Dynamic Monitoring of Soil Salinization in Oasis Regions Using Spatiotemporal Fusion Algorithms

Abstract

Accurate dynamic monitoring of soil salinization in arid oasis regions is crucial for sustainable regional development. Remote sensing is widely used for large-scale, long-term monitoring, but its effectiveness is often limited by image quality and spatiotemporal resolution. Spatiotemporal fusion algorithms, due to their low cost and accessibility, are frequently applied to generate missing images. However, the applicability of these fused images for soil salinization inversion, the impact of different fusion strategies on image quality, and the potential for using multiple fused images to improve model accuracy remain unclear. This study evaluates the performance of three typical spatiotemporal fusion algorithms on raw spectral bands and compares two fusion strategies: fusion-then-index (FI) and index-then-fusion (IF), for two vegetation indices (NDVI and EVI) and two salinity indices (SI and SI2) related to soil salinization. Additionally, the inclusion of multiple fused images during the sampling period is examined for its effect on model accuracy. The results show that (1) spatiotemporal fusion images are suitable for soil salinization inversion, with accuracy depending on image quality; (2) for vegetation indices (NDVI and EVI), the IF strategy yields better results, while for salinity indices (SI and SI2), the FI strategy is more effective; and (3) combining multi-year and multiple fused images significantly improves model accuracy, though using fused images as auxiliary datasets or variables does not further enhance accuracy. These findings provide valuable insights for large-scale, long-term monitoring of soil salinization in arid regions.

1. Introduction

Soil salinization is a widespread and persistent environmental problem in arid and semi-arid regions, posing serious threats to soil health, agricultural productivity, and ecosystem sustainability. It has been recognized as a key barrier to achieving the United Nations’ Sustainable Development Goals (SDGs). Globally, about 10 million km² of land is affected, with the highest concentration in drylands [1,2]. Beyond China, severe salinization challenges have been reported in Australia, Central Asia, the Middle East, Sub-Saharan Africa, and parts of South America, where remote sensing and spatiotemporal fusion techniques have been increasingly applied for large-scale, long-term monitoring [3,4,5,6,7].

Xinjiang, located in the arid region of northwestern China, is characterized by limited precipitation and high evaporation, leading to extensive soil salinization [8]. Dynamic monitoring of soil salinization in Xinjiang’s oasis regions is essential for understanding salinization mechanisms, optimizing water and nutrient management, and guiding restoration efforts. However, similar to other global drylands, challenges such as sparse field data, variable image quality, and insufficient spatiotemporal resolution limit monitoring accuracy. This study responds to these global and regional challenges by evaluating spatiotemporal fusion algorithms and strategies to improve soil salinization inversion in arid oases.

Traditional methods for monitoring soil salinization involve field collection of soil samples, which are then analyzed for salinity levels in the laboratory. This approach is time-consuming, labor-intensive, and only provides point-scale data, making it unsuitable for large-scale, long-term dynamic monitoring of soil salinization. Currently, remote-sensing techniques have become the mainstream for regional-scale monitoring of soil salinization dynamics, both domestically and internationally [9,10]. Optical remote-sensing imagery is cost-effective, easily accessible, and provides long-term and wide spatial coverage. However, its image quality is susceptible to various factors, such as weather conditions and sensor limitations, which make it challenging to obtain high-spatiotemporal-resolution imagery [11,12,13].

Spatiotemporal fusion techniques can compensate for the inability of remote-sensing images to simultaneously achieve both high spatial and temporal resolution. These methods are based on the principles of spatial invariance with temporal correlation and temporal invariance with spatial correlation, allowing for the fusion of one or more pairs of high temporal resolution and high-spatial-resolution images [14,15,16,17]. Numerous spatiotemporal fusion algorithms have been developed to meet different application needs and sensor types. Zhu et al. [18], in a review, classified existing spatiotemporal fusion models into five categories based on their technical principles. These techniques have been widely applied in land cover classification [19,20,21], agricultural dynamic monitoring [22,23], and surface parameter inversion [24,25,26]. Han et al. [27] examined the applicability of fused images for soil salinization modeling by comparing various spatiotemporal fusion methods. However, it remains uncertain whether higher accuracy can be achieved by first calculating vegetation indices through band operations and then applying spatiotemporal fusion (IF), or by first performing fusion on the original spectral bands and subsequently extracting the vegetation indices (FIs).

