Multi-Source Attention U-Net: A Novel Deep Learning Framework for the Land Use and Soil Salinization Classification of Keriya Oasis in China with RADARSAT-2 and Landsat-8 Data

Abstract

Soil salinization significantly impacts global agricultural productivity, contributing to desertification and land degradation; thus, rapid regional monitoring of soil salinization is crucial for agricultural production and sustainable management. With advancements in artificial intelligence, the efficiency and precision of deep learning classification models applied to remote sensing imagery have been demonstrated. Given the limited feature learning capability of traditional machine learning, this study introduces an innovative deep fusion U-Net model called MSA-U-Net (Multi-Source Attention U-Net) incorporating a Convolutional Block Attention Module (CBAM) within the skip connections to improve feature extraction and fusion. A salinized soil classification dataset was developed by combining spectral indices obtained from Landsat-8 Operational Land Imager (OLI) data and polarimetric scattering features extracted from RADARSAT-2 data using polarization target decomposition. To select optimal features, the Boruta algorithm was employed to rank features, selecting the top eight features to construct a multispectral (MS) dataset, a synthetic aperture radar (SAR) dataset, and an MS + SAR dataset. Furthermore, Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), and deep learning methods including U-Net and MSA-U-Net were employed to identify the different degrees of salinized soil. The results indicated that the MS + SAR dataset outperformed the MS dataset, with the inclusion of the SAR band resulting in an Overall Accuracy (OA) increase of 1.94–7.77%. Moreover, the MS + SAR MSA-U-Net, in comparison to traditional machine learning methods and the baseline model, improved the OA and Kappa coefficient by 8.24% to 12.55% and 0.08 to 0.15, respectively. The results demonstrate that the MSA-U-Net outperformed traditional models, indicating the potential of integrating multi-source data with deep learning techniques for monitoring soil salinity.

1. Introduction

Soil salinization is considered an environmental and socioeconomic challenge, profoundly impacting agricultural development and ecosystem health [1]. Salinity issues can arise in arid areas and may be the result of natural processes or human activities [2]. In general, the persistent lack of precipitation and improper irrigation lead to the migration of accumulated salts from saline geological deposits to the surface, leading to excessive accumulation of salts [3]. Over the past two decades, soil salinization has been a significant ecological challenge, mostly caused by the growing influence of climate change [4]. The global distribution of saline soils covers a significant land area, totaling approximately 932.2 million hectares. More than 100 countries are impacted by soil salinization [5,6]. In China, saline soils make up about 5% of the country’s arable land [4], with Xinjiang being recognized as one of the largest global regions affected by widespread salinization [7]. Therefore, an exact and real-time assessment of soil salinization extent is vital for the effective implementation of management strategies and the sustainable use of land resources.

Traditional methods for assessing regional soil salinization, primarily through field-based sampling and laboratory analysis, are time-consuming, labor-intensive, and limited in their capacity to offer large-scale, real-time dynamic monitoring [8]. In contrast, remote sensing technology provides an affordable and efficient approach to soil salinity assessment, providing rapid, timely, and relatively inexpensive data acquisition [9,10]. It has, thus, been largely adopted for soil salinity mapping and monitoring. Nevertheless, directly and accurately estimating soil salt content from optical imagery remains challenging due to constraints in spatial and spectral resolutions, as well as interference from vegetation cover and atmospheric effects [11]. In this context, SAR has significant advantages due to its strong penetrative capabilities and ability to capture data under all-weather, all-day conditions [12,13]. SAR technology holds considerable promise for soil salinity detection, as radar systems are highly sensitive to parameters such as electrical conductivity (EC), backscattering coefficient, soil moisture, and salinity [14,15].

With the advancement of SAR technologies, polarimetric SAR (PolSAR) has emerged as a high-resolution imaging system that provides comprehensive polarization scattering information about terrains, surpassing the capabilities of single-polarization SAR systems [16]. The full-polarization scattering information derived from PolSAR data is associated with the physical properties of surface objects, which makes a promising technology for accessing soil salinity [17]. For example, Zhang et al. retrieved and estimated soil salt content from dual-polarized PolSAR images using a one-dimensional convolutional neural network [18]. In addition, polarimetric target decomposition has found extensive application across multiple areas, such as monitoring geological disasters [19], assessing forests [20], estimating soil moisture and salinity [21], and performing land cover classification [17]. Overall, polarization decomposition offers a novel perspective for retrieving ground information, offering significant insights for the management and sustainable use of soils.

With the advent of the age of artificial intelligence, deep learning (DL) techniques have demonstrated considerable promise in extracting features from remote sensing images [22]. Numerous studies have demonstrated that, in contrast to traditional machine learning, the distinctive multi-layered architecture of neural networks enables the extraction of multi-scale and deep-level feature information, thereby allowing it to achieve a superior classification accuracy and effect [23,24]. DL approaches have gained considerable attention from researchers due to their superior classification performance on satellite imagery, particularly with complex data types such as PolSAR [25]. Fu et al. proposed an optimized fully convolutional network (FCN) model that utilizes a dilated convolution and multi-scale network architecture to achieve precise classification of high-resolution images [26]. Rajat Garg et al. utilized the DeepLab V3+ model and its transfer learning capabilities to achieve superior performance in land cover classification using PolSAR data, even with small datasets [24]. Despite the advancements made by FCN in semantic segmentation, they also present several drawbacks, including spatial information loss due to multiple pooling layers; high memory and computational demands; increased complexity in training, particularly with limited datasets; and challenges in accurately segmenting object boundaries [27].

The U-Net model, a symmetric FCN, was originally developed for image segmentation tasks in the field of biomedical imaging [28]. By integrating low-level spatial information from down-sampling with an up-sampling input via skip connections, U-Net significantly outperforms traditional FCNs [29]. In recent years, U-Net-based deep learning approaches have made significant strides in remote sensing data monitoring, largely due to U-Net’s ability to capture subtle spatial features and perform well under small-sample conditions [30]. Although both traditional and deep learning methods have achieved notable success in land cover classification, most studies still predominantly rely on either visible light or SAR data [31,32]. As several studies have pointed out, this reliance on a single data source often limits the consistency and reliability of the classification accuracy [33,34]. Specifically, single-source data may struggle to capture the full complexity of regions with significant variation in land cover characteristics, such as complex terrain and boundary areas [35]. Despite the challenges posed by data preprocessing issues—such as differences in resolution [36], registration [37], noise handling [38] and the difficulties associated with effectively integrating features from different data sources—research focusing on fully using the complementary advantages of multi-source data for more comprehensive feature extraction remains limited [39]. Integrating different data sources has the potential to combine the spectral details with the edge and texture information, thereby improving classification accuracy [40].

