Abstract
This study investigates the intercomparison of daily soil moisture (SM) retrieval from ASCAT (Advanced SCATterometer) observations using machine learning. The exploitation of spatial structure through convolutional neural networks (CNNs) is shown to significantly enhance retrieval performance compared to a standard multilayer perceptron (MLP), with spatial correlation with the target ERA5 SM increasing from 0.55 to 0.91 and temporal correlation from 0.61 to 0.73. Incorporating “localization” (i.e., a strategy to adjust the neural network (NN) behavior to local conditions) into the model is a key factor for improving retrieval quality, resulting in more accurate SM estimates, reduced regional biases, improved temporal dynamics, and more realistic representations of extreme SM events. Our NN-based retrievals show strong agreement with in situ SM measurements, achieving temporal correlations of 0.60 and 0.68 for the MLP and CNN models, respectively, in the contiguous United States (CONUS) during 2019. These findings underscore the critical role of spatial learning and localization in SM retrieval from remote sensing data such as ASCAT.
1 Introduction
Soil moisture (SM) is a vital component of the Earth system, regulating the exchange of water and energy between the land surface and the atmosphere. It serves as a critical factor in numerous applications, including agricultural monitoring, drought and flood forecasting, water resource management, and climate prediction (Seneviratne et al., 2010; Ochsner et al., 2013; Vereecken et al., 2015). Recognizing its importance, extensive efforts have been made to obtain SM estimates across different temporal and spatial scales. These methods include ground-based in situ measurements (), satellite-based retrievals (; ), and land surface model simulations (). While each method has its strengths, they also face limitations related to coverage, representativeness, and accuracy.
In the last two decades, data-driven methods, particularly neural networks (NNs), have shown strong potential for SM retrieval (; ; Rodríguez-Fernández et al., 2015; ). Multilayer perceptrons (MLPs) provide a robust and interpretable pixel-scale baseline, aligning with current operational retrieval frameworks (Rodríguez-Fernández et al., 2019; ). More advanced convolutional neural networks (CNNs) extend these models by exploiting the satellite data at the image scale. Unlike physically based inversion methods, these approaches are purely data-driven: they learn statistical relationships between remote sensing observations and soil moisture states, enabling flexible, scalable, and potentially more accurate retrievals (Zhang et al., 2025). However, their performance is highly dependent on the choice of input data. The inclusion of relevant predictor variables, whether physical (e.g., brightness temperature, vegetation indices, land surface temperature) or geographical (e.g., land cover, soil texture, topography), directly influences retrieval accuracy and generalization.
Previous retrieval studies have improved physical realism by integrating auxiliary variables to constrain the radiative transfer models better. For instance, included surface emissivities along with microwave satellite observations to improve surface temperature retrieval, effectively accounting for local variations such as humidity and vegetation. Due to the lack of reliable, real-time emissivity products at that time, first-guess emissivities from climatologies were used to provide a seasonal and location-specific constraint. In another example, as SM influences various satellite observations (Prigent et al., 2005), early efforts such as developed global NN retrieval models that leveraged synergies across visible/infrared, active, and passive microwave sensors. This global approach has continued in later studies, such as in , and has evolved to include a broader range of inputs. More recently, integrated data from a diverse collection of datasets (i.e., synthetic aperture radar, Landsat, MODIS, land cover, and other sources), with a total of 62 input features for a global NN-based SM retrieval. The growing trend in NN-based SM retrievals has been toward using increasingly large and diverse input sets. While this may help capture complex environmental interactions, it can sometimes obscure the role of localized factors. To address this, some studies have begun to explicitly introduce geo-localization information to enhance localization in NN models (Mahara and Rishe, 2023; Singh and Gaurav, 2023; ). Incorporating both physical (i.e., dynamic) and geographical (i.e., static) inputs has thus become a common strategy.
In this study, we evaluate the performance of NN-based models for SM retrieval from the Advanced SCATterometer (ASCAT) observations, with a particular focus on the role of “localization”—–the incorporation of region-specific information to allow NN models to adapt more effectively to local conditions (; Pellet et al., 2025). We compare MLPs and CNN2 with varying levels of localization, ranging from simple models based only on physical or geographical inputs to highly localized approaches such as pixel-wise MLPs and locally connected CNNs. Other modern architectures, such as convolutional neural network-long short-term memory (ConvLSTM) models, which account for both spatial and temporal dependencies, have demonstrated strong performance in forecasting SM (Wang et al., 2024; Rabiei et al., 2025b; Rabiei et al., 2025a). However, as this study focuses on retrieval rather than forecasting, such models were excluded from our analysis. Nonetheless, temporal modeling could also offer potential benefits for retrieval. Further extensions will therefore investigate temporal architectures that are better suited for retrieval tasks than ConvLSTM, such as transformers or embedding-based methods. We assess here the ability of MLPs and CNNs to retrieve daily SM and capture extreme values (minimum and maximum) using a 2019 validation dataset. Model performance is evaluated against in situ measurements from the International Soil Moisture Network (ISMN; ; ) and compared with the ASCAT-derived SM product. The analysis is focused on the contiguous United States (CONUS; 125W–70W, 25N–50N), a region characterized by dense in situ measurements and pronounced spatial variability (). While CONUS provides an ideal testbed for assessing model skill, the methodology is inherently generalizable and applicable to other well-instrumented regions, such as Europe.
