The IMXL spans 70–120° E longitude and 30–50° N latitude, covering 11 arid or semi-arid inland lakes (Ma et al., 2011). The topographic patterns of alternating mountains and basins direct surface water and groundwater into the depressions, forming numerous terminal lakes. The salinity of lakes in this region spans multiple magnitudes (Table 1). Referencing the classification criteria for lake salinity (Hammer, 1986; Zheng et al., 2002), the lakes were categorized as freshwater (< 1 g L⁻¹), brackish (1–3 g L⁻¹), and oligosaline (3–35 g L⁻¹) types. The salinity dataset to be produced covers the time span from 2016 to 2024 and the geographical coverage of 11 typical lakes within IMXL (Fig. 1). Notably, some years are missing data for Lake Hulun and Lake Juyan due to insufficient SAR imagery or multi-source data matching pairs.

Data products for each lake contain both raw data and derived data. The raw data includes Sentinel-1 SAR backscatter data and incidence angle, Sentinel-2 MSI remote sensing reflectance, and Landsat-8 Thermal Infrared Sensor (TIRS) temperature data. Derived data comprise daily, quarterly, yearly and all-season average salinity rasters, along with their mean and standard deviation statistical files (Table 2). All raw or derived raster data is stored in TIF format based on the WGS1984_UTM_Zone projection with a spatial resolution of 10 m. The statistical documents are compiled in Excel format.

From July 2017 to October 2024, 322 field survey data were collected from 11 lakes using two approaches: uniformly distributed field measurements and the China National Environment Monitoring Centre (CNEMC, https://www.cnemc.cn/, last access: 1 June 2025). These data were divided into Dataset one and Dataset two. Dataset one contains 257 field salinity data, with 69 sets of collected surface water samples (depth < 0.5 m) and simultaneously measured SDD and in situ spectral data. Dataset one was used for training, testing, and five-fold cross-validation (CV) of the ML salinity algorithm and also as the initial driver for the salinity mechanistic model. Dataset two contains salinity data transformed from conductivity measurements by CNEMC or the latest data not involved in modeling. The conversion function was established based on synchronized measured conductivity and salinity in the IMXL (R2= 0.99, RMSE = 1.76 ppt, and MAPE = 28.4 %) (Rusydi, 2018), as shown in Table 3. Dataset two (65 samples) was selected to ensure spatial independence (no overlap with Dataset 1's sampling points) and temporal coverage (2017–2024, including 10 % of 2024 data not used in modelling) to validate generalization. The surface water salinity (units ppt) and conductivity (units µS cm⁻¹) of the lake were in situ measured using a YSI multi-parameter water quality instrument (YSI ProDSS, USA). Spectral Evolution PSR-1100f (350–1050 nm, 1 nm spectral resolution) was used to measure water surface upward radiance (Lsw), sky radiance (Lsky), and gray plate radiance (Lp) at an observation direction of 40° from the nadir and 135 degrees from the Sun (Mobley, 1999; Mueller et al., 2003). These radiance data were further used to calculate Rrs(λ), with the formula given as follows (Mobley, 1999): 1Rrsλ=Lsw-ρ×Lsky×ρp/π×Lp where ρ represents the water-air interface reflectance, assumed to be 0.0028 under calm conditions (wind speed < 4 m s⁻¹), and ρp denotes the reflectance of the reference gray plate defined as 0.30. We calculated it by setting Rrs(λ) between 950 and 1050 nm to zero, since a low signal-to-noise ratio was observed within this range (Lee et al., 2016). Finally, convolve Rrs(λ) through the spectral response function (SRF) to simulate bands of the MSI.

Field water samples were stored in polyethylene bottles and rapidly transported back to the laboratory for analysis at the end of the cruise. Water samples were filtered using Whatman GF/F films with a 0.7 µm pore size to extract chlorophyll a (Chl a, µg L⁻¹), suspended particulate matter (mg L⁻¹), and CDOM samples. Laboratory measurements of these parameter concentrations can be found in Deng et al. (2024).