Building on its strong performance with small samples, several U-Net-based networks and improved U-Net networks have been optimized by integrating existing modules to enhance efficiency, robustness, and generalization. Attention mechanisms, such as the Convolutional Block Attention Module (CBAM) introduced by Woo et al., have been shown to improve deep learning models by adaptively focusing on relevant features and suppressing noise [41]. The CBAM integrates channel and spatial attention to enhance feature maps, making it particularly effective in improving convolutional neural networks for remote sensing tasks [41]. Specifically, channel attention strengthens the representation of critical features across different channels, allowing the model to focus on meaningful information in multi-source remote sensing data [42]. For instance, radar data channels typically contain pronounced edge and texture information, while spectral data channels capture land cover spectral details [43]. By enhancing key channels, channel attention can reduce interference from less relevant channels in soil salinity monitoring, potentially improving soil salinity monitoring under specific conditions. Spatial attention, on the other hand, adjusts feature responses at different spatial positions, potentially increasing the model’s ability to detect spatial patterns in soil salinity distributions. Therefore, this module can aid in capturing potential transitional distribution patterns of soil salinization.

In order to make full use of deep learning to extract multi-scale features from images by combining the rich texture information with the multispectral information, this paper proposes an innovative deep-fusion U-Net model utilizing multi-source data to map different degrees of salinized soil. The model, named Multi-Source Attention U-Net (MSA-U-Net) in this study, is based on a deep fusion architecture that integrates the CBAM into the skip connection, featuring a simple, lightweight structure capable of leveraging limited RS data for improved accuracy.

2. Materials

2.1. Study Site

The Keriya Oasis (35°14′–39°29′ N, 81°9′–82°51′ E) in southwestern Xinjiang, China, is located between Cele County to the west and Minfeng County to the east [33]. Covering an approximate total area of 39,500 km², the Keriya Oasis exhibits mountainous terrain in its southern region and flat plain desert landscapes to the north [44]. The region has a dry climate, low precipitation, and high evaporation rates, leading to extreme drought conditions [45]. As a typical inland arid desert region with a warm temperate climate, it also experiences abundant solar radiation and heat resources. The annual temperature stands at 11.6 °C, with a notable temperature variation between day and night. Based on the above reasons, water shortages and extremely fragile ecosystems have emerged as serious environmental issues in this region [46]. The location of the study area is shown in Figure 1.

The Keriya River, a seasonal watercourse, flows from south to north, originating in the Kunlun Mountains [47]. The meltwater provides a major water source for oasis agriculture in the midstream areas and natural oases in the downstream. However, due to intense evaporation, dissolved salts have accumulated on the surface, resulting in elevated salt concentrations. This has resulted in significant water shortages, which have contributed to soil salt accumulation, exacerbating ecological degradation in the Keriya Oasis [48]. Therefore, it is essential to develop a comprehensive understanding of the region’s characteristics.

2.2. Data Collection and Preprocessing

The RADARSAT-2 satellite was launched by the Canadian Space Agency (CSA) on 14 December 2007 [49]. The RADARSAT-2 data were collected on 6 May 2022. Additionally, optical data from the Landsat-8 OLI, launched by NASA in 2013 [50], were also used in this study. The main parameters are shown in

Table 1. The main parameters of remote sensing data.

Remote Sensing	RADARSAT-2	Landsat-8
Map Projection	WGS84 (DD)	UTM
Sensor	C-band synthetic aperture radar	Operational Land Imager
Data Observation Date	6 May 2022	18 May 2022
Product Type	SLC	L2SP
Nominal Resolution	5.5 m $\times$ 4.8 m	30 m
Incident angle/Orbit Inclination Angle	42.1°	98.2°
Revisit Time	24 d	16 d
Orbit Type	Sun Synchronous Orbit	Sun Synchronous Orbit
Satellite Attitude	798 km	705 km
Band Number	---	11
Polarizations	HH, HV, VV, VH	---

.

The RADARSAT-2 data were preprocessed using SNAP 9.0 software [51]. The main preprocessing steps are as follows: (1) radiometric calibration; (2) the generation of T3 polarization coherent matrix; (3) multi-view processing; (4) application of a 5 × 5 window refined Lee filtering for polarimetric speckle reduction; (5) geocoding; (6) resampling to the optimal resolution of 10 m.

The Landsat-8 OLI data were preprocessed by ENVI 5.3 and ArcMap 10.8.1 by performing the following operations: (1) radiometric calibration; (2) atmospheric correction; (3) image alignment; (4) resampling to the same resolution of 10 m; (5) spatial subset.

2.3. Field Data

Field sampling for this study took place in May 2022, with 60 sampling sites (Figure 2) distributed relatively evenly throughout the region. These sites represented a range of land use and soil salinization (LUSS) types. Soil type, soil texture, sampling time, number of sampling sites, surrounding vegetation types, and the corresponding GPS coordinate were recorded together at the sampling sites. The researchers securely sealed the soil samples in sealed bags with labels for laboratory analysis. The topsoil (0–20 cm) samples were air-dried, ground, and filtered through 1 mm sieves. Each soil sample was thoroughly mixed with distilled water at a soil-to-solution ratio of 1:5 in a flask and then shaken to achieve homogeneous dispersion. The mixture was allowed to settle for 24 h before the supernatant was carefully filtered to remove suspended particles. The electrical conductivity (${\mathrm{EC}}_{1:5}$) of the filtered solution was then measured using a digital multiparameter measuring apparatus (Multi 3420 Set B, WTW GmbH, Munich, Germany) [33]. According to previous studies, a highly significant linear relationship exists between the ECe and ${\mathrm{EC}}_{1:5}$ values for salinized soils in southern Xinjiang [52], with the ${\mathrm{EC}}_{1:5}$ value converted to the ECe value using an empirical equation, marked as EC [53,54]. Soil electrical conductivity has been found to be positively linked to its soluble salt content. Consequently, the researchers employed regression equations to estimate the total salt content in samples [55,56].

