The Xilinhot National Climate Observatory is located in the central Xilingol Grassland of Inner Mongolia, China (43.95°N, 116.12°E; elevation: 1,124 m), and serves as one of the national benchmark climate stations (Figure 1). The Xilingol region experiences a typical temperate continental climate with limited precipitation, with an annual rainfall of approximately 150–400 mm and annual evaporation ranging from 1,500 to 2,700 mm (Ye et al., 2021), indicating a pronounced water deficit characteristic of semi-arid regions. With long-term and continuous observational records, the station is representative of the climatic features of the typical temperate grassland ecosystem. The station is equipped with comprehensive observation systems, providing essential data support for regional climate change research and climate services. Grassland ecosystems are highly sensitive to soil moisture variability, as soil moisture directly affects vegetation growth, evapotranspiration, and overall ecosystem functioning (Su et al., 2020). Therefore, studying the calibration of soil moisture measurements from automatic stations at the Xilinhot National Climate Observatory is of significant importance.

Manual ground measured soil moisture data from the Xilinhot National Climate Observatory, covering the period from 18 May 2019 to 21 August 2025, were used to validate the soil moisture observations from automatic meteorological stations. Soil samples were collected at five depth intervals (0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm). The gravimetric method was employed to determine the soil gravimetric water content at each depth. Manual ground soil moisture data were collected annually from March 28 to October 28, with daily sampling performed once on the 8th, 18th, and 28th of each month.

The core device of the automatic soil moisture monitoring station is the DZN2 automatic soil moisture sensor (hereafter referred to as DZN2). Based on the FDR principle, the sensor measures changes in the frequency of electromagnetic waves emitted by the probe as they pass through media with different dielectric constants (Skierucha and Wilczek, 2010; Rasheed et al., 2022), and calculates the soil volumetric water content (VWC) using a mathematical model. The instrument features high resolution, up to 0.001 m³/m³, and measurement accuracy of ±0.02 m³/m³ VWC, covering a full range from 0 to soil saturation. It supports depth-profile measurements with 10 cm intervals and can monitor up to 16 soil layers. This system can be used to establish soil moisture monitoring networks, enabling real-time monitoring and dynamic analysis of soil moisture across different sites. In this study, the VWC data collected by the DZN2 span from 16 April 2020, to 23 August 2025, covering five soil layers at depths of 0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm, with measurements recorded daily.

The CRS-2000C regional soil moisture measurement system (hereafter referred to as CRS-2000C) is a professional instrument designed for mesoscale soil moisture monitoring, with its core measurement device being the COSMOS-001. The system indirectly estimates soil water content by detecting the concentration of fast neutrons produced near the ground during the moderation of cosmic rays. It can measure soil moisture to a maximum depth of approximately 70 cm, covering the full range from 0 to saturation. The system achieves a relative soil moisture measurement accuracy of ±0.03 m³/m³ and enables continuous and stable monitoring of soil moisture over large spatial scales. The VWC data obtained from this system span from 22 July 2022 to 23 August 2025. At the Xilinhot station, measurements were taken for the 0–50 cm soil layer, with a sampling interval of 30 min.

The soil temperature and moisture monitoring system (hereafter referred to as 5TM) primarily employs the 5TM sensor (Decagon Devices, United States) to measure soil VWC at depths of 0–50 cm. The 5TM sensor utilizes capacitance/frequency-domain technology to determine the soil dielectric constant, enabling accurate estimation of soil moisture. Operating at a frequency of 70 MHz, the sensor minimizes the influence of soil texture and salinity, ensuring reliable measurements across a wide range of soil types. It provides high measurement accuracy, with VWC precision of ±0.0008 m3/m3 within the 0–0.5 m3/m3 VWC range, and supports a measurement range of 0–1 m3/m3 VWC. The sensor also responds rapidly to dynamic changes in soil moisture, making it suitable for long-term, continuous monitoring. Additionally, the 5TM sensor features a compact and robust design, facilitating long-term field deployment and efficient data acquisition. The VWC data used in this study were collected by the 5TM from 18 May 2019, to 27 October 2024, at three soil layers: 0–10 cm, 10–20 cm, and 20–40 cm, with measurements recorded daily. The photographs of soil moisture monitoring instruments are shown in Figure 2.

