We employed the ERA5 reanalysis dataset (i.e., the fifth generation of ECMWF (European Centre for Medium-Range Weather Forecasts) reanalysis) (Hersbach et al., 2023), which provides consistent land–atmosphere variables at hourly temporal resolution and 0.25 ° spatial resolution, based on the ECMWF coupled assimilation system. Here, we chose ERA5 SM as the training target since it currently provides one of the most accurate and spatially coherent large-scale SM datasets available (Soci et al., 2024). This choice is a well-established practice in previous developments of machine-learning SM retrievals (Aires et al., 2005; Rodríguez-Fernández et al., 2019; Aires et al., 2021b; Pellet et al., 2025).

The following ERA5 variables relevant to SM dynamics were considered in this study:

Soil moisture (SM) – Volumetric SM ( m 3 /m³) is available at multiple depths. We selected the 0–7 cm layer to match the penetration depth of ASCAT observations.

Soil temperature (ST) – ST is important for retrieving SM due to its thermal properties (Campbell, 1985), which indirectly relate ST and soil water content. We used the topmost layer (0–7 cm), which is most relevant to ASCAT observations.

Leaf area index (LAI) – LAI was included as a vegetation descriptor due to its impact on the backscatter coefficient in microwave remote sensing (Prigent et al., 2005; Petchiappan et al., 2022).

Antecedent precipitation evaporation index (APEI) – APEI is a well-established predictor of SM, representing the balance between precipitation and evaporation integrated over a decaying temporal window (Shaw et al., 1997; Han et al., 2023). It is defined as: A P E I t = ∑ i = 0 N k i p t − i − e t − i , where k is an empirical factor to indicate the decay effect from the rainfall, p t − i and e t − i are the precipitation and evaporation values respectively, at i t h day before the reference day t ; N is the maximum number of days considered prior to day t . Following Han et al. (2023), we adopted k = 0.91, N = 33, as this configuration demonstrated a strong correlation with ERA5 SM in the preliminary analysis (see Section 2.5). Although these values were presented as globally optimized parameters, they have been validated against a long-term in situ SM network across the United States from the United States Climate Reference Network (USCRN). We also conducted additional tests (not shown), exploring alternative parameter combinations, but did not observe any significant improvements. Therefore, we have decided to use the same combinations of k and N as in Han et al. (2023), which are considered appropriate for our study region.

Amplitude of the diurnal surface temperature cycle (dT) – The diurnal temperature amplitude is closely linked to soil thermal inertia and, consequently, to SM content (Prigent et al., 2005). Therefore, alongside the aforementioned variables, dT serves as an auxiliary predictor in our SM retrieval framework. We computed dT as the difference between the daily maximum and minimum hourly ERA5 surface temperatures.

The original hourly ERA5 variables were aggregated to a daily scale. Precipitation and evaporation were converted to cumulative daily values, while all other variables were averaged over a 24-h period.

ASCAT is a C-band radar onboard the Metop-A, B, and C satellites, providing near-global coverage every 12 h from a sun-synchronous orbit at 817 km altitude. We used the Metop ASCAT Climate Data Record (CDR, H120 v7), which offers consistent SM-related products at a resampled 12.5 km resolution (from native 25–34 km, depending on swath position) (H SAF, 2021a; H SAF, 2021b). Two ASCAT variables were employed:

Backscatter ( σ 40 ) – Normalized radar backscatter ( σ 40 , in dB) was adjusted to a 40 ° incidence angle. Backscatter is highly sensitive to soil’s dielectric properties and thus widely used in SM retrieval (Wagner et al., 1999; El Hajj et al., 2016).

ASCAT soil moisture (SM) – Relative SM (%) derived by rescaling σ 40 between dry and wet reference conditions (H SAF, 2021b). To ensure comparability with ERA5 and retrieval results, we converted ASCAT relative SM into volumetric units ( m 3 /m³) using soil porosity data from the Global Land Data Assimilation System [GLDAS; Rodell et al. (2004)], following the approach of Saxton and Rawls (2006). This variable is hereafter referred to as the H120 SM product.

Both σ 40 and H120 SM data were collocated with ERA5 variables by regridding to 0.25 ° resolution and aggregating to daily values. The analysis covers a 4-year period (2016–2019).