By combining high-definition online satellite images with field photos, the LUSS types were classified into six categories based on our previous laboratory research [57] and the soil salinization classification standards outlined in [58]. These categories, determined by physical and chemical properties, especially electrical conductivity, include vegetation, water body, bare land, slightly salinized soil, moderately salinized soil, and highly salinized soil (

Table 2. Land use and soil salinization classes exhibiting different degrees of soil salinization, their characteristics, and corresponding pictures.

Symbol	Class	Characteristics	Typical Filed Photos
WB	Water Body	Salt lakes, rivers and tributaries, swamps, ponds, and reservoirs.
VG	Vegetation	Grassland, farmland, red willow, populus euphratica, reed, camel thorn, and dried riverbed.
BL	Barren Land	Gobi, desert, wasteland, and bare land.
SS	Slightly Salinized soil	EC value 2–4 (dS/m), covered by a thin salt crust (0–2 cm), and the vegetation coverage is around 30%.
MS	Moderately Salinized soil	EC value of 4–8 (dS/m), with a salt crust thickness of 1 to 4 cm and a vegetation coverage of around 5 to 15%.
HS	Highly Salinized soil	EC value of 8–16 (dS/m), covered by a thin salt crust (4–10 cm), and the vegetation coverage is less than 5%.

).

2.4. Collection of Training and Validation Data

Training data are a critical component of supervised learning, with most machine learning algorithms relying on a large number of samples. However, acquiring reference data from satellite images can be a difficult task [59]. To construct a deep learning training set, 180 sample plots (Figure 3) were created through visual interpretation, combined with field-collected landscape photos and sampling point data from the study area, false-color composites from Landsat-8 OLI, and spectral indices (e.g., vegetation or soil-related indices). These resources were used for visual interpretation and classification of different LUSS classes.

Since the vegetation, bare land, and water bodies of Keriya Oasis are obviously displayed in remote sensing images, some areas of the study area were labeled through visual interpretation. For salinized soil of different degrees, our research team members divided plots along the oasis edge’s highway at the junction of the oasis and desert and sampled the salinized soil, and the sampling points were evenly distributed within each plot. The distribution of some salinized land plots is shown in Figure 2. After laboratory analysis, the EC values were obtained and used as a reference for visual salinization levels at the sample sites.

The training data include 180 plots of 128 × 128 pixels, and the validation data consist of 70 plots of the same size. In order to ensure data consistency, deep learning and machine learning require the same set of training and validation. Due to computational memory limitations, it is not possible to train all machine learning model algorithms using the complete training set. Thus, we established a balanced dataset by randomly selecting sampling points from the training and validation plots, ensuring the same number of pixels in each class [60]. The statistics of training set and verification set are shown in

Table 3. The number of plots for each DL training and validation sample and the number of pixels selected for ML.

Class	Abbreviation	Training (70%)		Validation (30%)
		Plots	Pixels	Plots	Pixels
Vegetation	VG	42	4290	18	1810
Barren Land	BL	20	2877	9	1090
Water Body	WB	18	1576	8	1004
Slightly Salinized soil	SS	35	3542	15	1406
Moderately Salinized soil	MS	32	3195	14	1367
Highly Salinized soil	HS	33	3433	16	1743
Total	-	180	17,913	70	8420

.

3. Methods

This study developed MSA-U-Net, an enhanced architecture integrating the Convolutional Block Attention Module (CBAM) into U-Net’s skip connections to optimize multi-scale feature fusion. We conducted comprehensive comparisons of traditional machine learning methods (SVM, RF, KNN) and deep learning models (U-Net, MSA-U-Net) for soil salinization classification. The entire study’s flowchart is shown in Figure 4.

3.1. Vegetation and Soil-Related Indices

Optical remote sensing obtains surface information by transmitting and receiving electromagnetic wave energy reflected from targets [61]. Thus, the pre-processed Landsat-8 images contain rich spectral information, most of which is presented in the spectral reflectance values of the different visible bands of the images [62]. In salinized soils, soil salinity affects the growth of vegetation and the physicochemical composition of the soil, leading to changes in spectral reflectance at different salinity levels [63,64]. Thus, the spectral reflectance of saline soils or vegetation can indirectly indicate soil salinity [65,66].

In recent years, numerous vegetation and soil salinity indices have been developed for the effective monitoring of vegetation health and soil salinity respectively. These indices are crucial for understanding the relationship between optical imagery and soil characteristics. The spectral indices including vegetation and soil indices (listed in

Table 4. Summary of the selected spectral indices.