Current remote-sensing models for soil salinization typically rely on images from the sampling period, nearby periods, or multi-month composites to extract the spectral features of sample points [25]. However, Xinjiang, located in the data-scarce arid region of northwest China, faces challenges due to its vast area and limited human and material resources. These factors result in long sampling intervals and a limited number of samples, introducing considerable uncertainty into soil salinization modeling [28]. Spatiotemporal fusion techniques can generate high-resolution remote-sensing images in both time and space. A key question is whether generating multiple fused images within the sampling period can address the limitations of single-period and multi-year composite images in capturing temporal and spatial information, thus improving the accuracy and spatiotemporal generalization of inversion models.

To investigate the above issues, this study first selected three widely used spatiotemporal fusion methods and applied them in the study area. We then compared two vegetation indices (NDVI and EVI) and two salinity indices (SI and SI2), all of which are closely related to soil salinization, to evaluate which fusion strategy (FI or IF) produces better results over the study period. Next, fused images combined with field-measured soil samples were used to construct a soil salinization inversion model. Finally, considering the quality of low-spatial-resolution images within the sampling period, we generated multiple high-quality fused images. By leveraging the ability of these images to dynamically capture multi-temporal information, we examined whether incorporating multiple fused images into traditional models could improve model accuracy. The specific objectives of this study were as follows:

(1)

To explore whether images generated through spatiotemporal fusion can be applied to soil salinization inversion modeling in arid zone oases.

(2)

To determine which fusion strategy is more effective for different vegetation and salinity indices.

(3)

To assess whether spatiotemporal fusion-based modeling strategies can improve the accuracy of traditional soil salinization inversion models.

2. Materials and Methods

2.1. Study Area

The Ogan-Kucha River Oasis, located in the Aksu region of Xinjiang, China (82°12′–83°32′E, 41°00′–41°44′N), covers an area of approximately 12,100 square kilometers. It is a typical oasis agricultural area situated at the foothills of the Tianshan Mountains. The oasis lies on the southern slopes of the Tianshan Mountains, at an elevation ranging from 950 to 1300 m, with the terrain gradually sloping from north to south. The northern boundary is formed by the Tianshan Mountains, while the southern part borders the Taklamakan Desert [29]. The Ogan-Kucha River Oasis experiences a typical temperate continental arid climate, with an average annual temperature of about 11.6 °C and less than 150 mm of annual precipitation, both of which exhibit significant seasonal variations [30]. The general overview of the study area is shown in Figure 1.

2.2. Collection of Field Data

From 13 October to 27 October 2017, we conducted a two-week field-sampling campaign in the Kucha Oasis to collect soil salinity data. Based on principles of rational sampling layout and representativeness, we collected 83 surface soil samples (0–10 cm) using a soil auger. The GPS coordinates of each sampling point were recorded using a handheld GPS device, and the collected samples were sealed in bags and then transported to the laboratory. In the lab, the samples were air-dried, ground, and sieved through a 2 mm mesh for further analysis.

To determine soil salinity, 10 g of each air-dried sample was placed in a beaker with 50 mL of distilled water (in a 1:5 soil-to-water ratio). The mixture was stirred thoroughly and allowed to sit at room temperature for 24 h before being filtered to obtain the extract. The electrical conductivity (EC) of the extract was measured using a conductivity meter. To ensure accuracy, each sample was measured twice, and the average value was obtained. The conductivity meter was calibrated using standard conductivity solutions [34].

2.3. Acquisition and Preprocessing of Remote-Sensing Images

This study utilized Landsat 8 and MODIS data for spatiotemporal fusion of remote-sensing images, as these datasets offer long time series, similar bandwidths, and minimal radiometric inconsistencies [35], making them widely used in spatiotemporal fusion research. The spectral band information for both datasets is provided in

Table 1. Spectral bands and spatial resolutions of remote-sensing images.