In the following, Section 2 describes the datasets employed in this study and presents preliminary information content analysis between the target SM and the other considered predictors. Section 3 details the NN-based architectures, localization strategies, and evaluation metrics. Section 4 assesses model performance at the daily scale and investigates the effects of localization. Section 5 focuses on extreme SM retrievals, comparing localized and non-localized NN-based models. Section 6 reports the evaluation against in situ measurements. Finally, Section 7 summarizes the findings, highlights practical implications, and discusses the potential of deep learning and localization approaches for advancing remote sensing–based SM retrievals.
2 Datasets
2.1 ERA5 database
We employed the ERA5 reanalysis dataset (i.e., the fifth generation of ECMWF (European Centre for Medium-Range Weather Forecasts) reanalysis) (), 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 (; Rodríguez-Fernández et al., 2019; ; Pellet et al., 2025).
The following ERA5 variables relevant to SM dynamics were considered in this study:
Soil moisture (SM) – Volumetric SM (/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 (), 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; ). It is defined as:where is an empirical factor to indicate the decay effect from the rainfall, and are the precipitation and evaporation values respectively, at day before the reference day ; is the maximum number of days considered prior to day . Following , we adopted = 0.91, = 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 and as in , 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.
2.2 ASCAT database
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) (; ). Two ASCAT variables were employed:
Backscatter – Normalized radar backscatter (, 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; ).
ASCAT soil moisture (SM) – Relative SM (%) derived by rescaling between dry and wet reference conditions (). To ensure comparability with ERA5 and retrieval results, we converted ASCAT relative SM into volumetric units (/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 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).
2.3 In situ data from the international soil moisture network
In situ SM measurements from the International Soil Moisture Network [ISMN; ; ] 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; ). 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; ], and (3) the United States Climate Reference Network [USCRN; ]. 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 (/m3), were aggregated into daily averages. Stations with more than 100 missing records were excluded.
2.4 Soil texture and land-cover data
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 . 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.
2.5 Preliminary information content analysis
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.
FIGURE 1
The analysis reveals that ASCAT 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, , 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, 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 ().
Finally, the link between auxiliary variables and geographic coordinates also offers insights into spatial patterns—for example, 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.
3 Retrieval methods and localization strategies
3.1 Neural network models
NN retrieval models have been extensively applied in the field of remote sensing (; ; Maggiori et al., 2017). In this study, we investigated two widely used NN architectures—MLPs and CNNs—for the SM retrieval.
3.1.1 Multilayer perceptrons
An MLP, also known as a feedforward NN, consists of multiple layers of neurons (perceptrons) organized sequentially (Rumelhart et al., 1986). Each neuron performs a weighted summation of its inputs, followed by the application of a nonlinear activation function. In a fully connected architecture, every neuron in a given layer is connected to all neurons in the preceding layer. The output of neuron is computed as: , where are the inputs to neuron , are the synaptic weights associated with each input, is the bias term, and denotes an activation function (). Typically, neurons in the hidden layers (i.e., all layers except input and output layers) use a sigmoid activation (f(x) = ), while neurons in the output layer employ the linear activation function (f(x) = x).
3.1.2 Convolutional neural networks
Unlike MLPs, which process inputs as flattened vectors (e.g., a single pixel or feature at a time), CNNs are specifically designed for grid-like data such as images. A CNN processes inputs as multidimensional tensors, e.g., where and denote the spatial dimensions (image height and width) and represents the number of channels (or input depth). In our case, = 100 and = 200 pixels covering the CONUS domain, while ranges from two to seven, including and other auxiliary variables.
As suggested by the term convolutional, CNNs consist of multiple convolutional layers that transform the input tensor into successively higher-level feature representations. Each convolutional layer applies a set of learnable kernels (or filters) across the input. As the kernel slides over the tensor, it computes localized weighted sums that collectively form a feature map. Mathematically, this can be expressed as:where denotes the convolutional kernel, represents the convolution operation, is the local input, b is the bias term, and is the output feature. Unlike element-wise multiplications in MLPs, CNN kernels exploit spatial locality, making them effective for capturing spatial patterns. Typically, multiple kernels are used in each layer to extract diverse features.
The behavior of a convolutional layer is influenced by several architectural parameters:
Padding extends the input tensor by adding extra pixels (commonly zeros) around the borders, which allows control over the spatial size of the output. Padding is often used to preserve the input dimensions after convolution.
Stride specifies the step size with which the kernel slides across the input. A stride greater than one reduces the output resolution.
Dilation expands the convolutional kernel (also known as the receptive field) without increasing the number of parameters by involving pixel skipping with a given rate (typically 2 or 3), allowing the model to capture long-range dependencies efficiently (Yu and Koltun, 2016).
After convolution, activation functions are applied element-wise to introduce nonlinearity. Common activation functions include:
Rectified Linear Unit (ReLu), , is widely used due to its simplicity and effectiveness in mitigating the vanishing gradient problem ().
Tanh, , maps inputs to (−1,1), centered around zero, but can still suffer from vanishing gradients.
Softmax converts a vector of outputs into a probability distribution, typically applied only in the final layer for multi-class classification.
Beyond the core convolutional layers, a CNN model can include additional specialized layers or processing steps that enhance its ability to learn the input–target relationship:
Dropout layer temporarily deactivates a fraction of neurons during each training iteration, introducing randomness into the learning process and helping to mitigate overfitting. This technique promotes more uniform learning across neurons by preventing reliance on any single neuron. During inference, all neurons are active, but their outputs are scaled to account for the dropout applied during training.
Batch normalization layer normalizes the input of each layer for all training batches. This process helps stabilize and accelerate training by reducing internal covariate shift across all layers.
3.2 Localization approaches
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 (; 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.
