A Three-Dimensional Feature Space Model for Soil Salinity Inversion in Arid Oases: Polarimetric SAR and Multispectral Data Synergy

Abstract

Effective soil salinity monitoring is crucial for sustainable land management in arid regions. Most current studies face limitations by relying solely on single-source data. This study presents a novel three-dimensional (3D) optical-radar feature space model combining Gaofen-3 polarimetric synthetic aperture radar (SAR) and Sentinel-2 multispectral data for China’s Yutian Oasis. The random forest (RF) feature selection algorithm identified three optimal parameters: Huynen_vol (volume scattering component), RVI_Freeman (radar vegetation index), and NDSI (normalized difference salinity index). Based on the interactions of these three optimal features within the 3D feature space, we constructed the Optical-Radar Salinity Inversion Model (ORSIM). Subsequent validation using measured soil electrical conductivity (EC) data (May–June 2023) demonstrated strong model performance, with ORSIM achieving R² = 0.75 and RMSE = 7.57 dS/m. Spatial analysis revealed distinct salinity distribution patterns: (1) Mildly salinized areas clustered in the central oasis region, and (2) severely salinized zones predominated in northern low-lying margins. This spatial heterogeneity strongly correlated with local topography-higher elevation (south) to desert depression (north) gradient. The 3D feature space approach advances soil salinity monitoring by overcoming traditional 2D limitations while providing an accurate, transferable framework for arid ecosystem management. Furthermore, this study significantly expands the application potential of SAR data in soil salinization research.

1. Introduction

Soil salinization is recognized as a principal form of land degradation and a significant natural hazard in arid regions, substantially contributing to land desertification, impairing the efficient use of water and soil resources, and posing a threat to ecosystem stability. Consequently, soil salinization has become a pressing global environmental issue requiring immediate attention [1]. Although all soils contain some water-soluble salts essential for plant nutrient uptake, excessive salt accumulation in arid regions cannot be effectively alleviated through conventional irrigation or crop management practices. The misuse of irrigation techniques, compounded by climate change-induced aridity, has accelerated secondary soil salinization [2]. Furthermore, shallow-rooted crops extract limited groundwater, often leading to a rising water table and the accumulation of salts near the surface or within upper soil layers. This process creates highly unfavorable conditions for most crops, significantly reducing yields and directly threatening food security [3,4]. Globally, salinized soils constitute approximately 20% of all arable land, with projections suggesting this figure may reach 50% by 2050 [5,6]. As a result, soil salinization has been termed the “cancer” of the Earth, representing a critical obstacle to achieving sustainable development goals [7].

In China, salt-affected soils are primarily concentrated in arid and semi-arid regions, with Xinjiang being the most severely affected [8]. Oases within these environments serve as the core of regional ecosystems and play a crucial role in preserving local biodiversity and ensuring sustainable agricultural development [9]. However, persistent disruptions to the water–salt balance in oasis zones have rendered soil salinization management and control in Xinjiang increasingly challenging [10]. Thus, the efficient and accurate extraction and dynamic monitoring of soil salinization within oases are essential for ecological protection and the sustainable advancement of agriculture in Xinjiang.

Conventional techniques employed to assess soil salinization typically rely on on-site investigations and soil sample collection, followed by laboratory-based physicochemical analysis [11]. Although scientifically rigorous, these methods are time-consuming, labor-intensive, and spatially limited, often lacking representativeness, making them unsuitable for large-scale dynamic monitoring applications [12]. Since the 1970s, the rapid advancement of remote sensing (RS) technology has enabled new, large-scale monitoring approaches for soil salinization. RS provides several advantages, including high revisit frequency, broad spatial coverage, rapid data acquisition, and strong data timeliness, making it highly promising for monitoring soil salinization and extracting soil salinity information in arid regions [13]. However, optical RS is limited by cloud cover and atmospheric conditions, reducing data availability and continuity. The spatiotemporal variability and vertical heterogeneity of soil salinity further compromise the extraction accuracy of optical RS [14]. Additionally, since optical RS primarily detects the spectral characteristics of the soil surface, it is prone to the “same object, different spectra” and “different object, same spectrum” phenomena, leading to errors in the classification and extraction of soil salinization information [15]. The limited penetration depth of optical remote sensing further constrains its capacity to assess vertical salinity distribution in soils, underscoring the importance of integrating and optimizing multi-source remote sensing approaches [16,17].

To overcome these limitations, microwave RS has emerged as an increasingly valuable tool for saline–alkali land characterization [18,19]. As an active microwave sensing technology, radar RS provides distinct advantages, including: (1) negligible atmospheric attenuation, (2) unimpeded operation through clouds and precipitation, (3) continuous day-night monitoring capability, and (4) enhanced subsurface information acquisition through longer wavelength radiation [20,21]. These attributes make radar RS particularly advantageous for soil salinization monitoring compared to other techniques [17]. Soil salinity variations influence electrical conductivity, altering the dielectric constant and the radar backscatter coefficient. Synthetic aperture radar (SAR) demonstrates high sensitivity to dielectric property changes, enabling effective soil salinity monitoring [22]. Given this capability, SAR data hold significant potential for salinization research [23]. Additionally, SAR is highly responsive to soil moisture, a critical factor indirectly affecting salinity dynamics in irrigated oasis farmlands. Therefore, integrating multi-source RS datasets, especially those incorporating SAR imagery, offers a promising pathway for robust soil salinization monitoring in arid regions.