Category	Index	Calculation Formula	Reference
Vegetation Indices	Simple Ratio Vegetation Index (SR)	$\mathrm{SR}=\mathrm{NIR}/\mathrm{R}$	[68]
	Normalized Difference Vegetation Index (NDVI)	$\mathrm{NDVI}=(\mathrm{NIR}-\mathrm{R})/(\mathrm{NIR}+\mathrm{R})$	[69]
	Soil Adjusted Vegetation Index (SAVI)	$\mathrm{SAVI}=\left[\left(\mathrm{NIR}-\mathrm{R}\right)\times 1.5\right]/[\left(\mathrm{NIR}+\mathrm{R}+0.5\right)]$	[70]
	Green Normalized Difference Vegetation Index (GNDVI)	$\mathrm{GNDVI}=(\mathrm{NIR}-\mathrm{G})/(\mathrm{NIR}+\mathrm{G})$	[71]
	Differential Vegetation Index (DVI)	$\mathrm{DVI}=\mathrm{NIR}-\mathrm{R}$	[72]
	Normalized Difference Green Index (NDGI)	$\mathrm{NDGI}=(\mathrm{G}-\mathrm{R})/(\mathrm{G}+\mathrm{R})$	[73]
	Enhanced Normalized Vegetation Index (ENDVI)	$\mathrm{ENDVI}=(\mathrm{NIR}+\mathrm{SWIR}2-\mathrm{R})/(\mathrm{NIR}+\mathrm{SWIR}2+\mathrm{R})$	[74]
Soil-related Indices	Salinity Index (SI-T)	$\mathrm{SI}-\mathrm{T}=(\mathrm{R}/\mathrm{NIR})\times 100$	[75]
	Salinity Index (SI1)	$\mathrm{SI}1=\sqrt{\mathrm{R}\times \mathrm{G}}$	[76]
	Salinity Index (SI2)	$\mathrm{SI}2=\sqrt{{\mathrm{G}}^2\times {\mathrm{R}}^2\times {\mathrm{NIR}}^2}$	[77]
	Salinity Index (SI3)	$\mathrm{SI}3=\sqrt{{\mathrm{G}}^2\times {\mathrm{R}}^2}$	[77]
	Salinity Index (SI4)	$\mathrm{SI}4=(\mathrm{R}\times \mathrm{B})/\mathrm{G}$	[78]
	Salinity Index (S1)	$\mathrm{S}1=\mathrm{B}/\mathrm{R}$	[77]
	Salinity Index (S2)	$\mathrm{S}2=(\mathrm{B}-\mathrm{R})/(\mathrm{B}+\mathrm{R})$	[78]
	Salinity Index (S3)	$\mathrm{S}3=(\mathrm{R}\times \mathrm{G})/\mathrm{B}$	[78]
	Salinity Ratio Index (SAIO)	$\mathrm{SAIO}=(\mathrm{R}-\mathrm{NIR})/(\mathrm{G}+\mathrm{NIR})$	[79]
	Brightness Index (BRI)	$\mathrm{BRI}=\sqrt{{\mathrm{G}}^2+{\mathrm{R}}^2}$	[79]

) were selected for feature extraction based on their relevance to the salinization process and their proven effectiveness in previous studies [9,67].

3.2. Polarimetric Decomposition

The polarization target decomposition theory, first proposed in 1970 [80], laid the foundation for understanding scattering mechanisms in synthetic aperture radar (SAR) images. Numerous polarization decomposition methods have been developed and widely applied to analyze and process fully polarized SAR data [81,82,83]. These methods decompose each pixel of a SAR image into multiple scattering components based on known scattering mechanisms. To describe and analyze the different types of scattering mechanisms, we derive the polarization covariance and coherence matrix of the target. Polarization target decomposition methods are typically divided into two categories: coherent decomposition, which is based on the target scattering matrix, and incoherent decomposition, which utilizes the polarization coherence matrix, polarization covariance matrix, Mueller matrix, or Stokes vector [84].

In this study, the scattering matrix S, correlation matrix T, and covariance matrix C were derived from the RADARSAT-2 data. To fully utilize the PolSAR data, several polarization target decomposition methods were employed, including Pauli decomposition for coherent polarization target decomposition, Freeman–Durden decomposition [85], Yamaguchi decomposition [86], and Cloude decomposition [87] for incoherent polarization target decomposition, and Sinclair decomposition for common target polarization decomposition. Additionally, Van Zyl decomposition [88], H/A/Alpha decomposition [89], and Touzi decomposition [90] were also considered as part of the eight polarization target decomposition methods used in this study.

The scattering matrix S of the target is expressed as follows:

$$\mathit{S}=\left[\begin{array}{cc}〈S_{hh}〉& 〈S_{hv}〉\\ 〈S_{vh}〉& 〈S_{vv}〉\end{array}\right].$$

(1)

The polarization covariance matrix C of the target is expressed as follows:

$$\mathit{C}={\displaystyle \frac{1}{2}}\left[\begin{array}{ccc}〈S_{hh}·{S_{hh}}^{*}〉& \sqrt{2}〈S_{hh}·{S_{hv}}^{*}〉& 〈S_{hh}·{S_{vv}}^{*}〉\\ \sqrt{2}〈S_{hv}·{S_{hh}}^{*}〉& 2〈S_{hv}·{S_{hv}}^{*}〉& \sqrt{2}〈S_{hv}·{S_{vv}}^{*}〉\\ 〈S_{vv}·{S_{hh}}^{*}〉& \sqrt{2}〈S_{vv}·{S_{hv}}^{*}〉& 〈S_{vv}·{S_{vv}}^{*}〉\end{array}\right].$$

(2)

The polarization coherence matrix T of the target is expressed as follows:

$$\mathit{T}={\displaystyle \frac{1}{2}}\left[\begin{array}{ccc}〈{(S_{hh}+S_{vv})·(S}_{hh}+{S_{vv})}^{*}〉& 〈{(S_{hh}+S_{vv})·(S}_{hh}-{S_{vv})}^{*}〉& 2〈\left(S_{hh}+S_{vv}\right)·{S_{hv}}^{*}〉\\ 〈{(S_{hh}-S_{vv})·(S}_{hh}+{S_{vv})}^{*}〉& 〈\left(S_{hh}-S_{vv}\right)·{(S}_{hh}-{S_{vv})}^{*}〉& 2〈\left(S_{hh}-S_{vv}\right)·{S_{hv}}^{*}〉\\ 2〈S_{hv}{·(S}_{hh}+{S_{vv})}^{*}〉& 2〈S_{hv}{·(S}_{hh}-{S_{vv})}^{*}〉& 4〈S_{hv}·{S_{hv}}^{*}〉\end{array}\right]$$

(3)

where the symbol * denotes the conjugate.

3.3. Optimal Features Selection

Feature selection is a critical step in the feature construction process, involving the selection of the most important subset of variables for prediction or classification tasks. The Boruta algorithm is an efficient feature selection optimization technique that introduces shadow features and trains the features along with the original features in a unified feature matrix to evaluate the importance of the original features [91]. Specifically, the algorithm generates shadow features randomly and compares them with the candidate features to identify those in the dataset that are relevant to the dependent variable, ranking them according to their relative importance [92].