3.2.1 For multilayer perceptrons
MLPs are commonly applied in remote sensing as pixel-wise regressors, where a single model is trained over all spatial locations. This imposes a compromise solution that does not account for regional variability in land surface characteristics, often leading to regional biases. To mitigate these issues, several localization strategies have been proposed (
):
Physical variable augmentation: Inputs, such as soil ST or vegetation indices (e.g., LAI), are introduced to better represent the surface conditions. This approach enhances the model’s physical realism by better constraining the inverse radiative transfer and allows for more generalized models with global applicability (; ; ).
Geographic coordinate addition: Spatial static information such as latitude and longitude [or other coordinate-derived variables (Mahara and Rishe, 2023)] is introduced as proxy features to localize the model. This strategy helps partition the input space and simplifies the underlying relationships by implicitly capturing location-dependent processes. While less physically interpretable (latitude and longitude do not intervene in a radiative transfer model linking satellite observations and variables to retrieve), this statistical localization can be effective in improving regional accuracy in a global model () by modulating the NN behavior in several parts of the spatial domain.
The choice between these strategies depends on the intended application. For example, purely statistical approaches [e.g., CDF-matching at the site level ()] are commonly used in assimilation frameworks, where improving dynamics may outweigh physical interpretability. However, it is important to note that even physically motivated inputs may be exploited by the NN in a purely statistical way—leveraging correlations rather than causal relationships. This is evident when inspecting Jacobians (i.e., sensitivities) of trained models, which may show unrealistic patterns. Addressing this challenge may require the integration of physical constraints during training (; ) or the development of hybrid models where physical knowledge provides structural constraints and NNs learn the unknown parameters.
In practice, especially when the information content of the observations is not enough to fully characterize the variable to monitor, some form of localization is necessary to reduce regional biases. One can use only physical variables, but incorporating both physical and geographical inputs has become a common strategy (
Singh and Gaurav, 2023;
).
Pixel-scale modeling: At the most localized level, fully independent MLPs can be trained per pixel, allowing the model to specialize in the local observation-to-variable relationships. This “pixel-specific” modeling simplifies the retrieval task but sacrifices generalization and scalability. There is also a risk of overfitting due to the smaller sample size per model. Such a strategy may be appropriate in cases where auxiliary physical variables are unavailable or unreliable. It is worth noting that approaches such as CDF matching () or a linear regression (Wagner et al., 2013) of an SM index to a physical SM range (e.g., from ECMWF) are also examples of pixel-specific models.
3.2.2 For convolutional neural networks
Grid information: CNNs can inherently incorporate spatial information through convolutional layers that apply filters or kernels across the input images. When trained on a fixed domain (e.g., CONUS), the pixel location in the image is directly equivalent to the geographic coordinates information previously cited for MLPs. As such, traditional CNNs already embed a form of localization. However, standard CNNs rely on “weight sharing” (), where the same set of filters is applied across all spatial locations. This can limit their capacity to capture region-specific behaviors in large, heterogeneous domains. To address this, CNNs typically use multiple stacked convolutional layers to develop hierarchical features that generalize spatial variations. Nonetheless, this approach may require a large number of filters to approximate diverse land surface conditions.
Local connection: the locally connected layers were first proposed by and later adopted in remote sensing [e.g., in ; Pellet et al. (2025)]. These layers are convolutional layers but do not share weights across space. Instead, each spatial region has its own filter set, allowing the network to learn location-specific patterns. This architectural approach is analogous to the pixel-specific strategy used in MLPs, while benefiting the local neighborhood (within the kernel window) to enhance retrieval for the central pixel.
To comprehensively assess the impact of localization strategies, we explored a set of model architectures with varying input configurations for SM retrieval. The following section introduces these models and the experimental setups used to investigate how different forms of localization influence the spatial and temporal accuracy of the retrievals.
3.3 Proposed model architectures
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 and/or incorporates auxiliary inputs, including physical variables and geographic coordinates. Different architectures are explored, denoted as , where represents the number of physical variables and 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 -l with three physical inputs: , 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 16 kernels to extract low-level spatial features, followed by three dilated convolutional layers ( 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 ( 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 and/or additional auxiliary inputs. Similar to the standard MLP, the notation indicates a CNN with physical variables and 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 layer to capture local-specific patterns, without weight sharing. We considered in this case a -l model with three inputs, including three images of , 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.
3.4 Evaluation metrics
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, /m3), and bias (Bias, /m3). The standard deviation (Std, /m3) of the difference between two datasets is also reported.
In addition, we also evaluated the dampening and inflating effects () by comparing the range of values in the predicted SM dataset with that of the reference ERA5 dataset . A dampening effect occurs when , indicating a reduced range in the predictions. Conversely, an inflating effect is observed when . These effects are quantified as:
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 ().
4 Analysis of daily soil moisture retrievals
4.1 Overall performance of MLPs and CNNs
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.