Landsat 8			MOD09GA
Band Name	Wavelength (nm)	Resolution (m)	Band Name	Wavelength (nm)	Resolution (m)
SR_B2(Blue)	452–512	30	sur_ref1_b03	459–479	500
SR_B3 (Green)	533–590	30	sur_ref1_b04	545–565	500
SR_B4 (Red)	636–673	30	sur_ref1_b01	620–670	500
SR_B5 (NIR)	851–879	30	sur_ref1_b02	841–876	500
SR_B6 (SWIR 1)	1566–1651	30	sur_ref1_b06	1628–1652	500
SR_B7 (SWIR 2)	2107–2294	30	sur_ref1_b07	2105–2155	500

. The formulas for these indices are presented in Table 2.

To ensure spatiotemporal consistency between the Landsat 8 and MODIS images, the time difference between the corresponding observations was kept within 3 days [36]. This study selected MOD09GA images from 9 September and 25 October 2017, and a Landsat 8 image from 6 September 2017, for spatiotemporal fusion. A Landsat 8 image from 24 October 2017 was used to assess the quality of the fused image. It is noteworthy that all images used in this study were selected to minimize cloud cover, with cloud coverage limited to less than 10%.

Table 2. Spectral indices calculated from remote-sensing images, with abbreviations, formulas, and references.

Index Name	Calculation Equation	Reference
NDVI	$(NIR-Red)/(NIR+Red)$	(Rouse et al., 1974 [37])
EVI	$\begin{array}{c}2.5\times (NIR-Red)/(NIR+6\times Red\\ -7.5\times Blue+1)\end{array}$	(Huete et al., 2002 [38])
SI1	${\left(Blue\times Red\right)}^{0.5}$	(Douaoui et al., 2006 [39])
SI2	${\left({Green}^2+{Red}^2+{NIR}^2\right)}^{0.5}$

Table 2. Spectral indices calculated from remote-sensing images, with abbreviations, formulas, and references.

Index Name	Calculation Equation	Reference
NDVI	$(NIR-Red)/(NIR+Red)$	(Rouse et al., 1974 [37])
EVI	$\begin{array}{c}2.5\times (NIR-Red)/(NIR+6\times Red\\ -7.5\times Blue+1)\end{array}$	(Huete et al., 2002 [38])
SI1	${\left(Blue\times Red\right)}^{0.5}$	(Douaoui et al., 2006 [39])
SI2	${\left({Green}^2+{Red}^2+{NIR}^2\right)}^{0.5}$

2.4. Spatiotemporal Fusion Methods, Strategies, and Accuracy Assessment

The Spatiotemporal Adaptive Reflectance Fusion Model (STARFM) [40] is based on the concept of weighted neighboring similar pixels. Initially, similar neighboring pixels are identified, and according to the temporal invariance of spatial correlation (as described by the First Law of Geography) [41,42], the similarity is assumed to remain constant at the prediction time. The target pixel is then computed using a weighted combination of the neighboring pixels’ spatial distances and spectral similarities. This model serves as the basis for many spatiotemporal fusion algorithms [43,44,45,46,47].

The Fit-FC model [48] is primarily based on the principle of temporal invariance of spatial correlation, which suggests that for pure pixels, the temporal correlation of features remains unchanged across spatial scales [49]. The model operates in three key steps: First, a regression equation is established between coarse-resolution images at the initial and prediction times. Then, spatial filtering is applied to mitigate blocky effects [50,51]. Finally, residual downscaling is performed to preserve more spectral information.

The Reliable and Adaptive Spatiotemporal Data Fusion (RASDF) method, proposed by Shi et al. [52], introduces a spatial distribution reliability index that optimizes fusion strategies and reduces uncertainties caused by sensor discrepancies. The method combines global and local unmixing models to better capture strong temporal changes. The global unmixing model addresses large-scale changes, while the local unmixing model refines regions with significant local variations, significantly improving the accuracy of the fusion results.