Daily meteorological data from the Xilinhot station, including near-surface air temperature, wind speed, relative humidity, evaporation, and precipitation, were obtained for 1 May 2019 to 21 August 2025. The data were automatically recorded by a DZZ4 automatic weather station.

This study utilized the Google Earth Engine (GEE) platform to acquire the Normalized Difference Vegetation Index (NDVI) data from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors. Specifically, we used the 16-day composite MOD13Q1 product, which provides NDVI at a spatial resolution of 250 m and a temporal resolution of 16 days (Huete et al., 2002). NDVI values corresponding to the locations of the meteorological stations were extracted for the period from 1 May 2019, to 21 August 2025.

To ensure the reliability and usability of the soil moisture data, this study implemented strict quality control measures (González-Rouco et al., 2001; Hubbard et al., 2005; Dorigo et al., 2013). First, a reasonable threshold range was set, limiting the VWC to 0 m3/m3–0.6 m3/m3 to remove clearly erroneous or invalid values. Second, based on meteorological principles, soil moisture variations were compared with precipitation patterns to maintain temporal consistency; that is, soil moisture should not exhibit significant fluctuations during periods without precipitation, and obvious outliers were manually removed. Data showing no variation for 10 consecutive days were also considered invalid. Finally, to further identify and eliminate outliers, the 3σ rule was applied for outlier detection and removal, thereby maximizing the accuracy and representativeness of the data.

Moreover, to ensure the reliability of the NDVI data, quality control was performed on the MOD13Q1 NDVI product in this study. Specifically, MOD13Q1 provides pixel-level quality information (SummaryQA) to indicate data reliability. During processing, pixels affected by clouds, cloud shadows, aerosol contamination, or low-quality observations were removed, thereby retaining only high-quality pixels (SummaryQA = 0) for direct use. In addition, inspection confirmed that there were no missing NDVI values at the meteorological station locations, and thus no interpolation was required.

The accuracy of the automatic station soil moisture data was evaluated based on the manually observed soil moisture measurements. Ideally, manual ground and automatic soil moisture measurements should be matched at the same location, time, and depth. However, in practice, the manual ground sampling sites and the automatic station were not exactly at the same location, but both were situated within the same grassland area, where soil moisture conditions can be considered comparable. Since the automatic station records fewer soil layers, the manual ground observed multi-layer data were averaged to correspond to the station’s measurement depths. It should be noted that the manual ground measurements were originally recorded as gravimetric soil water content. To match the automatic station data, they were converted to volumetric water content using the corresponding soil bulk density for each layer. The formula is shown in Equation 1: θ v = θ g × ρ b (1) where θ_v is the volumetric water content (m3/m3), θ_g is the gravimetric water content (kg/kg), ρ_b is the soil bulk density (g/cm³).

For comparative analysis, the manual ground measured volumetric soil moisture was matched with the automatic station measurements at the same dates and soil layer depths, and line charts were plotted accordingly. The temporal variation patterns of both datasets were then compared.

Based on the paired dataset, the accuracy of volumetric soil water content measurements from the automatic stations was evaluated using the determination coefficient (R ²), mean absolute error (MAE), root mean square error (RMSE), and mean bias (MB). The corresponding formulas are shown in Equations 2–5: R 2 = 1 − ∑ i = 1 n y i − y ^ i 2 ∑ i = 1 n y i − y ¯ 2 (2) M A E = 1 n ∑ i = 1 n y i − x i (3) R M S E = 1 n ∑ i = 1 n y i − x i 2 (4) M B = 1 n ∑ i = 1 n y i − x i (5) where x_i is the manual measured volumetric soil water content, y_i is the automatic station measurements, x̄ and ȳ are their average values, and n is the number of paired samples.