Polarimetric synthetic aperture radar (PolSAR) can acquire comprehensive polarization information about surface features, requiring further interpretation for effective target identification and analysis [24]. With the advancement of PolSAR data acquisition technology, research efforts have increasingly focused on maximizing the utility of its abundant informational content [25]. Recent studies have employed various polarimetric target decomposition methods to extract polarization parameters associated with ground physical properties, facilitating detailed analysis of hidden information and surface scattering mechanisms using PolSAR data [26,27]. PolSAR target decomposition techniques elucidate the scattering mechanisms of terrestrial features by decomposing complex polarimetric information into physically meaningful scattering components, thereby improving the understanding of the physical characteristics of surface targets [28]. Through polarimetric decomposition of PolSAR data, relationships between specific scattering mechanisms and decomposition components can be investigated, providing crucial theoretical and methodological support for soil moisture and salinity research. Thus, polarimetric target decomposition represents an effective approach to derive soil salinization information from synthetic aperture radar, enabling continuous observation of soil salinity variations and offering scientific foundations and decision support for agricultural management and land use in arid regions [29].

Since the early 21st century, extracting soil information from a multi-dimensional spectral feature space has garnered increasing attention, and several inversion methods based on spectral feature space for soil moisture retrieval have been developed [30,31,32,33]. Many researchers have utilized feature space techniques to obtain soil salinity information [34,35,36]. However, most existing feature space models rely exclusively on optical or single-source RS data and fail to leverage the rich informational content of alternative data types, such as radar imagery, which exhibits significant potential for soil salinity inversion. Moreover, existing feature space models are predominantly limited to 2D models, overlooking the potential of 3D feature space for capturing more complex interactions.

To address the challenges associated with quantitative soil salinity inversion in arid regions and to enhance retrieval precision and efficiency, this study selected the Yutian Oasis as the research area. The primary data sources included fully polarimetric Gaofen-3 (GF-3) SAR data, multispectral Sentinel-2 optical RS imagery, and field-measured soil salinity. By integrating radar polarimetric target decomposition, the Radar Vegetation Index (RVI), optical indices, and feature space methodologies, this investigation aimed to introduce the theoretical framework of feature space into the realm of radar RS and develop a quantitative radar-based inversion model for soil salinization. Subsequently, a soil salinity spatial distribution map was generated for the study region.

This study investigates the synergistic potential of radar and optical RS data through a 3D feature space framework, intending to develop an integrated soil salinity inversion model. By leveraging multi-source RS data fusion, this research aims to establish an efficient, large-scale monitoring approach for soil salinization. The findings are expected to provide a theoretical foundation for sustainable oasis land resource management in arid regions.

2. Study Area

The Yutian Oasis (36°36′–37°18′ N, 81°10′–81°48′ E) is located in southwestern Xinjiang, China, within the central Eurasian continent. This ecologically significant region forms a transitional zone between the Kunlun Mountains’ northern foothills and the Taklamakan Desert’s southern margin, characterized by a warm temperate arid desert climate [37]. The area exhibits pronounced mountain–basin topography, creating a distinct oasis–desert ecotone that is particularly sensitive to environmental changes [38]. Geomorphologically, the southern sector features mountainous and hilly terrain, while the northern portion consists of desert plains, with elevations ranging dramatically from 1180 to 5460 m (a 4000 m vertical gradient) [39]. Climatically, the region demonstrates extreme aridity, with an annual precipitation of merely 14 mm in the plain oasis area and potential evaporation reaching 2500 mm. The region’s severe water deficit is partially alleviated by snowmelt from mountain glaciers and limited groundwater recharge. Coupled with the relatively high mean annual temperature (12.4 °C) [40], these extreme conditions have triggered significant ecological degradation, particularly through soil desertification and salinization processes. Such environmental vulnerability, combined with the area’s representative geoclimatic features, makes the Yutian Oasis a prime location for studying fragile oasis ecosystems [41]. These compelling characteristics formed the basis for selecting this region as our study area (Figure 1).

3. Materials and Methodology

3.1. Field Data Collection

Field sampling was conducted from 22 May to 9 June 2023. Sampling points were strategically selected based on the spatial heterogeneity and severity of soil salinization within the study area, with locations determined using GPS positioning supplemented by auxiliary data, including topographic and land use maps of Yutian County, to ensure comprehensive coverage of diverse land cover types. The collected samples represent various soil properties, such as texture, salinization type, and land use/cover categories. Primary sampling zones included the core oasis, its periphery, and the transitional belt between oasis and desert, resulting in a total of 69 field samples collected within the coverage area of GF-3 radar imagery.

Field sampling employed the five-point composite sampling method [42], where surface soil samples (0–10 cm depth) were obtained from a 10 m × 10 m quadrat at each site. Approximately 500 g of soil was collected from each sampling point, then sealed and transported to the laboratory for analysis. During laboratory processing, soil samples were air-dried, ground, and cleaned of debris. A soil–water suspension (soil–water ratio of 1:5) was prepared, vigorously shaken 200 times, allowed to settle for 4 h, and then filtered. The electrical conductivity (EC) of the filtrate was measured using a REX DZS-706F-A multiparameter analyzer (Shanghai REX Instrument Co., Ltd., Shanghai, China).