To obtain the NDVI, EVI, SI, and SI2 at the prediction time, this study employed two spatiotemporal fusion strategies: fusion-then-index (FI) and index-then-fusion (IF). The IF strategy involves first processing each band of the original images to compute the four indices, followed by applying three spatiotemporal fusion algorithms directly to these indices, resulting in single-band index images for the prediction time. In contrast, the FI strategy begins by using spatiotemporal fusion algorithms to fuse each band of the original images, producing multi-band images for the prediction time, from which the required indices are then calculated.

The accuracy assessment metrics used in this study included the average deviation (AD), correlation coefficient (CC), and structural similarity index (SSIM), all of which have been widely used to evaluate the performance of spatiotemporal fusion algorithms [53,54,55,56]. A lower AD, ideally approaching zero, indicates better fusion accuracy, as it reflects a smaller difference between the fused and reference images. The correlation coefficient (CC) assesses the linear relationship between the fused and reference images, offering a measure of how closely the pixel values of both images align. A higher CC indicates a stronger similarity between the two datasets. The SSIM, meanwhile, evaluates the perceptual quality of the fused image by comparing local luminance, contrast, and structure, and is especially sensitive to structural information. Together, these metrics provide a comprehensive evaluation of the fusion accuracy, considering both global differences and local structural similarities, which are critical for assessing the effectiveness of spatiotemporal fusion techniques.

2.5. Soil Salinization Inversion Model and Accuracy Assessment

Numerous studies have demonstrated that random forest regression (RFR) is widely applied in soil salinization inversion for oases in arid regions due to its high modeling accuracy and robust performance [57,58,59]. In this study, we employed the Scikit-learn library in Python 3.9 to build the RFR model, using default parameter settings to ensure a fair comparison of the accuracy between different inversion models. The sample data were randomly split into training and validation sets in a 7:3 ratio to assess the model’s generalization ability. The model’s accuracy was evaluated using the coefficient of determination (R²), root-mean-square error (RMSE), and relative prediction deviation (RPD). Higher R² and RPD values and lower RMSE values indicate greater modeling accuracy.

2.6. Composition of Multi-Temporal Remote-Sensing Images

The field-measured soil EC data for this study were collected in October 2017, a period when precipitation in the study area significantly decreased compared with September. Soil salinization is a dynamic surface parameter, strongly influenced by factors such as temperature and precipitation, and it exhibits significant differences between dry and rainy seasons [28]. This study selected seven Landsat 8 images with less than 10% cloud cover from October 2014 to October 2020, as shown in Figure 2. The median and mean values were used to composite the multi-temporal remote-sensing images. These methods are widely employed in multi-temporal remote-sensing studies due to their ability to effectively filter outliers and smooth noise [60,61].

3. Results

3.1. Statistical Description of Soil Samples

To assess the representativeness of the soil salinization samples and the distribution characteristics of the data, a descriptive statistical analysis was conducted on 83 soil samples (

Table 3. The statistical analysis of soil EC.

N	Min.	Max.	Mean	Median	IQR	SD	CV	Skewness	Kurtosis
83	0.113	114.200	17.295	5.940	15.308	25.847	1.495	2.057	3.641

). The results show that the mean electrical conductivity (EC) value is 17.295 dS m⁻¹, with a maximum value of 114.200 dS m⁻¹ and a minimum of 0.113 dS m⁻¹, indicating that the study area predominantly exhibits low salinity, while some localized areas exhibit higher salinity. The standard deviation (25.847) and coefficient of variation (1.495) are relatively high, reflecting significant spatial variation in salinity levels across the study area. The interquartile range (IQR) is 15.308, indicating that 50% of the data points are spread across a wide range of EC values. The skewness value of 1.495 suggests a right-skewed distribution, with most samples concentrated in the lower-EC value range, while a few high salinity samples elevate the overall distribution. The kurtosis value of 3.641 indicates a relatively peaked distribution, likely influenced by a few extreme values. Overall, the soil salinity in the study area shows clear spatial heterogeneity, with relatively large variability, highlighting significant differences in salinity levels across different regions. These findings provide essential insights for subsequent modeling analysis.