TABLE 1
| Scenarios | Input parameters | Statistical metrics | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ST | LAI | APEI | dT | lat | lon | (/m3) | (/m3) | Dampening/Inflating (%) | |||||
| MLP0,2 | | | | | | x | x | 0.44 | - | 0.109 | 0.44 | 0.115 | −100 |
| MLP1,0 | x | | | | | | | 0.37 | 0.24 | 0.110 | 0.41 | 0.118 | −74 |
| MLP1,2 | x | | | | | x | x | 0.51 | 0.24 | 0.104 | 0.54 | 0.110 | −78 |
| MLP3,0 | x | x | x | | | | | 0.59 | 0.61 | 0.096 | 0.55 | 0.103 | −45 |
| MLP3,2 | x | x | x | | | x | x | 0.64 | 0.60 | 0.092 | 0.60 | 0.098 | −53 |
| MLP5,0 | x | x | x | x | x | | | 0.74 | 0.52 | 0.091 | 0.71 | 0.086 | −25 |
| MLP5,2 | x | x | x | x | x | x | x | 0.74 | 0.53 | 0.081 | 0.71 | 0.085 | −25 |
| MLP3,0-l | x | x | x | | | | | 0.92 | 0.69 | 0.047 | 0.92 | 0.048 | −26 |
| CNN0,2 | | | | | | x | x | 0.64 | - | 0.063 | 0.65 | 0.074 | −100 |
| CNN1,0 | x | | | | | | | 0.85 | 0.50 | 0.065 | 0.85 | 0.067 | −29 |
| CNN1,2 | x | | | | | x | x | 0.87 | 0.52 | 0.062 | 0.87 | 0.064 | −39 |
| CNN3,0 | x | x | x | | | | | 0.91 | 0.73 | 0.055 | 0.91 | 0.057 | −36 |
| CNN3,2 | x | x | x | | | x | x | 0.92 | 0.73 | 0.050 | 0.91 | 0.052 | −32 |
| CNN5,0 | x | x | x | x | x | | | 0.91 | 0.72 | 0.058 | 0.91 | 0.059 | −44 |
| CNN5,2 | x | x | x | x | x | x | x | 0.90 | 0.71 | 0.055 | 0.89 | 0.057 | −30 |
| CNN3,0-l | x | x | x | | | | | 0.93 | 0.69 | 0.048 | 0.92 | 0.049 | −32 |
Performance of MLP and CNN retrieval models over the 2019 validation dataset, for different scenarios with different combinations of input parameters: Physical variables: ASCAT backscatter , soil temperature (ST), leaf area index (LAI), antecedent precipitation evaporation index (APEI), amplitude of the diurnal cycle of surface temperature (dT); Geographic coordinates: latitude (lat), longitude (long).
For example, is the MLP model with inputs including physical variables and geographical variables; -l is the localized version. The statistical metrics include: total correlation (); temporal correlation () and RMSE (, /m3); spatial correlation () and RMSE (, /m3); and dampening/inflating effects (%).
Bold values highlight the performance of the four representative models that are examined in the following sections.
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 —which uses only the ASCAT backscatter as an input—to 0.59 in . 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 captures some SM patterns in the southeastern part, 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.
FIGURE 2
On the other hand, CNNs exhibit a different behavior. While performance improves from to or , the model achieves a high total correlation of 0.91 using only three physical inputs (e.g., , 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 input allows 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 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).
FIGURE 3
It is also noted that when increasing the localization: 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.
4.2 Impacts of the localization information
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 . Incorporating coordinate information significantly enhances spatial correlation, as demonstrated by an increase from 0.37 in the model (model with only as an input) to 0.51 in the model (MLP with , 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 model with inputs including , ST, and LAI) instead of using coordinate information (as done in the model with three inputs: , 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 model, which incorporates , 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 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 configurations.
It is not surprising that the two more extreme localization cases—-l and -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 -l and -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 and . 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).
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.
5 Evaluation of extreme cases
5.1 Minimum and maximum soil moisture retrievals
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.
FIGURE 5
Both the non-localized models, and , 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, -l and -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.
5.2 Dampening and inflating effects
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 (; Skafte et al., 2019). However, under certain conditions—particularly when training data from different locations are pooled—regression models can also exhibit inflation (). 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 (-l and -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.
6 Evaluation of daily soil moisture retrievals against in situ measurements
In this section, we evaluated the performance of our daily NN-based SM retrievals using the in situ measurements from the ISMN dataset during the 2019 validation period at 638 sites in the study area. For this assessment, we selected two representative models— and —and compared them alongside the ERA5 and ASCAT H120 SM products against measurements. Each ISMN site corresponds to a specific geographic point, while the SM retrievals are provided at a 0.25 resolution. Therefore, for each site, we selected the nearest pixel cell to represent the corresponding model data for comparison. We first present time series examples from randomly selected sites, followed by a comprehensive statistical evaluation across all 638 locations.
6.1 Time series analysis
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.
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 (). 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), surpasses ERA5, with a correlation of 0.83 against the in situ data. In most cases, outperforms , further highlighting the advantages of convolutional architectures in capturing spatial and temporal patterns of SM.
6.2 Statistical comparison
Table 2 presents the statistical performance of four SM retrievals (H120 product, , , 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 (), bias (, /m3), and standard deviation (, /m3).
TABLE 2
| SM estimates | (/m3) | (/m3) | (/m3) | ||
|---|---|---|---|---|---|
| ASCAT retrievals | H120 | 0.61 | 0.108 | 0.024 | 0.080 |
| 0.60 | 0.136 | 0.108 | 0.068 | ||
| 0.68 | 0.107 | 0.067 | 0.061 | ||
| ERA5 | 0.75 | 0.106 | 0.061 | 0.060 | |
Performance metrics of soil moisture (SM) retrievals compared to the ISMN in situ measurements across 638 sites for the year 2019.
Metrics include temporal correlation (), root mean square error (, /m3), bias (, /m3), and standard deviation (, /m3), reported as the median values across all sites. The evaluated SM retrievals include the ASCAT-derived retrievals: H120 product, , ; and ERA5 reanalysis.