3.2. RS Data Acquisition and Pre-Processing

3.2.1. SAR Data

GF-3 data, temporally aligned with field sampling, were selected for this study. The GF-3 satellite operates in the C-band and is equipped with multi-polarization, high-resolution SAR sensors. It follows a sun-synchronous orbit at an altitude of approximately 755 km, with a revisit period of less than 5 days, providing high spatial resolution, wide swath imaging, multi-polarization, and various imaging modes [43]. The GF-3 system supports four polarization modes (HH, HV, VH, and VV) and offers 12 imaging modes [44]. The GF-3 data used in this study is a Level 1A Single Look Complex (SLC) fully polarimetric product acquired on 9 June 2023. Key imaging parameters are summarized in

Table 1. Technical specifications of Gaofen-3 (GF-3) SAR data used in this study, including acquisition parameters and system characteristics critical for soil salinity inversion.

Parameters	Data
Acquisition Date	9 June 2023
Product Type	Level 1A Single Look Complex (SLC)
Radar Center Frequency	5.4 GHz (C-Band)
Incident angle	35°~37°
Acquisition Type	Quad-Polarization Strip I (QPSI)
Nominal Resolution	5.54 m × 2.25 m (Range × Azimuth)
Polarization	Quad-pol (HH, HV, VH, VV)

.

3.2.2. Optical Multispectral Data

Sentinel-2 data were also utilized in this study. Sentinel-2 is a high-resolution multispectral imaging satellite system equipped with a Multi-Spectral Instrument (MSI) that encompasses 13 spectral bands. It operates at an orbital altitude of 786 km, with a single-scene swath width up to 290 km. The Sentinel-2 constellation consists of two satellites, Sentinel-2A and Sentinel-2B, each with a revisit period of 5 days [45]. For this study, two Level-2A scenes with less than 10% cloud cover, acquired on 10 June 2023, were selected. Detailed parameters are provided in

Table 2. Technical specifications of Sentinel-2B multispectral imagery used for soil salinity inversion, showing key acquisition parameters for optical-radar data fusion.

Parameters	Data
Acquisition Date	10 June 2023
Satellite	Sentinel-2B
Product Type	Level-2A
Number of bands	13 spectral bands
Resolution	10 m, 20 m, 60 m
Swath Width	290 km
Tile Numbers	T44SNG and T44SNF

.

3.2.3. Data Pre-Processing

The acquisition of RS imagery is often affected by various external environmental conditions and inherent sensor characteristics, which may result in data loss and distortion. Therefore, comprehensive pre-processing of RS data is required to mitigate these disturbances and enhance the reliability and utility of the resulting datasets [46]. In this study, both Sentinel-2 and GF-3 images underwent dedicated pre-processing procedures to maximize their effectiveness and applicability.

The Sentinel-2 imagery used in this research had already undergone radiometric calibration, atmospheric correction, geometric correction, and UTM WGS projection. The two image tiles covering the study area are labeled T44SNF and T44SNG. Further pre-processing, conducted using SNAP 9.0^®, ENVI 5.3^®, and other software, included: (1) super-resolution synthesis, (2) image mosaicking, and (3) regional subset extraction [47].

The GF-3 data consisted of Level 1A Single Look Complex (SLC) products. To meet the requirements of subsequent analysis, radar image pre-processing was performed using tools such as ENVI 5.3^®, ENVI IDL 8.5^®, and PolSAR Pro 6.0^® (Figure 2). Pre-processing of GF-3 data was divided into two main categories. For polarimetric information extraction, steps included (1) multi-looking, (2) speckle filtering, (3) geocoding, and (4) radiometric calibration and conversion to decibel values [48]. For polarimetric decomposition, the workflow comprised (1) format conversion and S2 matrix construction, (2) T3 matrix transformation and multi-looking, (3) speckle filtering, (4) polarimetric decomposition—resolving complex polarimetric information into distinct physical components for comprehensive analysis—and (5) orthorectification and geocoding [10].

All images were resampled to a standardized spatial resolution of 10 × 10 m and uniformly projected to the UTM WGS-84 coordinate system, Zone 44N (Figure 3).

3.3. Methodology

3.3.1. Target Decomposition of Polarimetric SAR

In recent years, target decomposition has garnered increasing interest, and development has accelerated because of its pivotal role in elucidating target scattering characteristics [49]. A variety of polarimetric decomposition methods have been developed to resolve the scattering characteristics of different surface objects in SAR data, further unlocking the wealth of information contained in SAR imagery [50,51,52,53,54,55]. The repertoire of polarimetric decomposition methods has continued to expand, enabling the extraction of various scattering components such as surface scattering, double-bounce scattering, and volume scattering, each of which is invaluable for distinguishing between different scattering elements [56]. SAR scattering refers to the process in which a portion of the radar-transmitted electromagnetic waves, upon interacting with the target surface, is scattered back toward the radar receiver [57]. This physical process is governed by three fundamental scattering mechanisms: volume scattering within multi-layer media, surface scattering at dielectric interfaces, and double-bounce scattering from dihedral structures [56]. In polarimetric decomposition, double-bounce scattering typically occurs from tree trunks or buildings, surface scattering is mainly associated with bare ground, and volume scattering is dominant in vegetation-covered areas [58].

