Inversion of soil pH during the dry and wet seasons in the Yinbei region of Ningxia, China, based on multi-source remote sensing data

Highlights

• Resample the Ground-spectrum data correlated well with the image data.

• Sensitive spectra extracted through multiple screening methods.

• Various models have been established to invert the pH value in the dry and wet seasons.

• The RPD of optimal model is 5.76.

• Quantitative inversion of soil pH value based on optimal monitoring model.

Abstract

Dynamic monitoring of soil pH changes across different seasons is crucial to understanding soil alkalization and preventing land degradation in arid regions. Here, we explored the potential use of ground-measured hyperspectral data and Landsat 8 OLI remote-sensing images in the inversion of soil pH during the dry and wet seasons. The measured and image data of spectral reflectance and corresponding soil pH data obtained in the laboratory were used to analyze the spatiotemporal patterns of soil pH in the Yinbei region, Ningxia, China. First, the hyperspectral data were resampled to match the range of the image bands; then 11 spectral indices were calculated based on the two sets of spectral data. Principal component analysis (PCA), stepwise regression (SR), and gray relational analysis (GRA) were used to select feature variables (sensitive bands and spectral indices) from the spectral data. The back propagation neural network (BPNN), support vector machine (SVM), ridge regression (RR), and geographically weighted regression (GWR) were used to construct soil pH inversion models. A total of 24 models established using the different methods were compared in terms of the determination coefficients of the calibration set (R_c²) and prediction set (R_p²), root mean square error (RMSE), and relative percent deviation (RPD). The results revealed that the mean soil pH in the study area was higher during the dry season (9.28) than the wet season (9.11). The resampled hyperspectral data correlated well with the image data under different levels of soil alkalization (R² > 0.9652). Of the different models examined, the global regression models (BPNN, SVM, and RR) were superior to the local regression model (GWR), with better inversion results obtained in the wet than the dry season. The BPNN and SVM models performed better than the RR model, and the PCA-SVM model based on measured data in the dry season achieved the best overall performance (R_p² = 0.9724 and RPD = 5.76). These findings provide valuable information on the distribution of Solonetzs land in the study area, facilitating management of soil alkalization and degradation in the Yinbei region.

Introduction

Like drought, high salinity is one of the major issues limiting global agricultural production, particularly in arid and semi-arid regions. It is estimated that almost 950 million hectares of soil worldwide has low fertility due to salinization (Aldabaa et al., 2015). Soil alkalization is a process of soil salinization during which the replaceable sodium in soil colloids undergoes chemical substitution with other cations, resulting in the absorption of a large amount of Na⁺ ions; Na⁺ then binds CO₃²⁻ to form Na₂CO₃, leading to increased soil pH and the formation of alkaline soil. Alkaline soil is characterized by a high pH and poor physical properties (e.g., poor structure and low permeability), resulting in soil compaction and retarded plant growth. Soil alkalization therefore has severe negative impacts on crop yield and agricultural productivity, threatening ecosystem health and economic sustainability (Li et al., 1998; Indrasumunar et al., 2012). According to statistical analyses, soil alkalinity is strongly affected by the monsoon climate (Li et al., 2016). Since seasonal precipitation in arid regions is unevenly distributed, the characteristics of alkaline soils show dynamic changes in terms of space and time. Thus, dynamic monitoring is essential in fully understanding the current status of soil alkalization and developing effective strategies aimed at soil restoration and land reclamation (Bannari et al., 2018). Zhang et al. (2018a) revealed that the annual mean precipitation and annual mean air temperature are strong predictors of regional soil pH levels (an indicator of soil acidity or alkalinity). Understanding the changes in soil pH that occur during different seasons is therefore important in understanding the dynamics of soil alkalization and preventing soil degradation.