FIGURE 7
Across 638 sites, the ACAST H120 SM product demonstrates a moderate median temporal correlation ( = 0.61), which is comparable to the other three evaluated SM products. Although it achieves the lowest median bias ( = 0.024 /m3) among all products, its standard deviation is relatively higher ( = 0.080 /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 ( = 0.75) and low errors ( = 0.106, = 0.061, = 0.060 /m3). This superior performance is attributed to its reanalysis nature, which combines model data with observations, thereby enhancing consistency with ground-based measurements (). Our NN-based retrievals, and , also show competitive performance. In particular, achieves a temporal correlation of 0.68, closely approaching ERA5, while records a lower correlation of 0.60. Notably, the bias and standard deviation distributions of 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 () of 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 and ( = () - ()). 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.
6.3 Effects of soil texture and land-cover types
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 (). Between two NN models, consistently outperforms , showing higher correlations and lower RMSE, bias, and standard deviation across all classes. Across soil texture groups (Figure 8), , 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), also outperforms ERA5 in correlation over mixed forest, and in RMSE over evergreen and deciduous forests, cropland, grassland, pasture, and developed areas.
FIGURE 8
FIGURE 9
7 Conclusions and perspectives
This study presented a comprehensive evaluation of various NN-based models (MLPs and CNNs) for daily SM retrieval, with a particular focus on the role of localization. Overall, NN-based models demonstrate strong performance in capturing both temporal and spatial dynamics of SM, showing high correlations with the target ERA5 SM as well as in situ measurements.
Among the two types of models evaluated, CNNs more effectively exploit spatial patterns and outperform MLPs. Our investigation of localization strategies suggests that physical variables should be preferred over purely geographical ones:
Latitude and longitude, while useful for capturing the overall structure, do not fully convey finer details. Specifically, they tend to smooth variations even when the variable displays spatial heterogeneity, as observed in surface variables such as SM.
Physical variables, particularly those linked to radiative transfer processes, preserve spatial heterogeneity and promote more physically consistent retrievals. Moreover, when CNNs are used, coordinate information can be implicitly encoded in input imagery, especially in localized CNN configurations.
Localized NN-based models (MLP-l and CNN-l) show improved performance in retrieving extreme values, highlighting their potential for specialized applications such as drought monitoring or flood forecasting. However, caution is needed concerning the risks of overfitting—especially for pixel-specific models like MLP-l—due to the limited sample size available for each localized model.
The study highlights several methodological considerations that are relevant to future SM retrieval research. ERA5 remains one of the most reliable large-scale reference datasets for SM, as it integrates physical modeling with the assimilation of diverse observational inputs (i.e., multiple satellite products, various precipitation datasets, and in situ measurements). As a result, NN-based SM retrievals cannot be expected to surpass ERA5 skill globally. It is nevertheless important to acknowledge that ERA5 exhibits local-dependent biases, and NNs trained on ERA5 inevitably inherit these biases. However, our results suggest that localization can mitigate some of these local errors. In certain cases, our retrieval (based on ASCAT backscatter and auxiliary variables) aligns more closely with in situ measurements than ERA5. Moreover, the performance improvements were observed across different soil textures and land-cover types. This improvement can arise from various factors: for example, the models may extract observation signal components not fully exploited in the reanalysis, correct specific local biases present in ERA5, or better capture high-frequency variability at certain sites. This finding aligns with previous work (Rodríguez-Fernández et al., 2019).
Overall, although a single-sensor retrieval cannot comprehensively outperform ERA5, our findings demonstrate that deep learning can extract high-quality SM information from ASCAT alone. This contributes to the advancement of independent satellite-based SM retrievals, which remain essential for climate analysis, model evaluation, long-term consistency assessments, and data assimilation (). Furthermore, in many practical applications, the retrieved SM will ultimately be assimilated into a land surface model whose climatology is already based on ERA-type reanalyses. This further supports the use of ERA5 as an appropriate and consistent training target for the NN retrieval.
For future works, these deep learning approaches can be extended to different regions or even to a global-scale analysis. This can ultimately support the development of a long-term SM record by leveraging observations from multiple satellite sensors. However, integrating data across different sensors poses notable challenges, including inter-sensor calibration and varying retrieval uncertainties. Addressing these complexities is essential for generating a coherent multi-sensor SM product. Incorporating in situ measurements into training (e.g., via regional fine-tuning) can be a promising approach, though technically challenging due to scale mismatches between point and satellite measurements. Furthermore, deep learning techniques have the potential not only to exploit spatial patterns, as demonstrated in this study using CNNs, but also to utilize temporal information. Therefore, we plan to explore the possibility of testing transformers and embedding representation methods to perform retrieval across a full sequence of images, rather than just a single image at a specific time. Another direction is the potential for forecasting, rather than just retrieval. Specifically, we might focus on short- and medium-term SM forecasting using ConvLSTM models (Wang et al., 2024; Rabiei et al., 2025a). Given that SM layers in the ground exhibit strong dependencies tied to soil texture and geology, our retrieval methods could also aim to estimate moisture levels in deeper layers beyond the surface layer we considered here.
The methods proposed in this study are general and hold potential for a wide range of applications beyond SM retrieval. For example, localized CNN approaches could be used to retrieve surface water extent at high spatial resolution. Moreover, these methods can also be applied in forward mode, using surface parameters such as land surface temperature, vegetation cover, and surface roughness to estimate land surface emissivities or even satellite-observed brightness temperatures. Such a forward model could be integrated into data assimilation frameworks to support numerical weather prediction systems. Finally, although this study focused on methodological development, at 0.25 and daily resolution, producing higher-resolution near-real-time SM maps would enhance real-world practical applications such as precision agriculture (for optimal watering), hydrological modeling, fire security during dry periods, and identifying areas prone to flooding (). This improvement can be achieved by using higher-resolution sensors (e.g., Sentinel) or by directly downscaling the obtained SM fields within a transformer-based retrieval scheme.