As shown in Figure 4, double-bounce scattering typically represents signals from vertical structures (e.g., tree trunks, buildings), surface scattering corresponds to bare soil, and volume scattering is primarily attributed to vegetation cover [58].

To effectively utilize the polarimetric information in GF-3 full-polarization radar data, this study drew on established feasibility assessments [10,56,59] and selected six representative polarimetric decomposition methods: Freeman, van Zyl, Cloude, Huynen, Yamaguchi, and AnYang. First, a T3 polarization matrix was generated. Second, six polarimetric decomposition techniques were applied to extract the scattering components from GF-3 data. Finally, 19 polarization features were extracted from the PolSAR data (

Table 3. Polarimetric decomposition methods applied to GF-3 SAR data and their corresponding scattering components.

Polarimetric Decomposition	Number of Components	Target Scattering Component
Freeman Durden	3	Freeman_odd, Freeman_vol, Freeman_dbl
van Zyl	3	van Zyl_odd, van Zyl_vol, van Zyl_dbl
Cloude	3	Cloude_odd, Cloude_vol, Cloude_dbl
Huynen	3	Huynen_odd, Huynen_vol, Huynen_dbl
Yamaguchi	4	Yamaguchi_odd, Yamaguchi_vol, Yamaguchi_hlx, Yamaguchi_dbl Yamaguchi_dbl, Yamaguchi_hlx
AnYang	3	AnYang_odd, AnYang_vol, AnYang_dbl

Note: In the table, _odd is surface scattering, _dbl is double-bounce scattering, _vol is volume scattering, and _hlx is helix scattering.

).

3.3.2. Surface Parameters Responsive to Soil Salinity

The scientific and rational selection of surface parameters indicative of soil salinity is a fundamental prerequisite for the efficient extraction of soil salinity information. In this study, representative surface parameters were divided into two principal categories: optical multispectral RS indices and radar RS indices. Optical indices are widely recognized for their utility in the analysis and interpretation of RS imagery [60], and are frequently applied in the monitoring and assessment of vegetation cover, soil salinity, soil moisture, and crop health [61]. In the context of arid and semi-arid regions, optical indices are indispensable tools that capture the spectral reflectance characteristics of surface objects, thereby directly or indirectly reflecting variations in soil salinity [62,63]. Their applications primarily fall into two domains: (1) vegetation-responsive indices, such as the Normalized Difference Vegetation Index (NDVI), which exhibits a strong negative correlation with soil salinity in arid environments [64]; and the Soil-Adjusted Vegetation Index (SAVI), as well as the Modified Soil-Adjusted Vegetation Index (MSAVI), which incorporate soil brightness adjustment factors to minimize the influence of bare soil on vegetation signals—making them particularly suitable for salinity inversion in sparsely vegetated areas [65,66]. (2) Salinity-sensitive indices, such as the Salinity Index (SI) and the Normalized Difference Salinity Index (NDSI), which utilize specific spectral band combinations to effectively differentiate soils across a gradient of salinity levels [67].

If vegetation cover exceeds 60%, optical indices are prone to spectral saturation, obscuring soil salinity signals; secondly, in bare or sparsely vegetated areas, spectral signals associated with salinity are easily confounded with those related to soil moisture or surface roughness [68]. Additionally, cloud cover can lead to data gaps during key crop growth periods. In light of these limitations, this study integrates optical indices, radar vegetation indices, and polarimetric decomposition scattering components to construct a multi-dimensional feature space, thereby enhancing the accuracy and reliability of soil salinity inversion. The Radar Vegetation Index (RVI), derived from the ratio of cross-polarized to co-polarized backscattering in fully polarimetric SAR data, effectively quantifies vegetation canopy structural complexity and exhibits superior penetration capability compared to optical indices [7,69]. Ultimately, six representative optical indices and six distinct radar vegetation indices were selected for this study. The specific parameters are presented in

Table 4. Optical vegetation and salinity indices derived from Sentinel-2 multispectral data, showing mathematical formulations and their references.

Optical Index	Formulation	Ref.
Normalized Difference Vegetation Index (NDVI)	$\left(\mathrm{NIR}-\mathrm{R}\right)/(\mathrm{NIR}+\mathrm{R})$	[70]
Soil Adjusted Vegetation Index (SAVI)	$\left(1+\mathrm{L}\right)\times \left(\mathrm{NIR}-\mathrm{R}\right)/(\mathrm{NIR}+\mathrm{R}+\mathrm{L})$	[71]
Modified Soil Adjusted Vegetation Index (MSAVI)	$\left(2\times \mathrm{NIR}+1-\sqrt{{\left(2\times \mathrm{NIR}+1\right)}^2-8\times \left(\mathrm{NIR}-\mathrm{R}\right)}\right)/2$	[68]
Salinity Index (SI)	$\sqrt{{\mathrm{G}}^2+{\mathrm{R}}^2{+\mathrm{NIR}}^2}$	[72]
Vegetation Soil Salinity Index (VSSI)	$2\times \mathrm{G}-5\times (\mathrm{R}+\mathrm{NIR})$	[73]
Normalized Difference Salinity Index (NDSI)	$\left(\mathrm{R}-\mathrm{NIR}\right)/\left(\mathrm{R}+\mathrm{NIR}\right)$	[74]

Note: Ref denotes reference. Band designations: B3 (Green, 560 nm), B4 (Red, 665 nm), B8 (NIR, 842 nm). L is the soil adjustment factor (typically 0.5 for moderate vegetation).

and

Table 5. Polarimetric radar vegetation indices derived from GF-3 quad-pol SAR data, showing decomposition-based formulations and their references.

