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

Abstract

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.

Keywords: Estimation model; Hyperspectral; Machine learning; Soil water-holding capacity; Wild fruit forest.

Fig. 1

Comparison of spectral characteristics of leaves under different SWHC levels: In the figure, A represents the curve of the mean leaf spectral reflectance corresponding to SWHC ≤ 0.49 g·cm⁻², B represents the curve of the mean leaf spectral reflectance corresponding to 0.49 g·cm⁻² < SWHC ≤ 0.65 g·cm⁻² and C represents the curve of the mean leaf spectral reflectance corresponding to SWHC > 0.65 g·cm⁻².

Fig. 2

Leaf spectral transformation diagram: (a) is the smoothed leaf spectral curve. (b) is the leaf spectral curve after the first derivative. (c) is the spectral curve after the second derivative. (d) is the spectral curve after the logarithm of the second derivative. (e) is the spectral curve after the third derivative.

Fig. 3

SWHC and leaf spectral correlation coefficient heatmap: (a) show the correlation coefficient heatmap between multiple vegetation indices and SWHC, (b) is the correlation heatmap between the original leaf spectrum and SWHC, (c) is the correlation heatmap between the first derivative and SWHC, (d) is the correlation heatmap between the second derivative and SWHC, (e) is the correlation heatmap between the logarithm of the second derivative conductivity and SWHC, (f) is the correlation heatmap between the third derivative and SWHC.

Fig. 4

Monte Carlo uncertainty analysis.

Fig. 5

SWHC distribution map of wild fruit forest in Emin county (Mapping: Created by ArcMap, version10.5 http://www.esri.com/. Satellite imagery data: Sentinel-2 A from https://dataspace.copernicus.eu/. Image processing and extraction: Created using SNAP, version 9.0.0 https://step.esa.int/main/download/snap-download/ and MATLAB, version R2022a https://www.mathworks.com/products/matlab.html. Boundaries made with free vector data provided by National Catalogue Service for Geographic Information(https://www.webmap.cn/commres.do?method=dataDownload)).

Fig. 6

Importance results of input variables for RF and BPNN models tested by MDA: X1 represents R’2136 nm, X2 represents R’’1959 nm, X3 represents R’’’488 nm, X4 represents LOG(R1247 nm)’’, X5 represents LOG(1/TCARI), and X6 represents LOG(1/MCARI).

Fig. 7

Study area. (a) The geographical location of Emin County, Xinjiang. (b) the location of the wild fruit forest scenic area. (c) the landscapes of the wild fruit forest, (d) the river basin landscape. (e) typical vegetation represented by wild apples, respectively. ((a–b) Mapping: Created by ArcMap, version10.5 http://www.esri.com/. Data source: Google Earth https://earth.google.com. Boundaries made with free vector data provided by National Catalogue Service for Geographic Information (https://www.webmap.cn/commres.do?method=dataDownload. (c–e) Photos were taken at the wild fruit forest study sites during the field campaign.)

Fig. 8

Technology roadmap.