Statements
Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: The ASCAT data supporting this study can be obtained from the Metop ASCAT SSM CDR (), https://hsaf.meteoam.it/ (last access: 22 January 2025). Porosity data (Rodell et al., 2004) are from https://ldas.gsfc.nasa.gov/gldas/soils (last access: 31 May 2025). The ERA5 reanalysis dataset can be downloaded from https://cds.climate.copernicus.eu (Hersbach et al., 2023) (last access: 15 March 2025). Finally, the International Soil Moisture Network data are available at https://ismn.earth/en/data/ () (last access: 3 March 2025).
Author contributions
LD: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review and editing.
Funding
The author(s) declared that financial support was received for this work and/or its publication. The CERISE project (grant agreement no. 101082139) is funded by the European Union.
Acknowledgments
The author thanks Filipe Aires (LIRA) and Victor Pellet (LMD) for valuable discussions and for their comments and suggestions on the original draft of the manuscript. She also thanks Patricia de Rosnay and Peter Weston (ECMWF) for interesting discussions.
Conflict of interest
Author LD was employed by Estellus.
Generative AI statement
The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI tools, including Grammarly and ChatGPT, were used solely for language editing and refinement. All ideas, interpretations, and original text were created by the author. AI assistance was limited to enhancing clarity, grammar, and style, without contributing to the conceptual or intellectual content of the work.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Author disclaimer
Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the Commission. Neither the European Union nor the granting authority can be held responsible for them.
References
1
AiresF.SchmittM.ChedinA.ScottN. (1999). The “weight smoothing” regularization of mlp for Jacobian stabilization. IEEE Trans. Neural Netw.10, 1502–1510. 10.1109/72.809096
2
AiresF.PrigentC.RossowW. B.RothsteinM. (2001). A new neural network approach including first guess for retrieval of atmospheric water vapor, cloud liquid water path, surface temperature, and emissivities over land from satellite microwave observations. J. Geophys. Res. Atmos.106, 14887–14907. 10.1029/2001JD900085
3
AiresF.PrigentC.RossowW. B. (2004). Neural network uncertainty assessment using bayesian statistics: a remote sensing application. Neural Comput.16, 2415–2458. 10.1162/0899766041941925
4
AiresF.PrigentC.RossowW. B. (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. Atmos.110. 10.1029/2004JD005094
5
AiresF.BoucherE.PelletV. (2021a). Convolutional neural networks for satellite remote sensing at coarse resolution. Application for the sst retrieval using iasi. Remote Sens. Environ.263, 112553. 10.1016/j.rse.2021.112553
6
AiresF.WestonP.de RosnayP.FairbairnD. (2021b). Statistical approaches to assimilate ascat soil moisture information—i. methodologies and first assessment. Q. J. R. Meteorological Soc.147, 1823–1852. 10.1002/qj.3997
7
AtkinsonP. M.TatnallA. R. L. (1997). Introduction neural networks in remote sensing. Int. J. Remote Sens.18, 699–709. 10.1080/014311697218700
8
BatchuV.NearingG.GulshanV. (2023). A deep learning data fusion model using sentinel-1/2, soilgrids, smap, and gldas for soil moisture retrieval. J. Hydrometeorol.24, 1789–1823. 10.1175/jhm-d-22-0118.1
9
BellJ. E.PaleckiM. A.BakerC. B.CollinsW. G.LawrimoreJ. H.LeeperR. D.et al (2013). U.s. climate reference network soil moisture and temperature observations. J. Hydrometeorol.14, 977–988. 10.1175/jhm-d-12-0146.1
10
BernhardtJ.CarletonA. M.LaMagnaC. (2018). A comparison of daily temperature-averaging methods: spatial variability and recent change for the conus. J. Clim.31, 979–996. 10.1175/JCLI-D-17-0089.1
11
BishopC. M. (1995). Neural networks for pattern recognition. Oxford University Press. 10.1093/oso/9780198538493.001.0001
12
BoucherE.AiresF. (2023). Improving remote sensing of extreme events with machine learning: land surface temperature retrievals from iasi observations. Environ. Res. Lett.18, 024025. 10.1088/1748-9326/acb3e3
13
BoucherE.AiresF.PelletV. (2023). Towards a new generation of artificial-intelligence-based infrared atmospheric sounding interferometer retrievals of surface temperature: part i – methodology. Q. J. R. Meteorological Soc.149, 1180–1196. 10.1002/qj.4447
14
CampbellG. S. (1985). Soil physics with basic: transport models for soil-plant systems. Amsterdam: Elsevier.
15
ChenY.López-MorenoI.SainathT. N.VisontaiM.ÁlvarezR.ParadaC. (2015). Locally-connected and convolutional neural networks for small footprint speaker recognition. Dresden, Germany: Interspeech.
16
DorigoW.XaverA.VreugdenhilM.GruberA.DostálováA.Sanchis-DufauA. D.et al (2013). Global automated quality control of in situ soil moisture data from the international soil moisture network. Vadose Zone J.12, 1–21. 10.2136/vzj2012.0097
17
DorigoW.HimmelbauerI.AbererD.SchremmerL.PetrakovicI.ZappaL.et al (2021). The international soil moisture network: serving earth system science for over a decade. Hydrology Earth System Sciences25, 5749–5804. 10.5194/hess-25-5749-2021
18
DruschM.WoodE. F.GaoH. (2005). Observation operators for the direct assimilation of trmm microwave imager retrieved soil moisture. Geophys. Res. Lett.32. 10.1029/2005GL023623
19
El HajjM.BaghdadiN.ZribiM.BelaudG.ChevironB.CouraultD.et al (2016). Soil moisture retrieval over irrigated grassland using x-band sar data. Remote Sens. Environ.176, 202–218. 10.1016/j.rse.2016.01.027
20
GoïtaK.Gonzalez-RubioR.BéniéG. B.RoyerA.MichaudF. (1994). Literature review of artificial neural networks and knowledge-based systems for image analysis and interpretation of data in remote sensing. Can. J. Electr. Comput. Eng.19, 53–61. 10.1109/CJECE.1994.6592069
21
GoodfellowI.BengioY.CourvilleA. (2016). Deep learning. MIT Press. Available online at: http://www.deeplearningbook.org.