In situ SM measurements from the International Soil Moisture Network [ISMN; Dorigo et al. (2013); Dorigo et al. (2021)] were used to assess retrieval performance in 2019. The ISMN provides a globally accessible archive of harmonized SM measurements, established through international collaboration and supported by the European Space Agency (ESA) funding and community contributions. This database serves as an essential reference for validating the retrieval of SM products (Rodríguez-Fernández et al., 2015; Batchu et al., 2023). For the CONUS, we selected data from three main networks, including: (1) the Soil Climate Analysis Network [SCAN; Schaefer et al. (2007)], (2) the SNOwpack TELemetry network [SNOTEL; Leavesley et al. (2010)], and (3) the United States Climate Reference Network [USCRN; Bell et al. (2013)]. We used measurements from the top layer (0–7 cm) to match the depth sensitivity of ASCAT and ERA5. The measurements, originally reported at hourly intervals in volumetric units ( m 3 /m³), were aggregated into daily averages. Stations with more than 100 missing records were excluded.

To further evaluate how retrieval performance varies across different environmental conditions, we incorporated soil texture and land-cover information into our analysis. Soil texture classes at the surface level (0 cm) were obtained from the United States Department of Agriculture (USDA) soil texture classification system provided by Hengl (2018). Land-cover information was derived from the Annual National Land Cover Database (NLCD) Collection 1 Land Cover 2019 product (U.S. Geological Survey, 2024).

We extracted the local soil texture and dominant land-cover type associated with each ISMN site based on its geographic coordinates. These ancillary datasets allow us to stratify the evaluation of SM retrievals across major soil texture classes and land-cover categories for all selected in situ sites across CONUS during 2019.

To assess the spatial and temporal dependencies among variables (Sections 2.2, 2.1), we performed a sensitivity analysis using ERA5 and ASCAT data aggregated at the daily scale across all pixels in the study domain from 2016 to 2019. Geographic coordinates (latitude and longitude) were also included. Figure 1 summarizes the results as correlation matrices, reporting total, temporal, and spatial correlations. Here, the “spatial correlation” denotes the temporal average of spatial correlation coefficients across the domain, while the “temporal correlation” corresponds to the spatial average of temporal correlations computed at each grid point.

The analysis reveals that ASCAT σ 40 backscatter and APEI exhibit moderate positive total correlations with ERA5 SM (0.39 and 0.45, respectively), whereas ST and dT show moderate and strong negative correlations (−0.37 and −0.67). Longitude also correlates positively with SM (0.38), suggesting the presence of large-scale geographic gradients. These results highlight dT, APEI, σ 40 , and longitude as particularly influential predictors in terms of their overall and spatial association with SM. On the other hand, temporal correlations further emphasize the relevance of ST (−0.44) and LAI (−0.30), underscoring their role in capturing day-to-day SM variability. As these correlations aggregate all locations and seasons, they offer a general overview of variable interactions. Still, examining total, spatial, and temporal correlations provides a comprehensive understanding of variable dynamics and helps inform predictor selection. There is a trade-off between capturing temporal versus spatial dynamics, which should be balanced when designing retrieval models.

Importantly, correlations also reflect confounding influences. For example, σ 40 correlates with LAI (0.40), indicating sensitivity to vegetation as well as SM. This mixed signal complicates large-scale retrieval, reinforcing the need to incorporate auxiliary (e.g., ERA5) and ancillary (independent from ERA5) data to improve SM estimates (Aires et al., 2021a).

Finally, the link between auxiliary variables and geographic coordinates also offers insights into spatial patterns—for example, σ 40 and longitude (0.36), or ST and latitude (−0.49). Recognizing such spatial and temporal dependencies should help in better designing models and selecting inputs, taking into account the underlying correlations that could either support or limit retrieval. These impacts will be discussed further in the following sections.

NN retrieval models have been extensively applied in the field of remote sensing (Goïta et al., 1994; Atkinson and Tatnall, 1997; Maggiori et al., 2017). In this study, we investigated two widely used NN architectures—MLPs and CNNs—for the SM retrieval.

NNs are statistical models typically trained to minimize global errors over large datasets. While effective on average, their performance can degrade in regions with complex surface conditions—such as areas with dense vegetation or significant topographic variability—leading to localized regional biases. These biases are often not due to insufficient data samples but rather a lack of input features that adequately represent the local surface conditions.

To address this challenge, the concept of “localization” has been introduced (Boucher et al., 2023; Pellet et al., 2025). Localization refers to strategies that enable NNs to adapt their behavior to local conditions. This can be achieved in several ways: by augmenting input features (e.g., adding physical or geographical variables) or by modifying the model architecture to account for spatial heterogeneity. In the following, we discuss several localization strategies for the two types of NN architectures considered: MLPs and CNNs.