Radar Vegetation Index	Formulation	Ref.
RVI_Kim	$8\times \mathrm{H}\mathrm{V}/(\mathrm{HH}+\mathrm{VV}+2\times \mathrm{HV})$	[75]
RVI_HH	$4\times \mathrm{H}\mathrm{V}/(\mathrm{HH}+\mathrm{HV})$	[76]
RVI_VV	$4\times \mathrm{V}\mathrm{H}/(\mathrm{VV}+\mathrm{VH})$	[77]
RNDVI	$\left(\mathrm{VH}-\mathrm{VV}\right)/\left(\mathrm{VH}+\mathrm{VV}\right)$	[78]
RVI_van Zyl	$4\times {\mathrm{\lambda }}_3/\left({\mathrm{\lambda }}_1{+\mathrm{\lambda }}_2{+\mathrm{\lambda }}_3\right)$	[79]
RVI_Freeman	${\mathrm{F}}_{\mathrm{V}}/({\mathrm{F}}_{\mathrm{S}}+{\mathrm{F}}_{\mathrm{D}}+{\mathrm{F}}_{\mathrm{V}})$	[58]

Note: Ref denotes reference. λ₁, λ₂, λ₃ represent eigenvalues from polarimetric decomposition. F_s, F_d, and F_v represent surface, double-bounce, and volume scattering components.

.

3.3.3. Optimal Feature Component Selection

The accuracy of soil salinity inversion based on feature space is highly dependent on the scientific selection of features responsive to soil salinity, with a particular emphasis on identifying optimal components. In this study, optimal component selection involved the rigorous screening of both polarimetric decomposition scattering components and RS indices. Polarimetric decomposition yields surface, double-bounce, and volume scattering components, which facilitate land cover discrimination by elucidating different scattering mechanisms [80]. Although polarimetric decomposition provides a wealth of scattering features, not all extracted components are suitable for soil salinity inversion, and the high dimensionality of features generated by decomposition methods can increase computational complexity and introduce redundancy. Moreover, the sensitivity, discriminative power, and applicability of radar and optical indices for soil salinity retrieval in arid regions remain insufficiently characterized.

To maximize model performance and reduce subjective bias, we employed the random forest (RF) feature selection algorithm to objectively identify the most salient parameters associated with soil salinity from among the polarimetric decomposition scattering components, optical indices, and radar indices, thereby supporting precise and reliable monitoring and visualization of soil salinity. The RF feature selection algorithm is renowned for its interpretability, ease of implementation, and robust resistance to overfitting, and has been widely adopted for feature selection tasks [56,81]. This algorithm quantifies the importance of each feature by measuring changes in out-of-bag (OOB) prediction error [56]. The specific steps are as follows: (1) Calculation of the original OOB error: each decision tree is trained on a bootstrap sample, and the unsampled data comprise the OOB set, which is used to evaluate the initial prediction error for each tree. (2) Feature perturbation and error reassessment: the feature under evaluation is randomly permuted in the OOB data, predictions are recomputed, and the new error rate is recorded. (3) Importance scoring: the difference in OOB error before and after feature permutation is calculated for each tree (a larger difference indicates a greater contribution of the feature to model prediction), and the average difference across all trees is taken as the feature’s importance score. The calculation formula is as follows:

$$VI\left(X_j\right)={\displaystyle \frac{1}{N_{trees}}}\sum _{t=1}^{N_{trees}}\left({Error}_t^{perturbed}\left(X_j\right)-{Error}_t^{original}\right)$$

(1)

In this formula, $VI\left(X_j\right)$ denotes the importance of the feature $X_j$, $N_{trees}$ is the total number of decision trees in the random forest, ${Error}_t^{original}$ represents the prediction error rate of the $t$-th decision tree on the original OOB data, and ${Error}_t^{perturbed}\left(X_j\right)$ signifies the prediction error rate of the $t$-th decision tree on the OOB data after permuting the feature $X_j$.

3.3.4. Feature Space

The feature space is a high-dimensional data representation framework constructed from surface parameters or indices derived from multiple RS sources, such as vegetation indices, salinity indices, land surface temperature, and others [82]. Its core principle is the careful selection of representative features (variables), allowing the physicochemical properties of terrestrial targets, such as soil salinity and moisture, to be mapped as distributions of points within this high-dimensional space. This multi-dimensional reconstruction facilitates the aggregation of sample points, enabling spatial geometric relationships—such as distance, density, and clustering—to quantify correlations between surface parameters and target variables [59]. As an extension of the index-based approach, the feature space method is recognized as a direct, rapid, and highly efficient strategy for RS estimation [82]. Extensive research has shown that composite parameters, obtained by integrating multiple surface indicators, generally deliver greater stability and reliability than single-parameter approaches [83].

