWHC values were relatively higher in areas with dense wild fruit forest coverage and valley terrain. This study is the first to jointly incorporate high-order spectral derivatives and uncertainty analysis into the modeling of SWHC in wild fruit forests, underscoring the advantages of spectral feature enhancement and variable perturbation analysis for improving model stability. The findings provide novel insights into SWHC inversion and offer valuable references for ecological restoration, enhancing the water conservation function of wild fruit forests, and formulating targeted management strategies.