3.2. Comparison of Spatiotemporal Fusion Methods

Visually, the fused images generated by the RASDF algorithm are the closest to the reference image in both spatial structure and spectral consistency. For instance, farmland boundaries and rural roads are better preserved (Figure 3a vs. Figure 3b), while the Fit-FC algorithm produces visibly blurred textures (Figure 3d), and STARFM shows muted tonal contrast (Figure 3c).

These visual impressions are supported by quantitative accuracy metrics in

Table 4. The prediction accuracy of each band in the fused images.

Band		RASDF			STARFM			Fit-FC
	AD	CC	SSIM	AD	CC	SSIM	AD	CC	SSIM
Blue	−0.002	0.855	0.896	−0.003	0.851	0.883	−0.003	0.827	0.874
Green	−0.012	0.915	0.886	−0.013	0.913	0.878	−0.013	0.896	0.872
Red	−0.011	0.923	0.867	−0.012	0.913	0.840	−0.020	0.905	0.849
NIR	−0.027	0.799	0.705	−0.029	0.727	0.626	−0.006	0.812	0.772
SWIR1	−0.003	0.889	0.814	−0.011	0.882	0.806	0.010	0.800	0.766
SWIR2	−0.002	0.917	0.843	−0.005	0.899	0.803	−0.003	0.899	0.812

and Figure 4. RASDF yields the highest correlation coefficients (CCs) and structural similarity index (SSIM) in most bands, particularly blue (CC = 0.855; SSIM = 0.896), red (CC = 0.923; SSIM = 0.867), and SWIR2 (CC = 0.917; SSIM = 0.843), with minimal absolute difference (AD) values, indicating lower spectral deviation.

In contrast, the Fit-FC algorithm performs best in the NIR band, achieving the highest accuracy (AD = −0.006, CC = 0.812, and SSIM = 0.772). This is consistent with Figure 5, which shows that the NIR band has the weakest correlation between the initial and prediction times, reflecting significant temporal variability. As a regression-based method, Fit-FC is better suited to handle such dynamic changes during vegetation senescence.

Notably, most AD values are negative across all bands, indicating that the fused images tend to be slightly darker than the reference image, which reflects a general trend of spectral underestimation. This aligns with visual observations in vegetated and transition zones.

The area highlighted in Figure 3 further shows that Fit-FC demonstrates relative robustness to cloud contamination, especially when the initial high-resolution input image is partially affected by clouds. This supports earlier findings that Fit-FC is well-suited for reconstructing NDVI time series under temporally unstable surface conditions [48,62].

3.3. Selection of Fusion Strategies

For the vegetation indices NDVI and EVI, the fused images obtained using the IF strategy are visually closer to the reference images (Figure 6 and Figure 7). Among the algorithms, Fit-FC performs the best under the IF strategy, both in terms of color tone and spatial detail. The RASDF algorithm, under the IF strategy, exhibits significant spectral underestimation in non-vegetated areas, such as water bodies and the desert–oasis transition zone, compared with the Fit-FC algorithm. The STARFM algorithm shows considerable tonal differences from the reference image, with severe spectral distortion.

The visual evaluation aligns with the quantitative analysis, as the results (

Table 5. Comparison of fusion accuracy for spectral indices under FI and IF strategies.

Strategy	Index		RASDF			STARFM			Fit-FC
		AD	CC	SSIM	AD	CC	SSIM	AD	CC	SSIM
FI	NDVI	−0.029	0.796	0.678	−0.033	0.722	0.538	0.015	0.821	0.760
	EVI	−0.027	0.808	0.721	−0.023	0.719	0.573	0.007	0.809	0.805
	SI	−0.004	0.890	0.919	−0.005	0.863	0.903	−0.004	0.864	0.905
	SI2	−0.006	0.882	0.863	−0.017	0.841	0.807	−0.007	0.875	0.831
IF	NDVI	−0.008	0.877	0.799	−0.008	0.837	0.624	0.000	0.912	0.833
	EVI	−0.013	0.857	0.817	−0.013	0.756	0.602	−0.002	0.921	0.890
	SI	−0.004	0.884	0.918	−0.008	0.880	0.911	−0.004	0.851	0.895
	SI2	−0.017	0.845	0.780	−0.018	0.815	0.759	−0.017	0.798	0.792