In recent years, the construction of multi-dimensional feature space models reflecting the synergistic effects among optimal surface parameters has demonstrated significant potential in the monitoring of surface elements such as soil moisture, drought, and salinity. For example, Ding et al. [34] used multispectral satellite imagery and spectral decomposition techniques to extract three principal indicators, coupled with vegetation cover and soil moisture content as proxies, to construct a 2D feature space for effective monitoring of soil salinization. Similarly, Liu et al. [35] developed feature space models such as Albedo-MSAVI, SI-Albedo, and SI-NDVI, successfully estimating the salinization status of representative saline soils in China. Guo et al. [36] extracted five typical desertification indices from optical imagery and constructed ten distinct feature space models, providing important decision support for desertification control. However, most existing feature space models depend solely on optical or single-source RS data, thus failing to harness the rich informational content of alternative data types, such as radar imagery, which holds significant potential for soil salinity inversion. In response, this study integrates radar polarimetric decomposition scattering components, radar vegetation indices, and optical indices to construct a 3D feature space, thereby comprehensively exploring the synergistic capability and prospects of multi-source optical and radar for soil salinity inversion. The primary research steps are illustrated in Figure 5.

4. Results

4.1. Construction of Optical-Radar Three-Dimensional Feature Space

The construction of an effective feature space requires screening for optimal feature components and applying appropriate normalization procedures [56]. Accordingly, in this study, guided by in situ soil salinity measurements, the RF feature selection algorithm was employed to rank the importance of radar polarimetric decomposition scattering components, radar vegetation indices, and optical indices (Figure 6).

Based on the feature importance ranking, Huynen_vol, RVI_Freeman, and NDSI were ultimately selected as the constitutive components for the 3D feature space. Before constructing the feature space, these features were normalized using the Max-Min normalization method [59], as detailed below:

$${\mathrm{Huynen}}_{{\mathrm{vol}}_{0~1}}={\displaystyle \frac{{\mathrm{Huynen}}_{\mathrm{vol}}-{\mathrm{Huynen}}_{{\mathrm{vol}}_{\min }}}{{\mathrm{Huynen}}_{{\mathrm{vol}}_{\max }}-{\mathrm{Huynen}}_{{\mathrm{vol}}_{\min }}}}$$

(2)

$${\mathrm{RVI}}_{{\mathrm{Freeman}}_{0~1}}={\displaystyle \frac{{\mathrm{RVI}}_{\mathrm{Freeman}}-{\mathrm{RVI}}_{{\mathrm{Freeman}}_{\min }}}{{\mathrm{RVI}}_{{\mathrm{Freeman}}_{\max }}-{\mathrm{RVI}}_{{\mathrm{Freeman}}_{\min }}}}$$

(3)

$${\mathrm{NDSI}}_{0~1}={\displaystyle \frac{\mathrm{NDSI}-{\mathrm{NDSI}}_{\min }}{{\mathrm{NDSI}}_{\max }-{\mathrm{NDSI}}_{\min }}}$$

(4)

Utilizing the normalized values, an integrated optical-radar 3D feature space—designated as the Huynen-RVI-NDSI 3D feature space—was established. In this space, Huynen_vol_0~1 (ranging from 0 to 1) serves as the X-axis, RVI_Freeman_0~1 (0 to 1) as the Y-axis, and NDSI_0~1 (0 to 1) as the Z-axis.

As illustrated in Figure 7, the points, randomly distributed in the Huynen-RVI-NDSI space, show pronounced interrelationships, indicating strong synergy among the selected features. This finding substantiates the advantages of multi-source RS for synergistic inversion. Specifically, the Normalized Difference Salinity Index (NDSI), as an optical indicator, is highly sensitive to surface soil salinity variations due to differences in reflectance between the red (620–750 nm) and near-infrared (840–870 nm) bands. RVI_Freeman, derived from fully polarimetric SAR data, is a radar vegetation index that indirectly reflects vegetation degradation under salinity stress by capturing canopy geometric structure. The volumetric scattering component (Huynen_vol), obtained from Huynen polarimetric decomposition, characterizes mixed vegetation scattering and is closely linked to canopy density, indirectly reflecting ground vegetation cover. Overall, scatter points in this 3D feature space radiate in specific directions, and the projection density gradient closely aligns with the gradation of soil salinization. The strong synergy further confirms the advantages of joint inversion using multi-source RS: optical data directly reveal surface soil salinity, while radar data indirectly indicate soil conditions through vegetation coverage.

Analysis of the projection on 2D planes (Figure 8) reveals a negative correlation between RVI_Freeman and NDSI, as well as between Huynen_vol and NDSI, and a positive correlation between Huynen_vol and RVI_Freeman. By further decomposing the 3D feature space into three 2D projection planes, the synergistic mechanisms among the multi-source RS parameters and the distribution characteristics of different degrees of salinization are elucidated (Figure 9).

Drawing upon high-resolution auxiliary imagery and a comprehensive archive of field photographs, the interpretative characteristics of the three projection planes within the 3D feature space are as follows: (1) On the RVI-NDSI plane, an increase in RVI_Freeman is accompanied by a decrease in NDSI, illustrating the negative feedback between salinity and vegetation; that is, weaker spectral salinity signals coincide with stronger radar vegetation scattering. (2) On the Huynen-NDSI plane, the scatter points exhibit a distinctive radial pattern—dense clustering in regions of low salinity and diagonal dispersion in higher salinity areas—indicating a negative correlation between radar-derived volumetric scattering and surface salinity. (3) On the Huynen-RVI plane, the distribution reveals a positive correlation between Huynen_vol and RVI_Freeman, with lower values corresponding to transitional desert zones and higher values indicative of vegetated areas with varying salinity status.