ERA5 land reanalysis data was published by the European Center for Medium-Range Weather Forecasts (ECMWF, https://cds.climate.copernicus.eu/, last access: 1 June 2025) and can be downloaded for free. This product provides hourly meteorological variables in GeoTIF format with a spatial resolution of approximately 11 km (Muñoz Sabater, 2019). It has been available since 1980 and provides a continuous record that can compensate for gaps in field meteorological observations. We downloaded hourly wind speed (WS, m s⁻¹), temperature (TEMP, °C), evaporation (EVP, mm) and precipitation (PRE, mm) data for each lake from GEE during 2016–2024 and calculated daily data on this basis. Additionally, nighttime light (NTL, nW cm⁻² sr⁻¹) in the Lake Daihai sub-basin was derived from the Visible Infrared Imaging Radiometer Suite (VIIRS). Population (POP) data were sourced from the statistical yearbook.

This study proposed a brand-new framework integrating microwave and optical data to estimate lake salinity, as illustrated in Fig. 2. It consists of three modules: data processing and feature building, model construction and ensemble, and lake salinity estimation. Module 2 is the core of the framework, aiming to use wave spectrum, backscatter coefficient, and dielectric constant models as forward models to build a salinity mechanistic model, then establish ML models, and finally construct an ensemble model by coupling mechanistic and ML results via a Generalized Additive Models (GAM) model.

The differences between estimated and measured salinity were evaluated by using several statistical metrics named R2, root mean square error (RMSE), mean absolute error (MAE), bias (system error), and mean absolute percentage error (MAPE). These metrics were calculated as follows: 9RMSE=1n∑i=1nyi-y^i210MAE=1n∑i=1nyi-y^i11Bias=1n∑i=1nyi-y^i12MAPE=1n∑i=1nyi-y^iy^i×100% where yi is the field measured salinity, y^i is the model estimated salinity, i denotes the ith sampling point data, and n is the number of sampling point pairs.

The constructed stacking model (N= 84, RMSE = 0.60 ppt, and MAPE = 2.3 %) outperformed four ML models and the mechanistic model, and the predicted salinity distributed consistently along the 1:1 line without significant underestimation or overestimation (Fig. 3). The five-fold CV result for the stacking model was close to the accuracy of the 30 % dataset test (N= 257, RMSE = 0.38 ppt, and MAPE = 6.9 %), with the range of estimated salinity consistent with measured salinity and no outliers observed, indicating that the model has good generalization and stability without significant dependence on the training set (Fig. 3j). The performance of the mechanistic model was second only to the ensemble model (N= 257, RMSE = 0.80 ppt, and MAPE = 13.3 %), better than the results of five-fold CV for each ML model (RMSE > 0.97 ppt and MAPE > 15.1 %) (Fig. 3m). The XGB model showed the best accuracy among the four ML models, followed by CNN and DNN, while the RFR model was the worst performer (Fig. 3). Additionally, the bar chart shows that the ensemble model outperforms the ML algorithms in terms of accuracy (RMSE < 0.55 ppt and MAPE < 8.7 %) (Fig. 3k, l), indicating that incorporation of the mechanistic model improves overall performance.

Figure 4 displays the SHAP values of the selected features for four ML salinity models. Among these models, lake area, theta, and VV exhibit significant contributions, as reflected in their higher SHAP values compared to other variables. The lake area contributes the most because it correlates with water volume, which directly influences the degree of salinity dilution (McGrath et al., 2025). VV and theta are important parameters because of their sensitivity to water surface scattering mechanisms, while salinity affects scattering intensity by altering the dielectric constant of the water (Reul et al., 2009). B6 reflectance ranks third in contribution within the XGB, RFR, and DNN models, indicating it is an optically sensitive band for salinity. In the CNN model, the VV band contributes most significantly, while alpha was the most influential optical index.