) indicate that the Fit-FC algorithm outperforms the other two spatiotemporal fusion algorithms under both the FI and IF strategies. Specifically, the Fit-FC algorithm achieves the best fusion performance for NDVI and EVI under the IF strategy, with AD values of 0.004 and −0.002, correlation coefficients (CCs) of 0.912 and 0.921, and SSIM values of 0.833 and 0.890, respectively. It is worth noting, however, that despite producing the best fusion results, the Fit-FC algorithm still smooths edge and texture details to some extent.

For the salinity index SI, the six fused images show minimal tonal differences compared with the reference image, with most of the discrepancies concentrated in the desert–oasis transition zone (Figure 8). In terms of spatial details, the FI strategy preserves more texture and edge information compared with the IF strategy, with the RASDF method producing the best fusion results under the FI strategy. This is consistent with the quantitative analysis (Table 5), where the RASDF_FI method achieves the best performance, with AD = −0.004, CC = 0.890, and SSIM = 0.919.

For the salinity index SI2, the overall tone of the fused images obtained using the FI strategy is closer to the reference image than those obtained using the IF strategy (Figure 9). The RASDF algorithm performs better than the STARFM and Fit-FC algorithms. The STARFM algorithm exhibits slight blurring in the transition zone and some high-salinity areas, while Fit-FC provides moderate detail preservation. Although the differences between Fit-FC and the reference image are reduced, its overall performance remains inferior to that of the RASDF algorithm. This is supported by the quantitative analysis, where the results show AD = −0.006, CC = 0.882, and SSIM = 0.863.

3.4. Applicability of Spatiotemporal Fusion Images

To investigate whether fused images from spatiotemporal fusion can be used for dynamic monitoring of soil salinization, this study employed fused images generated by three spatiotemporal fusion methods—RASDF, STARFM, and Fit-FC—along with reference images and field-measured soil EC values to build random forest regression models for quantitative soil salinization estimation. The accuracy of the four random forest models is shown in

Table 6. Comparison of modeling accuracy between fused images from three algorithms and reference image.

	Training R2	Validation R2	Training RMSE	Validation RMSE	Training RPD	Validation RPD
Reference	0.815	0.652	11.035	15.214	2.322	1.696
STARFM	0.821	0.548	11.772	13.295	2.362	1.487
RASDF	0.831	0.714	11.214	11.469	2.434	1.870
Fit-FC	0.829	0.543	11.537	13.026	2.416	1.480

. The results indicate that fused images from all three fusion algorithms can be used in soil salinization estimation models. Among them, the RASDF algorithm produced the highest modeling accuracy, with validation set values of R² = 0.714, RMSE = 11.469, and RPD = 1.870. The modeling accuracy of the fused images from the STARFM and Fit-FC algorithms did not reach the level achieved by the reference images.

This is likely due to the superior quality of the RASDF fused images, which resulted in minimal error propagation through the random forest regression model. The modeling results demonstrate that spatiotemporal fusion images can be effectively used for soil salinization estimation, and the model’s accuracy is directly influenced by the quality of the fused images. The higher the image quality, the better the model’s performance.

3.5. Comparison of Modeling Results for Multi-Temporal Composite Images

During the sampling period, significant changes occurred in the land cover of the study area, making it difficult to capture detailed surface changes and temporal dynamics using single-period images. Therefore, this study generated seven fused images within the sampling period by considering the quality of MODIS images (Figure 10). Next, median and mean composites were created from the seven fused images and seven multi-year October images. Additionally, we combined all 14 images and applied the same compositing methods (Figure 11). The composite images were then used for random forest regression modeling, with the model accuracy presented in

Table 7. Comparison of modeling accuracy for median and mean composites of fused images, multi-year monthly images, and combined images with reference image-based accuracy.