3.4. Supervised Classification

3.4.1. Machine Learning Algorithms

K-Nearest Neighbor (KNN)

KNN is a classic classification algorithm first proposed in 1967 [93]. The core principle of the KNN algorithm involves evaluating the features of test data and comparing them with those in the training set, subsequently identifying the $K$ nearest samples to perform classification or prediction tasks. The similarity between samples can be measured using various distance metrics [94,95,96]. KNN has been commonly applied to soil salinity monitoring due to its simplicity, ease of understanding, suitability for multiple classification tasks, and broad applicability [97,98].

Support Vector Machine (SVM)

The Support Vector Machine (SVM) algorithm, first introduced by Vapnik et al., is among the most commonly used kernel-based learning algorithms [99]. In remote sensing classification, the data sample to be classified is typically a single pixel extracted from multispectral or hyperspectral images. The corresponding set of band measurements for each pixel forms a feature vector. The core principle of SVM involves the identification of a hyperplane with the largest margin in a transformed feature space, derived from the feature vectors, with the objective of effectively partitioning the dataset into specified categories [100]. Furthermore, SVM is especially beneficial in remote sensing applications for its capability to manage limited training datasets, often resulting in a higher classification accuracy in comparison with traditional methods [101,102].

Random Forest (RF)

The Random Forest (RF) algorithm, proposed by Leo Breiman et al., is a widely utilized technique in machine learning [103]. RF has been extensively applied in soil salinity monitoring and information extraction [104,105,106]. This method employs ensemble learning, incorporating decision trees, with multiple trees trained on bootstrapped data samples [107]. The classification of each pixel is determined by the majority vote of all trees. In recent studies, RF has gained popularity for classification and regression tasks due to its high-speed training and high classification accuracy [108].

3.4.2. Deep Learning Algorithms

U-Net

The U-Net, introduced by Olaf Ronneberger et al. in 2015, is a convolutional neural network architecture [28,109]. The backbone of U-Net comprises two symmetrical parts: the encoder and the decoder. The encoder processes the input image through pooling operations and extracts features by stacking multiple convolutional layers. Pooling reduces the spatial dimensions by down-sampling while capturing increasingly abstract features. The decoder generates an up-sampled feature map (using operations such as transposed convolution) by combining features from different stages of the encoder. The final layer of the decoder is a convolutional layer that acts as a classifier, generating a probability matrix as output. The output is then thresholded, or the maximum probability value is selected, with each pixel being assigned to its corresponding category.

Multi-Source Attention U-Net (MSA-U-Net)

Based on the U-Net framework, this study introduced and modified the architecture to propose the Multi-Source Attention U-Net (MSA-U-Net). This innovative design enhances the original U-Net by incorporating the CBAM into the skip connections, thereby facilitating more effective feature fusion across different layers. A detailed overview of the MSA-U-Net is presented in Figure 5.

Each layer of the encoding section consists of a series of operations designed to extract and refine features from the input. It includes two main stages, each comprising a convolutional layer, followed by a rectified linear unit (ReLU) activation function [110], and a batch normalization layer. The ReLU activation function introduces non-linearity, and the batch normalization layer normalizes the activations to stabilize and accelerate the training process. The initial convolutional layer converts the input feature map from input to output channels, utilizing a 3 × 3 kernel with a padding of 1 to preserve the spatial dimension. The second stage repeats these operations, further refining the feature map while keeping the output channel number consistent. After the initial extraction of features, the network performs an early fusion of the optical and radar features through concatenation, followed by further encoding to integrate the multi-modal information deeply.

In the decoder, each stage begins with a 2 × 2 transposed convolution operation for up-sampling the feature map, followed by halving the number of feature channels. Subsequently, the up-sampled feature mapping is further refined using a sequence of two 3 × 3 convolutional layers. This up-sampling and convolutional refinement process is repeated four times, halving the number of filters at each stage. A 1 × 1 convolution operation at the end of MSA-U-Net is used to reduce the feature maps to make them equal to the number of classes in the dataset.

The attention mechanism is a process that focuses on the critical parts of information, effectively filtering out secondary or irrelevant details. In deep learning, the attention mechanism enhances the model’s ability to concentrate on key information by dynamically assigning weights to the input data, thereby enhancing processing efficiency and accuracy. Therefore, the MSA-U-Net introduces the Convolutional Block Attention Module (CBAM) (Figure 6) at various levels of the decoder pathway. The CBAM consists of two components: the channel attention mechanism (CAM) and the spatial attention mechanism (SAM).

Radar data capture the surface texture and structure using backscatter and polarization, reflecting roughness, moisture, and material type. Due to their high sensitivity to surface material composition and physical properties, spectral data can distinguish features such as vegetation, soil, and water by reflectance at different wavelengths. The channel attention mechanism assigns weights to important feature channels, prioritizing key information in both radar and spectral data (e.g., vegetation indices or radar polarization channels). This helps the model automatically select the most relevant features, extracting the texture and structure from the radar data, and material composition from the spectral data. When integrating both data types, channel attention balances their differences, enhancing detailed information. The spatial attention mechanism captures local spatial information, focusing on key regions within images, improving the model’s ability to differentiate surface areas and boost performance in multi-source data fusion.

In summary, the proposed MSA-U-Net model not only leverages the complementary strengths of optical and radar data but also enhances segmentation performance through the strategic incorporation of attention mechanisms. It is designed to be versatile and can be used for various tasks requiring high precision in multi-source image analysis, including remote sensing image segmentation.

In addition to the main classification models, we designed two U-Net variants, U-Net-SAM and U-Net-CAM, to evaluate the impact of different attention mechanisms on skip connections. These variants are discussed in the Discussion section.

The proposed models were developed in Python 3.11.8 using a Pytorch backend. All the experiments were conducted on an NVIDIA GeForce RTX 4060 GPU (NVIDIA, Santa Clara, CA, USA). The initial learning rate was assigned as 0.0001. The models were trained with mini-batches of size 16, and early stopping was applied during the training process [111].