22
GreifenederF.NotarnicolaC.WagnerW. (2021). A machine learning-based approach for surface soil moisture estimations with google earth engine. Remote Sens.13, 2099. 10.3390/rs13112099
23
GuoZ.DirmeyerP. A.HuZ.-Z.GaoX.ZhaoM. (2006). Evaluation of the second global soil wetness project soil moisture simulations: 2. Sensitivity to external meteorological forcing. J. Geophys. Res. Atmos.111. 10.1029/2006JD007845
24
H SAF (2021a). Algorithm theoretical baseline document (ATBD) metop ASCAT surface soil moisture climate data record v7 12.5 km sampling (H119) and extension (H120)
25
H SAF (2021b). ASCAT surface soil moisture climate data record v7 12.5 km sampling - metop. 10.15770/EUM_SAF_H_0009
26
HanQ.ZengY.ZhangL.WangC.PrikaziukE.NiuZ.et al (2023). Global long term daily 1 km surface soil moisture dataset with physics informed machine learning. Sci. Data10, 101. 10.1038/s41597-023-02011-7
27
HastieT.TibshiraniR.FriedmanJ. (2009). Neural networks. New York, NY: Springer, 389–416. 10.1007/978-0-387-84858-7_11
28
HenglT. (2018). Soil texture classes (usda system) for 6 soil depths (0, 10, 30, 60, 100 and 200 cm) at 250 m (v0.2). 10.5281/zenodo.2525817
29
HersbachH.BellB.BerrisfordP.BiavatiG.HorányiA.Muñoz SabaterJ.et al (2023). Era5 hourly data on single levels from 1940 to present. 10.24381/cds.adbb2d47
30
HuangJ.SehgalV.AlvarezL. V.BroccaL.CaiS.ChengR.et al (2025). Remotely sensed high-resolution soil moisture and evapotranspiration: bridging the gap between science and society. Water Resour. Res.61, e2024WR037929. 10.1029/2024WR037929
31
KolassaJ.AiresF.PolcherJ.PrigentC.JimenezC.PereiraJ. M. (2013). Soil moisture retrieval from multi-instrument observations: information content analysis and retrieval methodology. J. Geophys. Res. Atmos.118, 4847–4859. 10.1029/2012JD018150
32
KolassaJ.GentineP.PrigentC.AiresF. (2016). Soil moisture retrieval from amsr-e and ascat microwave observation synergy. part 1: satellite data analysis. Remote Sens. Environ.173, 1–14. 10.1016/j.rse.2015.11.011
33
LeavesleyG. H.DavidO.GarenD. C.Nrcs-UsdaN.GoodbodyA. G.LeaJ. K.et al (2010). A modeling framework for improved agricultural water-supply forecasting
34
LecunY.BottouL.BengioY.HaffnerP. (1998). Gradient-based learning applied to document recognition. Proc. IEEE86, 2278–2324. 10.1109/5.726791
35
LiuY.DorigoW.ParinussaR.de JeuR.WagnerW.McCabeM.et al (2012). Trend-preserving blending of passive and active microwave soil moisture retrievals. Remote Sens. Environ.123, 280–297. 10.1016/j.rse.2012.03.014
36
MadadikhaljanM.SchmittM. (2025). Geolocation-aware deep coding. PFG – J. Photogrammetry, Remote Sens. Geoinformation Sci.93, 3–18. 10.1007/s41064-024-00328-5
37
MaggioriE.TarabalkaY.CharpiatG.AlliezP. (2017). Convolutional neural networks for large-scale remote-sensing image classification. IEEE Trans. Geoscience Remote Sens.55, 645–657. 10.1109/TGRS.2016.2612821
38
MaharaA.RisheN. (2023). Integrating location information as geohash codes in convolutional neural network-based satellite image classification. IPSI Trans. Internet Res.19, 24–30. 10.58245/ipsi.tir.2302.04
39
OchsnerT. E.CoshM. H.CuencaR. H.DorigoW. A.DraperC. S.HagimotoY.et al (2013). State of the art in large-scale soil moisture monitoring. Soil Sci. Soc. Am. J.77, 1888–1919. 10.2136/sssaj2013.03.0093
40
PelletV.AiresF.BoucherE.VoldenE. (2025). Enhancing soil moisture statistical retrieval from smos using partial convolutions and localization strategies. J. Appl. Meteorology Climatol.64, 1951–1965. 10.1175/JAMC-D-25-0041.1
41
PetchiappanA.Steele-DunneS. C.VreugdenhilM.HahnS.WagnerW.OliveiraR. (2022). The influence of vegetation water dynamics on the ascat backscatter–incidence angle relationship in the amazon. Hydrology Earth Syst. Sci.26, 2997–3019. 10.5194/hess-26-2997-2022
42
PrigentC.AiresF.RossowW. B.RobockA. (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: relationship of satellite observations to in situ soil moisture measurements. J. Geophys. Res. Atmos.110. 10.1029/2004JD005087
43
RabieiS.BabaeianE.GrunwaldS. (2025a). Deep learning-based short- and mid-term surface and subsurface soil moisture projections from remote sensing and digital soil maps. Remote Sens.17, 3219. 10.3390/rs17183219
44
RabieiS.BabaeianE.GrunwaldS. (2025b). Surface and subsurface soil moisture estimation using fusion of smap, nldas-2, and solus100 data with deep learning. Remote Sens.17, 659. 10.3390/rs17040659