	Training R2	Validation R2	Training RMSE	Validation RMSE	Training RPD	Validation RPD
Reference image	0.815	0.652	11.035	15.214	2.322	1.696
Fused images–median	0.863	0.614	10.263	11.717	2.699	1.609
Fused images–mean	0.797	0.726	12.048	12.011	2.221	1.910
Multi-year images–median	0.862	0.600	10.251	12.447	2.687	1.581
Multi-year images–mean	0.857	0.724	10.505	10.347	2.640	1.905
Combined images–median	0.866	0.807	10.500	6.525	2.734	2.275
Combined images–mean	0.818	0.676	11.426	13.056	2.342	1.757

.

Overall, the multi-temporal fused images achieved better modeling accuracy compared with single-period images, whether they were spatiotemporal fusion composites within the sampling period or multi-year October composites. When modeling separately, mean composites outperformed median composites, as there were fewer outliers between datasets. However, when data from different sources were mixed, as shown in Figure 12, the number of outliers increased, and in this case, median composites performed better than mean composites.

Using spatiotemporal fusion images combined with multi-year October images further improved model accuracy, with the median composite achieving an R² of 0.807 on the validation set. This is because spatiotemporal fusion images provide high temporal resolution, allowing for a more accurate reflection of surface dynamics during the sampling period. Multi-year October composite images, meanwhile, provide long-term change information, smoothing out overall trends and mitigating the impact of individual image quality issues. The mixed images leverage both approaches, enabling the model to capture detailed temporal information while also accounting for long-term trends. This provides new insights for future dynamic inversion and precise estimation of soil salinization.

4. Discussion

4.1. Simulated Image Experiments

To eliminate the interference of sensor geometric and radiometric inconsistencies with the spatiotemporal fusion strategy results, this study designed a spatiotemporal fusion simulation experiment. High-resolution Landsat 8 images were aggregated into low-resolution simulated images, which were then used as inputs for the spatiotemporal fusion process. This method has been widely applied in evaluating the performance of spatiotemporal fusion algorithms [62,63,64,65].

The RASDF algorithm was selected for the simulation experiment due to its strong performance in both raw band and index fusion. A per-pixel comparison of the errors between the four fused index images obtained using the FI and IF strategies and the reference images was conducted, and the distribution and percentage of pixels where either FI or IF performed better for each index were calculated, as shown in Figure 13.

For the vegetation indices NDVI and EVI, the IF strategy produced superior fused images in most areas of the study region, as indicated in Figure 13a,b, with pixel percentages of 63.4% and 62.2%, respectively. This finding is consistent with the results of Zhou et al. [62], who also found that vegetation indices tend to benefit more from index-first fusion strategies. In contrast, for the salinity indices SI and SI2, the FI strategy performed better, as depicted in Figure 13c,d, with pixel percentages of 53.1% and 61.0%, respectively.

To elucidate the underlying mechanisms of the aforementioned phenomena, we conducted a statistical analysis of high-resolution imagery in the initial temporal phase. The calculation of global variances for original spectral bands and spectral indices (Figure 14) revealed that the vegetation indices (NDVI and EVI) exhibited significantly higher variances compared with both the original bands and salinity indices (SI and SI2). This discrepancy can be attributed to the inherent characteristics of the vegetation indices, which not only enhance vegetation-related signals but also amplify potential error components. Specifically, even minor spectral variations in high-resolution imagery may induce pronounced error fluctuations in vegetation indices through their non-linear transformation processes. Such an error amplification mechanism likely facilitates the identification and removal of these enhanced noise signals during spatial filtering procedures. Similar effects have been reported by Pôças et al. [66], who emphasized the vulnerability of vegetation indices to small spectral disturbances. Moreover, the elevated data variance in vegetation indices may contribute to more robust spatial interpolation outcomes, potentially explaining their superior performance under the IF strategy. This aligns with the conclusions of Liu et al. [23], who found that greater variance often improves spatial estimation quality.