4.2. Optical-Radar Soil Salinity Inversion Model

Field observations and empirical studies on salinization demonstrate that increasing soil salinity induces significant changes in surface conditions: areas with low salinity are generally well-vegetated; as salinity increases, vegetation cover diminishes; and under severe salinization, even halophytes become scarce or absent. Therefore, vegetation cover is highly responsive to changes in soil salinity, and both vegetation coverage and volumetric scattering components exhibit a marked decline with increasing salinity. This relationship provides a basis for using the position of different salinization levels within the feature space as an index for soil salinity inversion.

In this study, the position of various salinization degrees within the Huynen-RVI-NDSI optical-radar 3D feature space is used to quantitatively represent soil salinity. Specifically, the Euclidean distance from any pixel to an ideal reference point—defined by maximum values of Huynen_vol and RVI_Freeman and the minimum NDSI, corresponding to optimal vegetation cover and minimal salinity—serves as an indicator of salinity severity: the greater the distance, the more severe the salinization (see Figure 10).

Accordingly, the Optical-Radar Salinity Inversion Model (ORSIM) is proposed. The ideal low-salinity reference point is set at (1, 1, 0) in the feature space, and the ORSIM expresses the soil salinity of any point P (x, y, z) as the Euclidean distance to this reference point O (1, 1, 0), calculated as follows:

$$\mathrm{ORSIM}=\mathrm{P}\stackrel{\mathrm{L}}{\to }\mathrm{O}=\sqrt{{{\left(1-{\mathrm{Huynen}}_{\mathrm{vol}}\right)}^2+\left(1-{\mathrm{RVI}}_{\mathrm{Freeman}}\right)}^2+{\left(\mathrm{NDSI}\right)}^2}$$

(5)

4.3. Soil Salinity Inversion

To validate the ORSIM, we established a linear fitting relationship between in situ measured soil electrical conductivity (EC) and corresponding ORSIM values from 69 field samples (Figure 11). The fitted linear function (y = 33.539x − 12.788, where the x-axis represents ORSIM values and the y-axis represents measured EC) shows a strong correspondence (R² = 0.75), with an average deviation of 7.57 dS/m (RMSE). This demonstrates that: (1) the ORSIM effectively reflects soil salinity gradients, and (2) the optical-radar feature space approach captures meaningful spatial patterns of soil salinization. The results confirm the ORSIM’s utility for regional-scale salinity monitoring applications.

Additionally, the ORSIM was applied to generate a spatial inversion map of soil salinity across the entire study region. As shown in Figure 12, the spatial distribution of soil salinity is delineated by red-hued areas indicate zones of high salinity and elevated salinization risk, while dark green regions represent areas of lower soil salt content and minimal salinization. Mildly salinized areas are concentrated in the oasis center and its periphery, where moisture is sufficient and vegetation is dense. In contrast, more severely salinized zones are primarily located along the northern, low-lying margins, where high evaporation and poor drainage promote salt accumulation. Overall, soil salinity in the study area displays extensive yet discontinuous spatial patterns, with higher salinity levels in the north and lower levels in the south, reflecting a gradual north-south decline. This distribution closely corresponds to the topographic features of the Yutian Oasis, characterized by higher elevations in the south and low-lying desert in the north. Furthermore, significant spatial heterogeneity exists among different levels of salinization, revealing a complex and intricate landscape of soil salinity dynamics.

5. Discussion

5.1. Soil Salinity Distribution Characteristics Analysis

As depicted in Figure 12, saline soils are predominantly distributed along the oasis margins, the ecotone between oasis and desert, and the northern sector of the study area. This distribution pattern is largely attributable to the increased input of soluble salts from upstream of the Keriya River [56] in the northern Yutian Oasis, and aligns well with previous studies based solely on optical RS for large-scale salinity inversion [84]. Topography is also a decisive factor: the southern region’s higher elevation and the northern area’s lower elevation lead to northern runoff convergence and the formation of low-lying, poorly drained zones, where substantial evaporation accelerates surface salt accumulation [39]. Salinity generally increases moving from farmland toward the desert–oasis ecotone [10]. In the northern Gobi and desert areas, arid conditions drive high groundwater evaporation, and capillary action transports salts upward, resulting in significant surface salt accumulation [85]. Groundwater depth plays a critical role: shallow water tables enhance capillary rise and leaching, intensifying salinization in the north, while flat terrain and irrigation runoff retention further contribute to salt buildup. In the south, the main agricultural region, high salinity is mainly observed at the edges of cultivated fields, potentially due to secondary salinization arising from suboptimal farming practices [86]. The ORSIM effectively delineates the patchy distribution of severely, moderately, mildly, and non-salinized soils, capturing isolated saline patches within both agricultural and natural vegetation core areas. These spatial insights provide valuable references for arable land protection and ecological management.