Various model architectures can be exploited to retrieve SM.

Standard MLP: A simple MLP model consists of a single hidden layer with 10 neurons with sigmoid activation functions, followed by a linear output layer. The input can be limited to only σ 40 and/or incorporates auxiliary inputs, including physical variables and geographic coordinates. Different architectures are explored, denoted as MLP a , b , where a represents the number of physical variables and b the number of geographic variables.

Localized MLP (MLP-l): A pixel-specific MLP where, instead of pooling all pixels together, a separate MLP is trained for each individual pixel, allowing highly localized predictions. We chose to test here MLP 3,0 -l with three physical inputs: σ 40 , ST, and LAI, without including any geographic input, as a single model for each pixel.

Standard CNN: The proposed CNN consists of multiple convolutional layers, followed by a ReLU activation function. The architecture employs dilated convolutions to expand the receptive field without reducing spatial resolution. The network begins with a standard convolutional layer of 5 × 5 × 16 kernels to extract low-level spatial features, followed by three dilated convolutional layers ( 5 × 5 × D, where D corresponds to 32, 64, and 32 kernels) with rates of 1, 2, and 1, respectively, to capture broader contextual information. An additional standard convolutional layer ( 5 × 5 × 16 kernels) is included before the final output layer. Batch normalization is applied after each layer to enhance convergence, and dropout (30%) is introduced to mitigate overfitting. “Same” padding is considered here to preserve input dimensions. The input consists of a 2D image of σ 40 and/or additional auxiliary inputs. Similar to the standard MLP, the notation CNN a , b indicates a CNN with a physical variables and b geographic variables (although the latter case is not really meaningful).

Localized CNN (CNN-l): A localized variant that considers a locally connected convolutional layer, allowing the filter to learn unique weights for different spatial locations. The architecture includes a zero-padding to preserve the spatial dimensions, followed by a single locally connected 5 × 5 layer to capture local-specific patterns, without weight sharing. We considered in this case a CNN 3,0 -l model with three inputs, including three images of σ 40 , ST, and LAI.

All models use consistent hyperparameters: Adam optimizer (with an initial learning rate of 0.001), Glorot Uniform initialization with a fixed seed, and early stopping after 10 epochs of no validation loss improvement (maximum of 200 epochs), restoring best weights. The only variation is in the batch size, which depends on the model type: MLPs use a batch size of 64, while CNNs use 16. For the localized models, MLP-l uses a batch size of 16, whereas CNN-l uses 8. Since the dataset covers the entire CONUS domain, computationally intensive optimization strategies (e.g., Bayesian tuning) were infeasible, consistent with observations in previous large-sample remote sensing studies (Rabiei et al., 2025a). As a result, a structured manual tuning strategy is adopted here: key hyperparameters were first selected within standard ranges reported in the literature (e.g., Pellet et al., 2025; Rabiei et al., 2025a) and then iteratively refined based on validation performance. Further adjustments around the final configurations produced only marginal improvements, indicating that the selected architecture is adequate for capturing SM retrieval accuracy.

In all the experiments, the overall dataset was split sequentially rather than spatially: the 3-year 2016–2018 dataset was divided into 80% for training and 20% for validation, while 2019 was reserved for testing and performance evaluation. This approach ensures that the model is evaluated on a truly unseen period, thus providing a realistic assessment of generalization ability. Other temporal combinations (e.g., using 2016 as the test year) are also possible, provided the test set remains independent of the training period. A Monte Carlo-like test using four different test years, followed by averaging the test rates for each year, will be evaluated in upcoming work.

The ascending and descending satellite orbits were trained separately.

To evaluate the performance of various SM retrieval models, we used several statistical evaluation metrics, including: Pearson’s correlation coefficient (r, unitless), root mean square error (RMSE, m 3 /m³), and bias (Bias, m 3 /m³). The standard deviation (Std, m 3 /m³) of the difference between two datasets is also reported.

In addition, we also evaluated the dampening and inflating effects (Boucher and Aires, 2023) by comparing the range of values in the predicted SM dataset ( R a n g e pred ) with that of the reference ERA5 dataset ( R a n g e ref ) . A dampening effect occurs when R a n g e pred < R a n g e ref , indicating a reduced range in the predictions. Conversely, an inflating effect is observed when R a n g e pred > R a n g e ref . These effects are quantified as: E % = 100 × R a n g e pred − R a n g e ref R a n g e ref .