3.5. Model Evaluation Metrics

This study employed several evaluation metrics to measure the performance of different algorithms on the validation dataset. The specific evaluation metrics include User’s Accuracy (UA), Producer’s Accuracy (PA), Overall Accuracy (OA), Kappa coefficient, and F1 score [112]. The formulas are as follows:

$$\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}={\displaystyle \frac{{\mathrm{C}}_{\mathrm{ii}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{C}}_{\mathrm{ij}}}}=\mathrm{Producer}’\mathrm{s}\mathrm{Accuracy},$$

(4)

$$\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}={\displaystyle \frac{{\mathrm{C}}_{\mathrm{ii}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{C}}_{\mathrm{ji}}}}=\mathrm{User}’\mathrm{s}\mathrm{Accuracy},$$

(5)

$$\mathrm{OA}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{FP}+\mathrm{TN}+\mathrm{FN}}},$$

(6)

$$\mathrm{K}\mathrm{appa}={\displaystyle \frac{{\mathrm{P}}_0-{\mathrm{P}}_{\mathrm{e}}}{1-{\mathrm{P}}_{\mathrm{e}}}}{;\mathrm{P}}_0={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}};{\mathrm{P}}_{\mathrm{e}}={\displaystyle \frac{\left(\mathrm{TP}+\mathrm{FP}\right)\times \left(\mathrm{TP}+\mathrm{FN}\right)+(\mathrm{TN}+\mathrm{FP})\times (\mathrm{TN}+\mathrm{FN})}{{(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN})}^2}},$$

(7)

$${F1}_{score}=2\times {\displaystyle \frac{\mathrm{UA}\times \mathrm{PA}}{\mathrm{UA}+\mathrm{PA}}}={\displaystyle \frac{2\times \mathrm{TP}}{\mathrm{FP}+2\times \mathrm{TP}+\mathrm{FN}}}$$

(8)

where $n$ is the number of categories; $C_{ij}$ refers to the pixel of actual class $i$ and correctly classified as $j$; TP, FP, TN, and FN represent the total number of true positive, false positive, true negative, and false negative pixels, respectively.

4. Results

4.1. Polarimetric Decomposition of RADARSAT-2 Data

In this study, polarization target decomposition was performed on RADARSAT-2’s fully polarized images using SNAP software to extract feature parameters. First, the scattering matrix S, covariance matrix C, and coherence matrix T were calculated from the RADARSAT-2 data. Next, various polarization decomposition methods were applied to extract polarimetric scattering information. A total of 26 polarization scattering features were extracted and are shown in

Table 5. Polarimetric scattering features derived from RADARSAT-2 data.

Polarization Decomposition	Number of Features	Polarimetric Scattering Features
Pauli	3	Pauli_a/Pauli_b/Pauli_c
Cloude	3	Cloude_Dbl/Cloude_Odd/Cloude_Vol
H/A/Alpha	3	Alpha/Anisotropy/Entropy
VanZyl	3	VanZyl_Vol/VanZyl_Odd/VanZyl_Dbl
Freeman-Durden	3	Freeman_Durden_Dbl/Freeman_Durden_Odd/Freeman_Durden_Vol
Sinclair	3	Sinclair_Dbl/Sinclair_Vol/Sinclair_Surf
Touzi	4	Alpha/Psi*/Phi*/Tau*
Yamaguchi	4	Yamaguchi_Dbl/Yamaguchi_Odd/Yamaguchi_Vol/Yamaguchi_Hlx

Note: * indicates a feature that was excluded.

, and the standard RGB composition images are presented in Figure 7. By visually inspecting each polarimetric feature image, we identified and excluded those with significant noise and a high number of corrupted pixels, ultimately retaining 23 polarization features.

4.2. Feature Preprocessing and Selection of Optimal Feature Subset

The feature space in this study consists of 40 features, including 7 vegetation indices, 10 salinity indices, and 23 polarimetric scattering features. These polarimetric decomposed features and spectral indices were normalized to a range due to their significant differences. Before inputting training data into the machine learning models for training, each feature F was normalized to a scale of [−1, 1]. The normalized formula is expressed below:

$${\mathrm{F}}_{\mathrm{scaled}}={\displaystyle \frac{\mathrm{F}-{\mathrm{F}}_{\min }}{{\mathrm{F}}_{\max }-{\mathrm{F}}_{\min }}}$$

(9)

where $\mathrm{F}$ represents the pixel value.

A total of 40 variables were ranked by Z-Score values using the Boruta algorithm, as shown in Figure 8. Among these, eight variables were identified as having significance levels exceeding the shadow features’ maximum significance (significance of 2.402). These variables including SI2, ENDVI, Sinclair_Dbl, NDGI, Cloude_Vol, SI, Cloude_Dbl, and Freeman_Durden_Odd were subsequently selected for the classification model. The MS dataset comprised four optical features (e.g., SI2, ENDVI, etc.). The SAR dataset consisted of four radar-derived polarimetric scattering features (e.g., Sinclair_Dbl, Cloude_Dbl, etc.). To balance the amount of information from the different datasets, a combined MS + SAR dataset was created by integrating the top two features from each group. The selected features strike a balance between data diversity and redundancy minimization, providing a robust foundation for accurate soil salinization classification.

4.3. Classification Results of MSA-U-Net

To quantitatively evaluate the performance of MSA-U-Net compared to other methods for monitoring salinized soil, the classification accuracies of the five classifiers were analyzed. This evaluation includes a detailed statistical summary and the corresponding confusion matrix, offering a comprehensive performance assessment.

As demonstrated in Figure 9, the classification results of the MS data (Figure 9(A-1–A-5)) display pixel misclassification, with mixed distributions observed among vegetation, water bodies, and saline soil. On average, the Overall Accuracy (OA) across the five classification methods for MS data is 77.11% (

Table 6. Performance comparison of KNN, SVM, RF, U-Net, and MSA-U-Net models with respect to PA, UA, OA, and Kappa across classes (WB, VG, BL, SS, MS, HS).