The VWC recorded by automatic soil moisture stations was corrected against manual ground measurements. VWC is often influenced by surrounding environmental conditions such as vegetation cover, soil texture, and microclimatic variability. Moreover, its relationship with near-surface meteorological factors (e.g., mean air temperature, precipitation, wind speed, and evapotranspiration) is inherently complex and nonlinear. Machine learning models can automatically capture these nonlinear relationships, thereby producing predictions that more closely reflect actual ground measurements. By incorporating auxiliary variables such as vegetation indices and meteorological data, the models are also able to account for systematic biases in automatic station measurements arising from local environmental heterogeneity. Different algorithms exhibit distinct strengths in capturing feature interactions and error patterns; therefore, to improve the accuracy of the automatic station measurements, multiple machine learning methods were employed for model fitting, including Cubist, Random Forest, XGBoost, CatBoost, and further developed a linear ensemble model based on these approaches.

The Random Forest algorithm, proposed by Breiman, is an ensemble learning method that employs decision trees as base learners (Breiman, 2001; Salman et al., 2024). It builds upon the Bagging framework by introducing random feature selection during the construction of each decision tree. The final classification or regression outcome is determined through majority voting or averaging across all trees. Random Forest is characterized by its simplicity, robustness, and reduced tendency to overfit.

The Cubist algorithm is a rule-based regression model developed as an extension of Quinlan’s M5 model tree. At each leaf node of the Cubist decision tree, a multivariate linear regressionmodel is fitted to the subset of data covered by the corresponding rule set (Chen et al., 2020). By combining decision trees with local linear regression, Cubist retains the interpretability of rule-based partitions while enhancing local fitting capability, making it particularly suitable for small-to medium-sized datasets.

XGBoost, proposed by Chen, is a gradient boosted decision tree (GBDT) algorithm (Chen and Guestrin, 2016). It builds an ensemble of decision trees sequentially, with each new tree fitted to the residuals of the previous trees, thereby progressively improving predictive accuracy. The algorithm is computationally efficient and can effectively handle large-scale datasets.

CatBoost, developed by Yandex, is another gradient boosting decision tree algorithm (Dorogush et al., 2018). It employs symmetric trees as the base structure and uses feature-splitting preprocessing to handle categorical data. In combination with an ordered boosting strategy, it effectively mitigates gradient bias and prediction shift issues common in gradient boosting (Cai et al., 2024). CatBoost requires fewer hyperparameters, is user-friendly, and is well suited for complex datasets while reducing the risk of overfitting (Fu et al., 2024)^.

To further improve predictive accuracy and fully leverage the strengths of individual models, this study developed an ensemble model based on a generalized additive model (GAM) (Hastie and Tibshirani, 1995). The GAM flexibly characterizes the nonlinear relationships between input features and the response variable through smooth functions. Specifically, the predictions of the four base models were used as new input features to train a second-level GAM, which learns the optimal combination of the base model outputs. Compared with individual models, this ensemble approach achieves a secondary optimization of the prediction results, effectively reducing both bias and variance and significantly improving the robustness and overall accuracy of the predictions (Ganaie et al., 2022; Mienye and Sun, 2022)^.

The appropriate selection of variables plays a critical role in ensuring the accuracy and stability of the model. In this study, meteorological, vegetation, and soil-related variables were selected as inputs for the automatic soil moisture correction model according to physical considerations and previous studies, including near-surface air temperature, surface temperature, wind speed, relative humidity, evapotranspiration, precipitation, atmospheric pressure, NDVI, and the soil moisture of adjacent upper layers as model input variables (Hide, 1954; Seneviratne et al., 2010; Trenberth, 2011; Vicente-Serrano et al., 2010; Wang et al., 2007; Zhang et al., 2018; Froidevaux et al., 2014; Wu et al., 2025; Elmotawakkil et al., 2025). These variables characterize atmospheric forcing, vegetation conditions, and soil state that jointly influence soil moisture variability, and were used to construct the model training dataset to develop four base models.