Negative values indicate dampening, while positive values indicate inflation. The range quantifies the difference between the average SM of the upper and lower 10% quantiles at each pixel (Boucher and Aires, 2023).

As described in Section 3.3, we investigated various model configurations based on MLP and CNN architectures, incorporating different combinations of physical and geographical input variables for both MLPs and CNNs. These configurations were designed to assess the models’ sensitivity to varying degrees of localization and input diversity. Table 1 summarizes the evaluation metrics for all MLP and CNN retrieval models, applied to the 2019 testing dataset, covering eight scenarios for each model type. Detailed results, including spatial maps and time series plots, are presented in the following for selected representative models.

The performance of MLP models improves with the addition of more auxiliary features. For example, as shown in Table 1, the addition of ST and LAI data results in a significant increase in total correlation, rising from 0.37 in MLP 1,0 —which uses only the ASCAT σ 40 backscatter as an input—to 0.59 in MLP 3,0 . Corresponding improvements are also observed in temporal correlation (increasing from 0.24 to 0.61) and spatial correlation (increasing from 0.41 to 0.55). Overall, model performance continues to enhance with the inclusion of additional features such as APEI, dT, and geographic coordinates. This trend is further illustrated in Figure 2, which shows SM retrievals from different models for the ascending orbit on 20 January 2019. Although MLP 1,0 captures some SM patterns in the southeastern part, MLP 3,0 yields improved retrievals that, while not perfect, are closer to the ERA5 reference data. This enhancement can be attributed to the MLP model’s ability to treat all input features as independent, thereby effectively integrating additional information.

On the other hand, CNNs exhibit a different behavior. While performance improves from CNN 1,0 to CNN 1,2 or CNN 3,0 , the model achieves a high total correlation of 0.91 using only three physical inputs (e.g., σ 40 , ST, LAI), suggesting it was already near its capacity. Adding more features results in only marginal gains and, in some instances, slight performance degradation. This phenomenon can be attributed to the architectural characteristics of CNNs, which are designed to extract spatial hierarchies and local patterns through convolutional layers. As shown in Figure 2, even a single σ 40 input allows CNN 1,0 to effectively capture meaningful SM patterns. For example, it identifies very dry conditions (indicated in red) over the Rocky Mountains and the southwestern part, while showing wetter conditions (in blue) across the southeastern region–these patterns closely resemble the ERA5 reference data for 20 January 2019. Therefore, the CNN 3,0 model is recommended here.

CNN consistently outperforms the equivalent MLP models, as evidenced by the statistical metrics reported in Table 1. This superiority is further illustrated in Figure 3, which compares the spatial correlation scores of various MLP and CNN models. Across all configurations, CNNs demonstrate higher spatial correlation than classical MLP models. This improvement can be attributed to the CNNs’ ability to exploit spatial patterns, which is particularly advantageous when such patterns are strongly expressed in the target variable, such as SM. However, both model types still have difficulty capturing abrupt events (i.e., sudden SM peaks). For example, during a pronounced wetting event in the eastern United States around late May 2019, all models showed a reduced spatial correlation compared to the surrounding periods (see Figure 3).

It is also noted that when increasing the localization: MLP 3,0 → MLP 3,2 → MLP 5,0 → MLP 5,2 the total correlation increases (from 0.59 to 0.74) at the detriment to the temporal correlation (from 0.61 to 0.53). This means that the localization (by adding geographic or auxiliary features) allows to constrain better the overall SM pattern, but the SM dynamics are degraded locally. This is a true issue because even if we need a correct spatial pattern of SM, the SM dynamics are an essential piece of information, for instance, for assimilation. There is clearly a balance that needs to be done between a better spatial pattern, reducing local biases, and a satisfactory local dynamics. We strongly suggest that any new product should be evaluated using these two diagnostics.

We examined a simplified case where only geographic coordinates (latitude and longitude) were used as input variables. Although not practical for real-world SM retrieval, this setup highlights the role of spatial information in model performance. The results showed non-negligible total correlations: just latitude and longitude information provides a total correlation of 0.44 for MLP 0,2 . Incorporating coordinate information significantly enhances spatial correlation, as demonstrated by an increase from 0.37 in the MLP 1,0 model (model with only σ 40 as an input) to 0.51 in the MLP 1,2 model (MLP with σ 40 , latitude, and longitude as inputs). In contrast, the temporal correlation does not change and is equal to 0.24. This addition of the geographic features only improves the spatial correlation, reducing the local biases, without increasing the temporal correlations. This is natural, as latitude and longitude are static.