In contrast, the salinity indices (SI and SI2) demonstrated distinct characteristics, showing variance levels comparable to original spectral bands without significant signal amplification. Their relatively flat spectral response patterns and weaker signal intensity render them more susceptible to interference from other land cover features and environmental noise. These inherent properties might fundamentally account for their different behavioral patterns compared with vegetation indices. This finding aligns with Chen et al. [64] concerning the critical influence of index characteristics on fusion performance, and further emphasizes the necessity of adopting appropriate fusion strategies tailored to specific spectral index types when reconstructing index imagery through spatiotemporal fusion algorithms. The differential responses between vegetation and salinity indices underscore the importance of index-specific optimization in remote-sensing data fusion applications.

4.2. Incorporating Auxiliary Datasets and Variables

Due to the vast size of the study area, field sampling requires extended periods, resulting in a limited number of measured sample points, which is insufficient for modeling purposes. Although the random forest model can reduce the risk of overfitting, it is still impacted by the issue of small sample size.

Spatiotemporal fusion technology can generate multiple remote-sensing images during the sampling period. Compared with using a single reference image, incorporating multiple fused images as auxiliary datasets or features could theoretically improve the accuracy of traditional soil salinization models.

Table 8. Comparison of model accuracy after incorporating auxiliary datasets and variables with reference image accuracy.

	Training R2	Validation R2	Training RMSE	Validation RMSE	Training RPD	Validation RPD
Reference image	0.815	0.652	11.035	15.214	2.322	1.696
Auxiliary datasets	0.892	0.636	8.555	14.983	3.044	1.657
Auxiliary features	0.826	0.626	11.155	14.022	2.399	1.636

Note: While training metrics improve after incorporating auxiliary datasets and features, the validation R² decreases despite reductions in RMSE and RPD fluctuations. This inconsistency may reflect a loss of generalization due to high feature redundancy, increased dimensionality, and limited sample size, potentially leading to overfitting.

presents the accuracy of these two methods. However, compared with single-image inversion, the model accuracy did not improve. The potential reasons for this are as follows:

(1)

When additional high-resolution images are input as auxiliary datasets, the temporal proximity of the images leads to high sample correlation (strong autocorrelation), causing the model to learn more noise than useful features during training, thus reducing its generalization ability. Additionally, errors or inconsistencies introduced during the spatiotemporal fusion process may accumulate, affecting the model’s performance.

(2)

When used as auxiliary features, the new features may have high redundancy with the existing ones, offering little additional useful information and increasing model complexity, which leads to more noise during training. Furthermore, the increase in feature numbers introduces high-dimensional data issues, especially when the sample size remains unchanged, making it difficult for the model to capture the underlying structure and reducing predictive performance.

In the future, these strategies may be more effective under conditions such as access to more diverse data sources, improvements in the accuracy and consistency of spatiotemporal fusion techniques, the use of advanced feature extraction and selection methods, and the availability of larger sample sizes and greater computational power, which could enhance the model’s predictive performance.

5. Conclusions

In this study, we aimed to evaluate the applicability of three widely used spatiotemporal fusion algorithms (RASDF, Fit-FC, and STARFM) and different modeling strategies for the dynamic inversion of soil salinization in arid oasis regions. The results show that fused images generated by these algorithms effectively compensated for missing data, with RASDF achieving the highest overall accuracy across most spectral bands, while Fit-FC performed best in the NIR band. All three algorithms produced fused images suitable for inversion modeling, and in some cases, fusion-based models outperformed those using single-reference imagery. For the vegetation indices (NDVI and EVI), the strategy of calculating indices first and then applying fusion (IF) yielded better results, whereas for the salinity indices (SI and SI2), the opposite was observed—the fusion-first (FI) strategy proved more effective. This contrast, identified for the first time, highlights the importance of index-specific strategy selection. Moreover, using median composites of fused and multi-year historical images improved model performance, but directly incorporating multiple fused images as auxiliary datasets or features did not further enhance model accuracy, likely due to noise accumulation and feature redundancy. Overall, this study confirms the feasibility and potential of spatiotemporal fusion imagery for dynamic soil salinization monitoring, while also underscoring the importance of selecting suitable fusion strategies and modeling approaches based on index characteristics. Future research should focus on identifying better fusion algorithms, more suitable fusion strategies, and improved modeling approaches for dynamic monitoring and precise inversion of soil salinization.