45
RodellM.HouserP. R.JamborU.GottschalckJ.MitchellK.MengC.-J.et al (2004). The global land data assimilation system. Bull. Am. Meteorological Soc.85, 381–394. 10.1175/bams-85-3-381
46
Rodríguez-FernándezN. J.AiresF.RichaumeP.KerrY. H.PrigentC.KolassaJ.et al (2015). Soil moisture retrieval using neural networks: application to smos. IEEE Trans. Geoscience Remote Sens.53, 5991–6007. 10.1109/TGRS.2015.2430845
47
Rodríguez-FernándezN.de RosnayP.AlbergelC.RichaumeP.AiresF.PrigentC.et al (2019). Smos neural network soil moisture data assimilation in a land surface model and atmospheric impact. Remote Sens.11, 1334. 10.3390/rs11111334
48
RumelhartD. E.HintonG. E.WilliamsR. J. (1986). Learning representations by back-propagating errors. Nature323, 533–536. 10.1038/323533a0
49
SaxtonK. E.RawlsW. J. (2006). Soil water characteristic estimates by texture and organic matter for hydrologic solutions. Soil Sci. Soc. Am. J.70, 1569–1578. 10.2136/sssaj2005.0117
50
SchaeferG. L.CoshM. H.JacksonT. J. (2007). The usda natural resources conservation service soil climate analysis network (scan). J. Atmos. Ocean. Technol.24, 2073–2077. 10.1175/2007jtecha930.1
51
SeneviratneS. I.CortiT.DavinE. L.HirschiM.JaegerE. B.LehnerI.et al (2010). Investigating soil moisture–climate interactions in a changing climate: a review. Earth-Science Rev.99, 125–161. 10.1016/j.earscirev.2010.02.004
52
ShawB. L.PielkeR. A.ZieglerC. L. (1997). A three-dimensional numerical simulation of a Great Plains dryline. Mon. Weather Rev.125, 1489–1506. 10.1175/1520-0493(1997)125<1489:ATDNSO>2.0.CO;2
53
SinghA.GauravK. (2023). Deep learning and data fusion to estimate surface soil moisture from multi-sensor satellite images. Sci. Rep.13, 2251. 10.1038/s41598-023-28939-9
54
SkafteN.Jø rgensenM.HaubergS. R. (2019). “Reliable training and estimation of variance networks,” in Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems. Vancouver, BC, Canada, December 8–14, 2019. Editors WallachH. M.LarochelleH.BeygelzimerA.d’Alché-BucF.FoxE. B.GarnettR. (NeurIPS). 6323–6333.
55
SociC.HersbachH.SimmonsA.PoliP.BellB.BerrisfordP.et al (2024). The era5 global reanalysis from 1940 to 2022. Q. J. R. Meteorological Soc.150, 4014–4048. 10.1002/qj.4803
56
U.S. Geological Survey (USGS) (2024). Annual nlcd collection 1 science products. 10.5066/P94UXNTS
57
VereeckenH.HuismanJ. A.Hendricks FranssenH. J.BrüggemannN.BogenaH. R.KolletS.et al (2015). Soil hydrology: recent methodological advances, challenges, and perspectives. Water Resour. Res.51, 2616–2633. 10.1002/2014WR016852
58
WagnerW.NollJ.BorgeaudM.RottH. (1999). Monitoring soil moisture over the canadian prairies with the ers scatterometer. IEEE Trans. Geoscience Remote Sens.37, 206–216. 10.1109/36.739155
59
WagnerW.HahnS.KiddR.MelzerT.BartalisZ.HasenauerS.et al (2013). The ascat soil moisture product: a review of its specifications, validation results, and emerging applications. Meteorol. Z.22, 5–33. 10.1127/0941-2948/2013/0399
60
WangY.ShiL.HuY.HuX.SongW.WangL. (2024). A comprehensive study of deep learning for soil moisture prediction. Hydrology Earth Syst. Sci.28, 917–943. 10.5194/hess-28-917-2024
61
YuF.KoltunV. (2016). “Multi-scale context aggregation by dilated convolutions,” in Conference: international conference on learning representations (ICLR).
62
ZhangX.SunX.LinZ. (2025). Improving soil moisture prediction using gaussian process regression. Smart Agric. Technol.11, 100905. 10.1016/j.atech.2025.100905
Summary
Keywords
ASCAT, deep learning, localization, neural network, soil moisture retrieval
Citation
Dinh LA (2026) ASCAT soil moisture retrieval using deep learning: a focus on localization strategy. Front. Remote Sens. 6:1718353. doi: 10.3389/frsen.2025.1718353
Received
03 October 2025
Revised
11 December 2025
Accepted
15 December 2025
Published
23 January 2026
Volume
6 - 2025
Edited by
Amen Al-Yaari, Université Paris-Sorbonne, France
Reviewed by
Xiangzhuo Liu, INRA Centre Provence-Alpes-Côte d’Azur, France
Saman Rabiei, University of Florida, United States
Updates
Copyright
© 2026 Dinh.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Lan Anh Dinh, lan-anh.dinh@obspm.fr
† Present addresses: Lan Anh Dinh, LIRA, Observatoire de Paris, Université PSL, Sorbonne Université, CNRS, Paris, France
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.