On the other hand, substituting physical variables (e.g., in the MLP 3,0 model with inputs including σ 40 , ST, and LAI) instead of using coordinate information (as done in the MLP 1,2 model with three inputs: σ 40 , latitude, and longitude) enhances both temporal and spatial correlations. These physical features carry inherent spatiotemporal information, allowing the model to better capture variations across both dimensions. This shows that physical variables are more SM information-carrying features than just latitude and longitude.

While coordinate information benefits MLPs, it has minimal impact on CNNs. For example, in the CNN 1,2 model, which incorporates σ 40 , latitude, and longitude as inputs, all correlation metrics (total, temporal, and spatial) increase by less than 0.02 compared to a standard CNN model that uses only σ 40 as an input feature. This limited improvement occurs because the grid position of the pixels in the input image is exploited directly by the CNN as latitude and longitude information when trained on a fixed domain as CONUS. We will not comment further on the CNN _ , 2 configurations.

It is not surprising that the two more extreme localization cases— MLP 3,0 -l and CNN 3,0 -l—outperform all other experiments, as presented in Table 1 with high total correlations of 0.92 and 0.93, respectively. An example of Figure 2 also shows that the MLP 3,0 -l and CNN 3,0 -l retrieved SM patterns more closely align with the ERA5 target for the ascending orbit on 20 January 2019, compared to other classical models. Both architectures demonstrate similar spatial and temporal performance, as illustrated in Figures 3, 4, respectively, and they far surpass classical NN models such as MLP 3,0 and CNN 3,0 . The local bias in these localized models is notably reduced, with near-zero bias errors and very low standard deviations observed in most regions of the study area (see Figure 4).

Despite these promising results, it is important to acknowledge the limitations of such localization strategies. By design, these models primarily learn localized statistical relationships, which may limit their ability to generalize or represent broader physical processes. Consequently, they may inadvertently propagate inherent biases in the target dataset. Furthermore, the reliability of the models is highly dependent on the quality of training data at each location; areas with sparse observations may experience instability, leading to issues such as overfitting or underfitting.

We further assessed the models’ ability to capture SM extremes, both dry and wet conditions, by comparing their retrievals of minimum and maximum values. Figure 5 shows the corresponding maps derived from several NN-based models alongside the ERA5 reference. The results from the ASCAT product (i.e., H120) are also presented for comparison purposes.

Both the non-localized models, MLP 3,0 and CNN 3,0 , tend to underestimate SM extremes. Notably, they struggle to represent dry conditions in the western region and high SM values in the eastern part, because these are anomalies compared to standard cases.

In contrast, the localized models, MLP 3,0 -l and CNN 3,0 -l, demonstrate improved estimates of extreme values. Localization enables the models to adapt more effectively to regional variability, allowing them to detect local anomalies with higher accuracy. The enhanced ability of localized models to capture extreme wet and dry events highlights their potential value for targeted applications such as drought monitoring or flood forecasting.

In general, statistical models (e.g., NNs) tend to dampen the range of values, as regression-based models typically capture a smaller portion of the total variability than the original data (Hastie et al., 2009; Skafte et al., 2019). However, under certain conditions—particularly when training data from different locations are pooled—regression models can also exhibit inflation (Boucher and Aires, 2023). Inflating effects are more commonly associated with non-statistical models.

As illustrated in the third column of Figure 5, the ASCAT H120 product shows a widespread inflating effect (depicted in green), largely due to an excessive number of pixels with a minimum SM of zero. It is important to note that ASCAT SM is not trained on ERA5; it is more “independent” than the NN-based results presented here. In contrast, our NN-based retrievals (MLPs and CNNs) rarely show inflation. These models mainly exhibit dampening (shown in magenta), consistent with their statistical nature. Among the NN-based models, the localized versions ( MLP 3,0 -l and CNN 3,0 -l) show a reduced dampening effect, as also seen in Table 1, indicating better characterization of SM extremes. By tailoring the model to regional characteristics, localization improves the representation of both minima and maxima, further enhancing retrieval realism.

Figure 6 presents examples of daily SM time series in 2019 from six randomly selected in situ sites, along with the correlation scores between model outputs and ground measurements. Among these six examples, the ASCAT H120 product demonstrates a relatively good correlation with the in situ data; however, it exhibits a pronounced bias in absolute values, as shown by the consistently offset range (orange lines) in Figure 6.