A simpler positive correlation (edf = 2.19) with salinity was found for s5 (Mechanistic) compared to the other variables (edf > 2.76) in the PDP of the ensemble model (Fig. 5). The PDP curve of the mechanistic model has a broader range of variation on the y axis than the ML models, with its 95 % confidence region encompassing most of the sample data, suggesting that the model contributes significantly and reliably to salinity prediction for the ensemble model. At a salinity range of 0–4.71 ppt, XGB, RFR, CNN, and Mechanistic models contribute negatively to the ensemble model, but their contributions turn positive at salinities exceeding 9.52 ppt. The contribution pattern of the DNN model differed from those of the aforementioned models (Fig. 5c). Overall, the integrated salinity model works admirably by effectively coupling the virtues of both mechanistic and ML models, making it successful for estimating lake salinity.

A single-scene comparison for lakes with matching field sampling points was used to examine the spatial quality of salinity data produced by the stacking model, ML model, and mechanistic model (Fig. 6). Lake salinity maps generated by the stacking model show a smooth transition from the shore to the center in 10 lakes, whereas a slight discontinuity of salinity was observed in each lake using ML or mechanistic models. Especially in nearshore waters, DNN, CNN, and mechanistic models exhibit outliers due to land adjacency effects, while the stacking model corrects predictions by combining accurate salinity derived from XGB and RFR, such as in Lake Hongjiannao, as shown in Fig. 6h2, h3. However, salinity values in Lake Bosten's nearshore pixels were not fully corrected because the DNN and CNN models overestimate salinity influenced by mixed pixels (Fig. 6c4, c5, c7). The stacking model successfully captured the spatial variations of freshwater dilute salinity by combining the capabilities of ML models in salinity estimates at river inlets into lakes, compensating for the limitations of mechanistic models, as shown in Fig. 6i7. Furthermore, the RMSE of the stacking model was observed to be lower than other algorithms across nine lakes in single-scene comparisons, with only the mechanistic model performing better in Lake Juyan. Finally, it was observed that the stacking model effectively suppressed salinity outliers to guarantee the quality of the dataset. In brief, the stacking model produced maps of lake salinity with smoother spatial variations and richer detail, yielding a higher-quality dataset.

Comparisons between the MSI-based XGB salinity algorithm (Deng et al., 2025) and the SAR-based mechanistic algorithm were performed for each lake (Fig. 7), and it can be revealed that the stacking algorithm has higher and more stable accuracy with an average RMSE of 0.24 ppt. Although the XGB and the mechanistic models outperformed the stacking algorithm in some lakes, including Lake Ulungur, Lake Juyan, and Lake Nanhaizi, both models (with RMSE of 0.45 and 0.57 ppt, respectively) still showed slightly lower precision across the entire region. The stacking model and the mechanistic model outperformed the XGB salinity model in oligosaline-type lakes, suggesting that the mechanistic model improves the accuracy of the stacking algorithm under salinity exceeding 3 ppt. In addition, the points for the three models were concentrated on the Taylor plots of several lakes. It can be deduced that the mechanistic model and XGB model also exhibit reliable performance to provide rational data support for constructing the stacking model. Overall, the proposed algorithm combines the strengths of both the physically constrained model and the data-driven model in that it avoids the single model or data source as well as improving the precision in complex inland water.

To further objectively evaluate the accuracy and scientific validity of the proposed framework, independent validation was performed using Dataset 2, which was not involved in model training, testing, or CV. The validation density was insufficient due to the absence of some lake stations and historical salinity data. The independent validation (Fig. 8) demonstrates that salinity estimates from the integrated algorithm predominantly align well with measured salinity along the 1:1 line (N= 65, R2= 0.97, RMSE = 0.89 ppt, and MAPE = 37.6 %). Only Lake Ulansuhai showed a slight overestimation, likely affected by aquatic vegetation pixels. An underestimated validation point was observed in the southern part of the Juyan Lake, associated with the complex water characteristics in the river inlet region. No significant underestimation or overestimation was observed in other lakes. These results confirm that the integrated algorithm combining microwave and optical data has considerable accuracy in retrieving salinity in inland lakes.