5.2. The Potential and Advantages of the Model

The salinization issues in the Yutian Oasis are the result of a complex interplay among climatic, hydrological, vegetative, edaphic, and anthropogenic factors, establishing the region as a prototypical area affected by soil salinization. In the context of soil salinity inversion using feature space, optical RS data provide abundant surface spectral information, facilitating the rapid and efficient mapping of soil salinity through the strategic selection of feature indices for feature space construction. However, models relying exclusively on optical RS are constrained by atmospheric disturbances, spectral saturation, and mixed pixel effects, which limit their inversion accuracy. To overcome these limitations, this study integrates optical data with SAR observations, exploiting the penetration capabilities of GF-3 SAR to enhance sensitivity to both vegetation and surface characteristics. By incorporating radar scattering mechanism parameters (Huynen_vol), a radar vegetation index (RVI_Freeman), and an optical index (NDSI) into a multiparametric 3D feature space, this approach effectively compensates for the shortcomings of single-source RS data.

Furthermore, most previous research has been confined to 2D feature spaces, with relatively little attention paid to the construction and utility of 3D spaces. In practice, 3D feature spaces enable the inclusion of additional parameters, thereby expanding the potential for in-depth mining and interpretation of RS data. Grounded in feature space theory, this study applies polarimetric decomposition and radar index extraction from PolSAR data to systematically investigate the synergistic inversion capabilities of optical and radar parameters, ultimately achieving efficient soil salinity inversion in arid regions. When compared to the 2D optical-radar ORSDI inversion model proposed by Muhetaer et al. [56] (maximum R² = 0.656), the newly developed 3D optical-radar feature space ORSIM exhibits a significant improvement in accuracy (R² = 0.75), thus demonstrating the effectiveness and superiority of the 3D feature space approach.

5.3. Limitations and Future Work

The 3D feature space soil salinity inversion model developed in this study—based on the synergistic integration of radar and optical RS—demonstrates both feasibility and application potential for monitoring soil salinization in arid environments, combining high efficiency with clear model interpretability. Nonetheless, several limitations remain.

First, although SAR speckle filtering was employed to suppress noise, complete elimination remains challenging, and filtering may inadvertently attenuate high-frequency signals, resulting in the loss of fine surface features. Future studies should systematically evaluate various filtering algorithms—including non-local means and deep learning-based approaches—focusing on noise reduction and feature preservation, and optimize performance through parameter sensitivity analysis and validation across different scenarios.

Second, radar backscatter intensity is highly sensitive to soil dielectric properties (strongly linked to soil moisture) and surface roughness, which may introduce confounding effects in quantitative salinity prediction. Additionally, residual speckle noise can further degrade data quality. Future research should exploit the complementary strengths of multi-source RS by integrating polarimetric decomposition, temporal feature enhancement, and coupled physical modeling to optimize model performance. Such strategies will help mitigate the effects of soil moisture, vegetation cover, surface roughness, and speckle noise on salinity inversion accuracy, thereby improving monitoring precision and reliability.

Additionally, to achieve spatial consistency between optical and radar data, this study utilized cubic convolution interpolation to resample radar images to a 10 m resolution. While this ensured data compatibility, it also resulted in the loss of some fine micro-textural information inherent to higher-resolution radar images. Future efforts should investigate super-resolution reconstruction, multi-scale feature fusion, or spatiotemporal collaborative fusion algorithms to further enhance the spatial and informational quality of RS imagery.

Finally, although field sampling was designed to capture typical salinization categories and surface heterogeneity, the overall sample size remains limited, potentially introducing sampling bias and affecting the generalizability of the model. Future field campaigns should aim to increase sampling density and coverage to obtain more comprehensive and granular data regarding the spatial variability of soil salinity.

6. Conclusions

This study developed the Optical-Radar Salinity Inversion Model (ORSIM), a novel 3D feature space framework for soil salinity estimation in arid regions. By synergistically integrating polarimetric GF-3 SAR data and Sentinel-2 multispectral imagery, we constructed a Huynen-RVI-NDSI feature space through random forest-based feature selection, enabling quantitative salinity mapping across the Yutian Oasis. The main conclusions are as follows:

(1) Random forest (RF) analysis identified three optimal indices with high salinity sensitivity: (i) Huynen_vol (volume scattering component from Huynen decomposition); (ii) RVI-Freeman (Freeman decomposition-based Radar Vegetation Index); and (iii) NDSI (Normalized Difference Salinity Index). These indices formed the foundation for ORSIM’s 3D feature space.

(2) Linear fitting between ORSIM values and in situ soil electrical conductivity (EC) demonstrated strong agreement (R² = 0.75, RMSE = 7.57 dS/m), confirming the model’s reliability for quantifying salinity gradients across low-to-high salinity conditions.

(3) The inversion results revealed three salient spatial characteristics: (i) A topography-driven north-south gradient, with severe salinity in northern low-lying deserts and lower salinity in southern highlands; (ii) ecotone vulnerability, with critical salinity zones concentrated in the fragile oasis–desert transition areas; and (iii) high spatial heterogeneity, reflecting complex interactions between evaporation, groundwater dynamics, and landscape position.

In summary, this study confirms that the 3D feature space model, based on the synergistic integration of optical and radar RS data, enables rapid and effective inversion of regional surface soil salinity, thereby providing robust multisource and multi-dimensional support for dynamic, high-precision, and real-time monitoring of soil salinization in arid regions.