On the other hand, the other three retrievals—our two NN-based models and ERA5—show more reasonable SM ranges and better capture observed temporal dynamics. ERA5 demonstrates the highest agreement with ISMN data, with correlation coefficients exceeding 0.80 across all six sites, consistent with its assimilation of extensive in situ and satellite observations (Aires et al., 2005). CNN 3,0 also performs well, closely following the ERA5 time series. This is consistent with expectations, given that ERA5 SMs served as the training target for the models. Interestingly, at one of the six randomly selected sites (last panel of Figure 6), CNN 3,0 surpasses ERA5, with a correlation of 0.83 against the in situ data. In most cases, CNN 3,0 outperforms MLP 3,0 , further highlighting the advantages of convolutional architectures in capturing spatial and temporal patterns of SM.

Table 2 presents the statistical performance of four SM retrievals (H120 product, MLP 3,0 , CNN 3,0 , and ERA5) evaluated against in situ measurements from 638 sites during the 2019 validation period. To complement this tabular summary, Figure 7A illustrates the distributions or the probability density functions (PDFs) of three temporal metrics: correlation ( r tem ), bias ( Bias tem , m 3 /m³), and standard deviation ( Std tem , m 3 /m³).

Across 638 sites, the ACAST H120 SM product demonstrates a moderate median temporal correlation ( r tem = 0.61), which is comparable to the other three evaluated SM products. Although it achieves the lowest median bias ( Bias tem = 0.024 m 3 /m³) among all products, its standard deviation is relatively higher ( Std tem = 0.080 m 3 /m³), suggesting greater variability compared to the alternatives. Additionally, Figure 7A shows that the H120 product exhibits a longer negative tail in its bias distribution, and its standard deviation values extend beyond the typical range observed in the other retrieval methods.

In contrast, ERA5 performs best among all retrievals, with the highest median correlation ( r tem = 0.75) and low errors ( RMSE tem = 0.106, Bias tem = 0.061, Std tem = 0.060 m 3 /m³). This superior performance is attributed to its reanalysis nature, which combines model data with observations, thereby enhancing consistency with ground-based measurements (Aires et al., 2005). Our NN-based retrievals, MLP 3,0 and CNN 3,0 , also show competitive performance. In particular, CNN 3,0 achieves a temporal correlation of 0.68, closely approaching ERA5, while MLP 3,0 records a lower correlation of 0.60. Notably, the bias and standard deviation distributions of CNN 3,0 retrieval closely resemble those of ERA5 (as presented in Figure 7a), indicating that it provides more stable and reliable predictions than the other ASCAT-derived products.

Figure 7b provides spatial insights by mapping the temporal correlation ( r tem ) of MLP 3,0 with the in situ measurements over 638 considered sites. Lower correlation values, indicated by yellow markers, are primarily found in the Rocky Mountains. A corresponding improvement map highlights the change in temporal correlation between CNN 3,0 and MLP 3,0 ( Δ r tem = r tem ( CNN 3,0 ) - r tem ( MLP 3,0 )). The observed increases in correlation (shown in red) over the mountainous areas underscore the ability of CNN architectures to better capture spatial variability, outperforming MLPs in complex environmental conditions, such as those presented by varied topography and heterogeneous mountain landscapes.

To further investigate environmental controls on model performance, we evaluated SM retrievals within individual soil texture groups and land-cover classes against in situ SM measurements. Figures 8, 9 present the comparison metrics for the major soil texture classes and dominant land-cover types, respectively. Overall, ASCAT H120 exhibits smaller biases but with higher standard deviations than both NN-based retrievals and ERA5 across most soil texture and land-cover categories, likely due to the retrieval’s calibration to local conditions (Aires et al., 2021b). Between two NN models, CNN 3,0 consistently outperforms MLP 3,0 , showing higher correlations and lower RMSE, bias, and standard deviation across all classes. Across soil texture groups (Figure 8), CNN 3,0 , although trained on ERA5, demonstrates better performance compared to ERA5, particularly for clay loam and silty clay loam in terms of correlation, and for most texture classes in terms of RMSE. Across land-cover types (Figure 9), CNN 3,0 also outperforms ERA5 in correlation over mixed forest, and in RMSE over evergreen and deciduous forests, cropland, grassland, pasture, and developed areas.