Various techniques have been used to study soil pH, including traditional soil sampling and laboratory analyses, as well as field measurements using newly-developed sensors. The recent development of remote sensing technology (e.g., satellite platforms, global positioning systems, unmanned aerial vehicles, and ground sensors) combined with geographic information systems have provided new opportunities for research into soil science (Chai et al., 2015). Furthermore, with the application of spectral analysis, remote sensing has become a major approach during large-scale monitoring and evaluation of soil alkalization (Roelofsen et al., 2015; Wang et al., 2018a; Davis et al., 2019; Ren et al., 2020; dos Teixeira et al., 2020), and a number of studies have used remote sensing images to estimate soil pH (Liu et al., 2013; Li et al., 2018; Fu et al., 2020). For example, Miao et al. (2018) used the gray correlation model to estimate soil pH values across Hengshan County in Shanxi Province, China, based on hyperspectral data, while Zhang et al. (2018a) combined ground-measured hyperspectral and satellite image spectra to establish a soil pH estimation model with improved accuracy. Although the spectral indices used in previous studies can provide accurate estimates of soil alkalization, their effectiveness varies with the environmental conditions, land cover, soil type, and salt content (Wang et al., 2017; Farahmand et al., 2020). During spectral analysis, establishment of a reliable quantitative model is a prerequisite for accurate predictions of the type, composition, and nature of unknown samples. However, due to the complicated causes of soil alkalization and the sophisticated composition of soil pH, contradictory results have been obtained from different regions (Wang et al., 2017; Wang et al., 2019). Appropriate selection of feature variables is therefore crucial for quantitative analysis of spectral data. The most commonly used methods of variable selection include Pearson's correlation (Wang et al., 2019), Principal component analysis (PCA) (Xu et al., 2018), stepwise regression (SR) (Wang, 2019), gray relational analysis (GRA) (Wang, 2019), and variable importance in projection (Wang, 2019).

With inversion models, estimations of soil properties using soil spectra tend to be based on statistical analyses, such as multivariate stepwise regression (MLR) and partial least squares regression (PLSR) (Luan et al., 2013). In comparison, machine learning inversion models have strong nonlinear fitting capabilities and excellent data mining qualities that increase the use of spectral reflection data (Li et al., 2020). For example, Wang et al. (2018b) found that the machine learning algorithm showed strong data mining capability when a large number of variables were involved, and their correlations with soil salt content were relatively low. Li et al. (2017) found that the back propagation neural network (BPNN), support vector machine (SVM), and random forest (RF) also showed high inversion accuracy, far superior to statistical analysis models. However, the above-mentioned machine learning models tend to regard the relationship between independent and dependent variables as globally unchanged, ignoring those changes caused by spatial variability (Wu et al., 2016). To solve this issue, Fotheringham et al. (1996) proposed a local model, geographically weighted regression (GWR), which improved the inversion accuracy and effectiveness by introducing an unsteady relationship between the independent and dependent variables. To date, a number of models have been established to predict levels of soil salinization (Yuan et al., 2019; Li et al., 2019); yet, few reports have compared their prediction accuracy on alkalized soil when pH is used as the major indicator.

The Yinbei region in Ningxia, China, is an arid area with wide-spread occurrence of saline-alkaline soil. The total area of salinized soil in the Yinbei region is ~11.4 × 10⁴ hm² (The alkaline represented takyr solonetzs), approximately 2.15 times the area of non-saline farmland. Because of its flat terrain and deep soil layer, the Yinbei region is considered an important reserve of arable land in Ningxia. Accurate and efficient analyses of local soil salt contents and pH levels are therefore critical in assessing soil health and optimizing land management in this region. A number of studies have examined the mechanisms, and carried out dynamic monitoring and prediction of soil salinization on a local scale (Zhang et al., 2018b; Zhang et al., 2019a); however, most of these studies focus on the total soil salt content over a specific period of time (Sidike et al., 2014; Zhang et al., 2019b). Monitoring the spatiotemporal patterns of soil pH during different seasons is essential in understanding the dynamics of soil alkalization and facilitating soil improvements in the Yinbei region.

The objectives of this study were: (1) to explore the differences in soil pH during the dry and wet seasons, and determine the relationship between ground-measured hyperspectral data and Landsat 8 OLI image spectral data in the Yinbei region; and (2) to construct an optimal soil pH inversion model using BPNN, SVM, RR, and GWR based on different input variables selected by PCA, RS, and GRA. The results could provide a useful framework for the establishment of a spectral quantitative model of Solonetzs in this region.