	KNN		SVM		RF		U-Net		MSA-U-Net
	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)
WB	90.32	88.89	87.90	91.60	90.32	87.50	87.10	85.71	88.33	85.48
VG	83.45	81.12	82.01	87.69	84.17	82.98	79.86	82.84	82.14	82.14
BL	72.43	79.76	97.30	63.83	72.43	81.71	97.30	75.63	92.18	89.19
SS	66.67	62.18	51.11	93.88	70.37	75.25	72.22	92.20	86.96	88.89
MS	61.04	61.04	57.14	65.67	79.22	63.21	68.18	82.68	81.99	85.71
HS	80.38	81.41	83.03	78.61	85.44	91.22	92.41	83.91	91.67	90.51
OA (%)	74.79		77.66		79.10		82.98		87.34
Kappa	0.69		0.73		0.74		0.76		0.84

). Similarly, the classification results of the SAR data (Figure 9(B-1–B-5)) exhibit a lower classification accuracy, with evident effects of noise and outliers on the results, leading to an average OA of 53.23%. In contrast, the multi-source classification results (Figure 9(C-1–C-5)) show better outcomes, with reduced noise in the radar bands and fewer mixed land cover pixels, achieving the highest average OA of 80.37%.

The results in Table 6 show that the proposed MSA-U-Net yielded the highest OA (87.34%) and Kappa (0.84), outperforming traditional classifiers such as KNN (OA: 74.79%, Kappa: 0.69), SVM (OA: 77.66%, Kappa: 0.73), RF (OA: 79.10%, Kappa: 0.74), and the standard U-Net (OA: 82.98%, Kappa: 0.76). Notably, the MSA-U-Net also demonstrated superior PA and UA across most salinization classes.

4.4. Comparison and Analysis of Classification Results Based on Multi-Sources Data

To assess the performance of five methods (KNN, SVM, RF, U-Net, and MSA-U-Net) on these datasets and evaluate the impact of data integration, OA, Kappa, and F1 score were used.

Table 7. Evaluation index values of different methods under different data sources.

Data Source	Classification Methods	OA (%)	Kappa	F1 Score
Landsat-8 OLI (MS)	KNN	76.38	0.71	0.77
	SVM	73.08	0.67	0.72
	RF	77.36	0.75	0.80
	U-Net	79.15	0.74	0.79
	MSA-U-Net	79.57	0.75	0.80
RADARSAT-2 (SAR)	KNN	51.38	0.41	0.51
	SVM	52.98	0.43	0.51
	RF	55.64	0.46	0.56
	U-Net	50.85	0.40	0.50
	MSA-U-Net	51.49	0.41	0.51
Landsat-8 + RADARSAT-2 (MS + SAR)	KNN	74.79	0.69	0.75
	SVM	77.66	0.73	0.77
	RF	79.10	0.75	0.80
	U-Net	82.98	0.79	0.82
	MSA-U-Net	87.34	0.84	0.87

presents the classification performance of all tested models. On the MS dataset, traditional machine learning methods (KNN, SVM, RF) achieved moderate performance, with OA ranging from 73% to 77%. However, the deep learning methods (U-Net and MSA-U-Net) outperformed them, achieving an OA of 79.15% and 79.57%, respectively.

For the SAR dataset, there was a noticeable decrease in the extraction accuracy (OA, Kappa, and F1 score) compared to the MS dataset. Specifically, SAR classification results for deep learning methods showed a more significant decrease in accuracy compared to machine learning methods, with OA dropping by 5% (23% vs. 28%), Kappa by 0.04 (0.28 vs. 0.32), and F1 score by 0.04 (0.25 vs. 0.29).

When combining MS and SAR data, all models showed improved performance, with MSA-U-Net achieving the best results. OA increased significantly, rising from 79.57% to 87.34%. Additionally, Kappa improved by 0.09 (0.75 vs. 0.84), and the F1 score increased by 0.07 (0.80 vs. 0.87).

4.5. Characteristics of Spatial Distribution of Soil Salinity in Keriya Oasis

The classification results indicated a distinct spatial distribution of salinized soils in the Keriya Oasis. Highly salinized soils (HSs) are predominantly concentrated in the central and northeastern parts of the region, corresponding to areas with limited vegetation cover. These areas are typically characterized by bare soil or sparse vegetation, where the accumulation of salts is facilitated by higher evaporation rates and shallower groundwater tables. Moderately salinized soils (MSs) are mainly distributed in the ecotone between the Oasis and the Taklimakan Desert. Slightly salinized soils (SSs) are primarily located in the western and southern regions, often overlapping with vegetated zones, as the presence of vegetation is indicative of lower surface salt accumulation.

5. Discussion

5.1. Classification Performance Across Different Remote Sensing Data Sources

As demonstrated in Figure 9 and Table 6 and Table 7, the MS + SAR data significantly enhanced the classification performance compared to the single-source data (MS and SAR). The MS data, which capture spectral indices sensitive to vegetation characteristics, proved highly effective in identifying vegetation-dominated areas and low-salinity soils [113]. However, their limited sensitivity to soil structure posed challenges for accurately classifying saline soils. In contrast, the SAR data excelled in capturing soil texture and structure, enabling more effective differentiation of salinity levels [114]. Nevertheless, the radar data performed poorly in distinguishing vegetation types, highlighting the limitations of radar’s sensitivity to spectral characteristics and the negative impact of high-frequency noise, such as speckle, on classification results [115]. The combination of optical and radar data improved classification accuracy by utilizing the complementary strengths of both sources [92]. Multi-source data fusion significantly enhanced OA, Kappa, and F1 scores across all classification models, particularly for challenging categories like moderately and highly salinized soils. These results highlight the critical role of multi-source data fusion in improving classification performance, especially in complex landscapes characterized by diverse land cover types and varying salinity levels.

5.2. Comparison of MSA-U-Net with Traditional Machine Learning and Baseline Models

As shown in Table 7, DL methods, especially MSA-U-Net, were superior in extracting salinization information compared to the other three ML models in MS and MS + SAR data. It may be due to the fact that U-Net and MSA-U-Net are based on convolutional neural networks (CNNs), which effectively capture subtle differences in spectral features and reflectance associated with different salinity levels [116]. However, for the SAR dataset, both ML and DL methods obtained poor classification results, likely due to the model’s high sensitivity to high-frequency noise (e.g., speckle) [117].