Pixel-based histogram statistics were performed on the salinity raster generated by the stacking model for each lake (Fig. 9). Using frequency instead of pixel counts in traditional histograms, the mean salinity, standard deviation (STD), and frequency proportions of different salinity ranges were calculated. The distribution patterns of salinity can be visualized through a frequency histogram, which helps identify outliers and objectively assess salinity map quality. Outliers are usually found dispersed and dramatically changed, appearing as discontinuities in histograms. A single peak pattern of salinity histogram was observed in most freshwater or brackish lakes, including Lake Sayram, Lake Ulungur, Lake Bosten, Lake Nanhaizi, and Lake Hulun. These minimal spatial and interannual variations within the lakes align with the field measurements found. A double-peak characteristic was observed in Lake Juyan and Lake Ulansuhai, with the primary and subsidiary peaks distributed consecutively, caused by the differences between the northern and southern regions for the lake salinity. However, the subpeaks in the bimodal distribution of Lake Chagannaoer showed low frequency and discontinuity, indicating that this portion of the data may be abnormal. The oligosaline-type lakes (e.g., Lake Hongjiannao, Lake Daihai, and Lake Dalinor) commonly showed multi-peak characteristics and wide fluctuation ranges, suggesting high interannual or spatial variations in salinity.

Further analyze image quality by examining the pixel frequency across different salinity ranges. The pixel frequency in Lake Sayram, Lake Ulungur, Lake Bosten, and Lake Nanhaizi was more than 90 % within the interval of 0–2 ppt. Lake Juyan (46.6 %), Lake Ulansuhai (89.1 %), and Lake Daihai (88.2 %) accounted for the highest pixel proportions within salinity ranges of 4–6, 0–4, and 8–16 ppt, respectively. Lake Hongjiannao and Lake Dalinor had the highest pixel proportions in the range of 4–8 ppt, with 68.5 % and 70.4 %. These high-proportion intervals align with the distribution of field measurement salinity, and the small proportion of abnormally low-frequency pixels demonstrated that the salinity data generated by the stacking model had a good image quality with only a few outliers.

The annual scale salinity map of 11 lakes was shown in Fig. 10, with its color fluctuations effectively delineating the spatial differences in water salinity. The water salinity exhibits spatial homogeneity within several lakes, such as Lake Sayram, Lake Ulungur, and Lake Bosten. Lake salinity discontinuities were commonly observed in shore or river inflow zones, such as the western part of Lake Ulansuhai, the southern region of Lake Juyan, the northern part of Lake Chagannaoer, and the eastern part of Lake Hulun, which were primarily affected by mixing pixels or freshwater dilution (Han et al., 2021). This phenomenon with freshwater dilution of water salinity is prevalent in inland lakes receiving surface runoff, as well as lakes on the Tibetan Plateau (Wang et al., 2023a). In addition, the salinity in Lake Nanhaizi was higher in the east than in the west. Seasonal salinity patterns showed greater spatial heterogeneity than annual patterns across all lakes (Fig. 11), particularly in Lake Ulansuhai, Lake Nanhaizi, and Lake Chagannaoer. It indicates that the spatial pattern of salinity is regulated by external environmental factors, such as seasonal variations in precipitation, runoff, and evaporation (Rimmer et al., 2006; Yihdego and Webb, 2012; Jiang et al., 2022).

Interannual salinity trends were calculated using the Mann-Kendall trend test (significance level α= 0.05). Lake Daihai (slope = 0.48 ppt yr⁻¹, p<0.01) and Lake Dalinor (slope = 0.22 ppt yr⁻¹, p<0.05) showed significant increasing trends (Fig. 12a), and the management agency requires attention to this occurrence. Other lakes had non-significant interannual trends, with a comparable proportion of them increasing or decreasing in salinity. Significant seasonal variations of salinity (p<0.05) were observed in some lakes (Fig. 12c), with higher salinity in summer and autumn compared to spring due to enhanced evaporation during summer, including Lake Ulungur, Lake Ulansuhai, Lake Hongjiannao, and Lake Chagannaoer. The other lakes did not present significant seasonal changes, and the freshwater and brackish lakes had lower seasonal salinity differences compared to oligosaline lakes, likely due to their greater water storage capacity and more reliable water supply sources (Rusuli et al., 2016).