Precipitation events not only directly replenish soil moisture but also exert a sustained influence on VWC over the following days, resulting in a clear lagged response in soil moisture dynamics. Similarly, evaporation continuously depletes soil water, producing a cumulative effect on VWC in subsequent days. Therefore, considering only the precipitation or evaporation on a single day may fail to accurately capture the true soil moisture dynamics when analyzing the water balance and its driving factors. To address this, we introduced a time-weighting approach that accounts for the effects of precipitation and evaporation over the preceding 10 days on the current-day VWC. A linearly decreasing weighting scheme was applied (Equations 6, 7), giving progressively greater weight to more recent days so that their influence on soil moisture is emphasized. This approach allows a more comprehensive characterization of soil moisture responses to water inputs and losses, effectively capturing the delayed effects of water replenishment and depletion and improving the model’s ability to simulate the spatiotemporal dynamics of soil moisture. X t * = ∑ i = 1 10 w i · X t − i (6) where X t * is the weighted cumulative precipitation (or evaporation) over the preceding 10 days for day t; X t − i is the daily precipitation (or evaporation) on the ith day prior to day t; w i is the weight coefficient assigned to the ith day, calculated as follows: w i = 11 − i ∑ j = 1 10 11 − j (7)

During the modeling process, this study employed RandomizedSearchCV (Bergstra and Bengio, 2012) to optimize the hyperparameters of four base learners (Random Forest, Cubist, XGBoost, and CatBoost) with the aim of achieving optimal model performance. Parameter search spaces were constructed according to the structural characteristics of each model. For the Cubist model, the key parameters included neighbors, n_rules, and n_committees. The Random Forest model primarily tuned n_estimators, max_depth, min_samples_split, min_samples_leaf, and max_features. The XGBoost model focused on n_estimators, max_depth, learning_rate, subsample, colsample_bytree, and min_child_weight. The CatBoost model optimized core parameters such as iterations, depth, and learning_rate. All parameter search spaces were explored using RandomizedSearchCV with random sampling and five-fold cross-validation to identify the optimal parameter combinations. The detailed search ranges are presented in Table 1.

This study used manually measured soil moisture as the dependent variable and constructed the training dataset with independent variables including the automatic station measurements for the same layer, near-surface air temperature, wind speed, relative humidity, the weighted cumulative precipitation over the past 10 days, the weighted cumulative evaporation over the past 10 days, the calibrated soil moisture from the adjacent upper layer (for deeper layers), and the NDVI index. Based on this training dataset, base models were constructed using machine learning algorithms such as RF, Cubist, XGBoost, and CatBoost, which were then integrated using a GAM model. Five-fold cross-validation was employed to evaluate model performance. Specifically, the entire dataset was randomly divided into five subsets. In each iteration, one subset was used as the validation set (without replacement), while the remaining four subsets served as the training set, and this process was repeated five times. Model performance was assessed using four metrics: the R2, MAE, RMSE, and MB.

Figure 3 present the comparison between manual measurements and the DZN2, the CRS-2000C, and the 5TM sensor. Overall, the sensor data follow similar trends as the manual ground measurements, although they exhibit fluctuations around the ground values, showing a certain degree of deviation.

Figure 4 show scatter plots between manual ground observations and the DZN2 (a-e), CRS-2000C (f), and 5TM sensor (g-i), respectively. For the DZN2, the R2 values in the 0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm soil layers were 0.418, 0.313, 0.218, 0.148, and 0.179, respectively, with a max–min difference of 0.270. The corresponding MAE values were 0.053, 0.057, 0.056, 0.058, and 0.062 m3/m3, with a max–min difference of 0.009 m3/m3. The RMSE values were 0.068, 0.069, 0.070, 0.075, and 0.079, with a max–min difference of 0.011 m3/m3. The MB values were −0.026, −0.023, 0.013, 0.022, and 0.037 m3/m3, with a max–min difference of 0.063 m3/m3. These results indicate that the agreement between DZN2 and manual ground observations was highest at the 0–10 cm depth, but declined notably with increasing soil depth. Moreover, automatic stations tended to underestimate soil water content in the shallow layers while overestimating it in the deeper layers.