From the comparative analysis results in Table 6, the MSA-U-Net achieved the best performance, especially in classifying saline soils. In contrast to RF and U-Net, the extraction accuracy for the slightly salinized soil improved from 70.37% and 72.22% to 86.96%, and for moderately salinized soil, it increased from 79.22% and 68.18% to 81.99%. In general, the MSA-U-Net also showed significant improvements in OA and Kappa. These improvements can be attributed to the integration of the CBAM. The CBAM effectively highlights regions with significant spectral variations, such as salinity transitions, and emphasizes spectral indices closely linked to soil salinity. By incorporating the CBAM, the MSA-U-Net achieves a well-balanced synergy between feature preservation and enhancement, resulting in superior performance in multi-source data fusion tasks.

5.3. Impact of Attention Mechanisms on Skip Connections

To evaluate the impact of integrating the CBAM into the U-Net model for extracting and analyzing multi-source data, we compared its classification performance against four established approaches. The baseline U-Net was used as the control, alongside three enhanced variants: the U-Net incorporating a Spatial Attention Module (SAM) in the skip connections (referred to as U-Net-SAM), the U-Net incorporating a Channel Attention Module (CAM) in the skip connections (referred to as U-Net-CAM), and the MSA-U-Net. To ensure a consistent comparison, all models shared identical encoding and decoding architectures, maintaining structural equivalence across all approaches.

The aforementioned algorithms utilized multi-source features as input.

Table 8. Classification accuracies of the U-Net, the U-Net-SAM, the U-Net-CAM, and the MSA-U-Net for the study area.

Models	U-Net		U-Net-SAM		U-Net-CAM		MSA-U-Net
Classes	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)
VG	79.86	82.84	84.89	80.82	83.45	82.27	82.73	82.14
BL	97.30	75.63	92.43	89.53	88.65	90.61	89.19	92.18
WT	87.10	85.71	84.68	88.98	84.68	89.74	85.48	88.33
SS	92.41	83.91	90.51	91.08	91.14	90.57	90.51	91.67
MS	68.18	82.68	83.12	79.01	84.42	80.75	85.71	81.99
HS	72.22	92.20	83.89	90.96	84.44	83.98	88.89	86.96
OA (%)	82.98		86.81		86.28		87.34
Kappa coefficient	0.79		0.84		0.83		0.84
F1 score	0.82		0.86		0.86		0.87

presents a comparison of the classification results across four different models. The MSA-U-Net model achieved the highest performance, with an OA of 87.34%, a Kappa coefficient of 0.84, and an F1 score of 0.87, outperforming the other three models (Table 8). The U-Net-SAM and U-Net-CAM models showed comparable performance, each achieving an OA of approximately 86%. However, the baseline U-Net model demonstrated relatively poor performance, highlighting the limitations of the standard architecture without attention mechanisms.

The performance improvements in U-Net-CAM, U-Net-SAM, and MSA-U-Net over the baseline U-Net can be attributed to the incorporation of attention mechanisms. The CAM selectively amplifies important features across channels, refining feature representation [118]. The SAM enhances the model’s focus on critical spatial features, improving boundary detection, particularly in tasks with complex spatial patterns like salinization zones. The MSA-U-Net, which integrates both channel and spatial attention through the CBAM, demonstrates the best performance. This dual-attention mechanism effectively captures interdependencies across feature maps, allowing the model to distinguish subtle variations, such as differences in vegetation or salinity levels, which are crucial for accurate classification in multi-source data [119]. Moreover, integrating the CBAM into the skip connections helps preserve refined features during fusion, further boosting the model’s performance.

5.4. Potential and Limitations

Soil salinization, a significant global environmental issue, remains a key research focus. In the Keriya Oasis, salinity levels increase from the oasis toward the desert, a trend consistent with previous studies [33,44]. The superiority of deep learning combined with multi-source remote sensing over traditional machine learning approaches aligns with prior findings [57]. However, few studies have explored attention-based feature fusion strategies, which are crucial for effectively integrating multi-source data in soil salinization mapping. MSA-U-Net addresses this gap by incorporating such a strategy, enhancing its ability to capture complex spatial and spectral relationships. Specifically, the Convolutional Block Attention Module in skip connections refines feature representation by selectively emphasizing spatially and spectrally important information, improving the differentiation of subtle salinity variations. These factors likely contribute to its superior accuracy, yet further comparative studies are needed to fully validate its effectiveness and generalizability across different salinization contexts.

However, this study has several limitations. First, the labeling of saline soils is inherently subjective, and distinguishing between different salinity levels remains challenging, potentially leading to discrepancies between labeled data and actual conditions, thereby affecting classification accuracy. Second, while SAR data enhance classification performance, residual speckle noise in radar images may still introduce errors. Additionally, attention-based fusion strategies struggle to filter redundant information in real time and are highly sensitive to differences between optical and SAR images, making the effective fusion of multi-source data a persistent challenge in remote sensing. Research on land use and soil salinity monitoring using deep learning and multi-source remote sensing data remains limited. Future studies should focus on improving the deep feature fusion of optical and SAR images and exploring more advanced deep learning models (e.g., Vision Transformer [120], Vision Mamba [121]) and multi-source feature fusion strategies (e.g., linear feature fusion [122], gated-based feature fusion [123]) to enhance the accuracy and reliability of soil salinity monitoring.

6. Conclusions

A novel deep fusion U-Net model, MSA-U-Net, is proposed based on multi-source data to optimize salinized soil mapping performance across different degrees of salinization. To improve mapping accuracy, the model integrates the Convolutional Block Attention Module (CBAM) into the skip connections, combining the benefits of spatial and channel attention mechanisms. This dual-attention approach allows the model to capture both spatial and channel-wise interdependencies effectively, leading to more accurate feature extraction and classification. With an OA of 87.34%, a Kappa of 0.84, and an F1 score of 0.87 in the study area, the MSA-U-Net demonstrated excellent potential for precise and reliable salinized soil mapping. These findings underscore the advantages of integrating attention mechanisms into U-Net for multi-source remote sensing tasks.

The findings of this research offer a novel method for assessing soil salinization in the Keriya Oasis, aiding the region’s ecological stability. Furthermore, the MSA-U-Net model demonstrates its potential for broader remote sensing applications where multi-source data integration is critical, offering a novel perspective for monitoring soil salinity.