The development of remote sensing algorithms for water salinity has long presented technical challenges, particularly in inland lakes. The lake salinity dataset generated using the proposed framework exhibits a spatial resolution of 10 m, significantly higher than ocean salinity products such as SMOS, SMAP, and Aquarius (> 40 km) (Hu and Zhao, 2022; Jang et al., 2022; Zhang et al., 2023), thereby enabling much greater spatial detail. It also demonstrates higher accuracy (RMSE = 0.60 ppt) compared to even a regional-scale Tibetan Plateau lake salinity product (RMSE = 12.51 g L⁻¹) (Xu et al., 2024). To further confirm the scientific validity of the GAM algorithm, it was compared with Bayesian Model Averaging (BMA) (Hoeting et al., 1999), a method suited for addressing model uncertainty, with the BMA model's key hyperparameters listed in Table 4. The GAM model outperformed the BMA method (N= 84), with lower RMSE (0.60 ppt vs. 0.88 ppt) and MAPE (2.3 % vs. 12.6 %) in both the test and five-fold CV (Figs. 3 and 13). It suggested that the GAM algorithm more effectively handled the nonlinear relationship between the outputs of the mechanism and the ML model, with its average edf of 6.58 (n_splines = 20) implying moderate model complexity while avoiding overfitting.

In Lake Juyan's southern river estuary, the stacking model showed higher RMSE (0.41 ppt) compared to open water (0.19 ppt), due to suspended particulate matter interference. This underestimation accounts for ∼ 5 % of total pixels in estuary zones. Subsequent works will focus on integrating Sentinel-2 shortwave infrared (SWIR) band data for correcting this effect (Knaeps et al., 2015). And the spatial heterogeneity stemming from the 3 d matching interval between TIRS and Sentinel data may introduce errors (Jin et al., 2016), which future work will mitigate by combining Sentinel-3 Sea and Land Surface Temperature Radiometer (SLSTR) for temperature correction. Limited by the salinity gradient of the training set (< 35 ppt), the salinity algorithm has insufficient applicability in polysaline (35–50 g L⁻¹) and hypersaline lakes (> 50 g L⁻¹) and will expand the boundaries of the model by adding high-salinity samples in the future. The K&S model was designed for seawater, and its application to complex ionic lakes (e.g., Lake Chagannaoer) introduces about 5 % uncertainty (Fig. 6f7). This uncertainty can be reduced in the future by developing lake-specific dielectric constant models.

Salinity significantly increased (slope = 0.48 ppt yr⁻¹, p<0.01) in Lake Daihai (Fig. 12a), which was selected for driving force analysis to demonstrate the potential application of the dataset. The generalized linear model was used to quantify the relative contribution of each factor, and the correlation coefficient (r) was applied to examine the relationship between salinity and factors. The lake area accounts for 34.1 % of the variation in Daihai salinity, followed by POP (32.5 %) and TEMP (17.0 %), with other factors contributing less (Fig. 14a). Lake area (r= -0.89, p<0.01) and POP (r= -0.71, p<0.05) showed negative correlations with salinity, while TEMP (r = 0.69, p<0.05) and NTL (r = 0.95, p<0.01) exhibited positive correlations, and other variables were not significant (Fig. 14b). These indicate that the lake area closely related to the water volume is the dominant factor in the salinity change of Lake Daihai, while climatic warming will aggravate the salinization of inland lakes (Jeppesen et al., 2020). Overall, the dataset supports UN SDG 6.3 (improve water quality) by providing high-resolution salinity data for IMXL – an arid region with 30 % of China's saline lakes – enabling local policymakers to track salinization progress (e.g., Lake Daihai's ecological water replenishment with 2.57 × 10⁷ m³ in 2024 under SDG 6.6) (Liangcheng County, 2025).