For the 0–50 cm soil layer, CRS-2000C measurements had a R2 of 0.774, a MAE of 0.027 m3/m3, a RMSE of 0.033 m3/m3, and a MB of 0.008 m3/m3 compared with manual ground observations. These results indicate that the CRS-2000C slightly overestimated VWC, particularly under higher soil moisture conditions, but achieved the highest measurement accuracy among the three systems.

For the 0–10 cm, 10–20 cm, and 20–40 cm soil layers, the 5TM sensor showed R2 of 0.673, 0.535, and 0.200, respectively, when compared with manual ground measurements. The MAE were 0.039, 0.032, and 0.076 m3/m3, the RMSE were 0.048, 0.041, and 0.099 m3/m3, and the MB were −0.027, −0.013, and 0.072 m3/m3, respectively. The 5TM sensor exhibited higher accuracy in shallow soil layers and lower accuracy in deeper layers. The measurement bias varied with depth: the system slightly underestimated VWC in the g and h layers, whereas it overestimated VWC in the 20–40 cm soil layer.

Figure 5 present scatter plots comparing the corrected measurements from the DZN2 (a-e), CRS-200C (f), and 5TM (g-i) systems with the manual ground observations. As shown in Figure 5, after correction, the VWC measured by the DZN2 across the 0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm soil layers exhibits R2 values of 0.683, 0.507, 0.382, 0.437, and 0.294, respectively. The corresponding MAE values are 0.031, 0.027, 0.024, 0.020, and 0.018 m3/m3, while the RMSE values are 0.037, 0.036, 0.031, 0.026, and 0.024 m3/m3. The MB of all soil layers is close to zero. Overall, the correlations for all five layers improved substantially, with the largest increase in R2 reaching 0.289 (30–40 cm soil layer), indicating that the correction markedly enhanced the consistency between DZN2 and manual ground observations. Compared with the uncorrected data, the error metrics also decreased significantly. The maximum reduction in MAE reached 0.044 m3/m3 (40–50 cm soil layer), while the maximum decrease in RMSE reached 0.055 m3/m3 (40–50 cm soil layer). These results demonstrate that the correction effectively improved the measurement accuracy of the DZN2 across all soil depths, with particularly notable improvements in the deeper layers.

As shown in Figure 5, the corrected R² of the CRS-2000C increased to 0.849, representing an improvement of 0.075 compared with the uncorrected results. Meanwhile, the MAE and RMSE were reduced to 0.027 m³/m³ and 0.033 m³/m³, decreasing by 0.0143 m³/m³ and 0.0167 m³/m³, respectively, with the MB approaching zero. The These results indicate that the correction method significantly improved the measurement accuracy of the CRS-2000C and enhanced its agreement with manual ground observations.

As shown in Figures 5g–i, after correction, the 5TM achieved R² values of 0.800, 0.683, and 0.447 for the for the 0–10 cm, 10–20 cm, and 20–40 cm soil layers, respectively, representing improvements of 0.127, 0.148, and 0.247 compared with the uncorrected results. The corresponding MAE were 0.022, 0.020, and 0.019 m³/m³, decreased by 0.0167, 0.0121, and 0.0572 m³/m³, respectively. The RMSE values were reduced to 0.030, 0.028, and 0.024 m³/m³, corresponding to reductions of 0.018, 0.013, and 0.075 m³/m³, respectively. Overall, the correction substantially improved the consistency between 5TM measurements and manual ^ground observations, with notably greater enhancements in the deeper soil layer.