High Spatial Resolution Soil Moisture Mapping over Agricultural Field Integrating SMAP, IMERG, and Sentinel-1 Data in Machine Learning Models

Abstract

Soil moisture content (SMC) is a critical parameter for agricultural productivity, particularly in semi-arid regions, where irrigation practices are extensively used to offset water deficits and ensure decent yields. Yet, the socio-economic and remote context of these regions prevents sufficiently dense SMC monitoring in space and time to support farmers in their work to avoid unsustainable irrigation practices and preserve water resource availability. In this context, our study addresses the challenge of high spatial resolution (i.e., 20 m) SMC estimation by integrating remote sensing datasets in machine learning models. For this purpose, a dataset made of 166 soil samples’ SMC along with corresponding SMC, precipitation, and radar signal derived from Soil Moisture Active Passive (SMAP), Integrated Multi-satellitE Retrievals for GPM (IMERG), and Sentinel-1 (S1), respectively, was used to assess four machine learning models’ (Decision Tree—DT, Random Forest—RF, Gradient Boosting—GB, Extreme Gradient Boosting—XGB) reliability for SMC mapping. First, each model was trained/validated using only the coarse spatial resolution (i.e., 10 km) SMAP SMC and IMERG precipitation estimates as independent features, and, second, S1 information (i.e., 20 m) derived from single scenes and/or composite images was added as independent features to highlight the benefit of information (i.e., S1 information) for SMC mapping at high spatial resolution (i.e., 20 m). Results show that integrating S1 information from both single scenes and composite images to SMAP SMC and IMERG precipitation data significantly improves model reliability, as R² increased by 12% to 16%, while RMSE decreased by 10% to 18%, depending on the considered model (i.e., RF, XGB, DT, GB). Overall, all models provided reliable SMC estimates at 20 m spatial resolution, with the GB model performing the best (R² = 0.86, RMSE = 2.55%).

1. Introduction

Soil moisture content (SMC) in the unsaturated soil zone [1] plays a pivotal role in agriculture, influencing seed germination, plant growth, and nutrient uptake in the root zone [2,3]. Actually, SMC is a key indicator of soil health and agricultural productivity, affecting both the physical and chemical properties of soil [4,5,6]. Understanding and accurately monitoring SMC is essential for optimizing irrigation practices, improving crop management, mitigating the impacts of droughts, and minimizing water stress on plants to achieve agriculture resilient to the effects of climate change, especially in arid and semi-arid regions [7,8,9].

Conventionally, SMC is monitored through direct measurements such as the gravimetric method, which determines SMC by weighing a soil sample and calculating the water fraction relative to its total weight [10]. Despite its high reliability, this method relies on soil sample collection from the field prior to laboratory analysis, which results in labor-intensive and cost-prohibitive monitoring [3,8]. In consequence, this method is not adapted to intensive SMC, yet is required to adequately capture SMC variation in space and time related to complex interactions among multiple factors such as (i) soil texture and structure, (ii) topographic characteristics, (iii) land cover patterns, (iv) vegetation properties, and (v) meteorological forcing [11,12,13]. To overcome this issue (i.e., large scale and intensive monitoring in space and time), an alternative measurement consists in the use of a TDR (Time Domain Reflectometer) [14], a device that rapidly measures soil moisture using electromagnetic waves, but its accuracy is affected by salinity [15,16]. It is useful for point measurements but not for large-scale monitoring, as it requires physical displacement and does not offer continuous coverage in space and time.

In recent decades, remote sensing technology has transformed the way we observe and analyze the Earth’s surface, providing continuous, large-scale, and non-invasive data acquisition capabilities. Active microwave sensors in low frequencies (X, C, and L-bands) are well suited for SMC retrieval because the backscatter is very sensitive to soil dielectric properties, a proxy to soil water content [3,17,18]. As SMC increases, the dielectric constant of the soil–water mixture also rises, a variation that can be detected using microwave sensors [17,19]. In comparison to higher-frequency bands (i.e., C and X), L-bands have a greater penetration depth into the soil and are less sensitive to soil roughness, vegetation, and atmospheric conditions [20,21]. In this context, two dedicated satellite missions equipped with L-band microwave radiometers have been deployed for monitoring earth’s surface SMC: ESA’s Soil Moisture and Ocean Salinity (SMOS), launched in 2009, and NASA’s Soil Moisture Active Passive (SMAP), launched in 2015 [20]. However, because of their limited spatial resolution (35 km and 10 km for SMOS and SMAP, respectively), SMOS and SMAP SMC estimates are limited to regional studies [20,22,23,24], and their use for local applications is challenging [9].

In this context, many studies have proposed the use of optical and/or thermal passive sensors data to obtain SMC estimates at a finer spatial resolution. Two main methods stand out. One links SMC to land surface temperature (LST) and vegetation indices [25], while the other links SMC to red and near-infrared (NIR) reflectances [26]. It is worth mentioning that an adaptation of the first method consists in replacing LST data with short-wave infrared (SWIR) data [27]. Continuously improved, these techniques have been successively applied to MODIS, Landsat, and Sentinel-2 data to enable soil SMC monitoring at increasingly finer spatial resolutions. However, due to cloud cover, LST, along with red, NIR, and SWIR data, remains prone to considerable gaps in both space and time, limiting SMC retrieval to clear sky conditions.

To overcome cloud cover difficulties, Sentinel-1 (S1) C-band SAR images offer an interesting alternative, as C-band signal is not altered by cloud cover. Various techniques have been successfully adapted to S1 data to retrieve SMC at high spatial resolution, such as the use of the Water Cloud Model (WCM) [28], Change Detection (CD) technique [29,30], and machine learning approaches [31,32,33]. While S1 images provide high spatial resolution (10 m), the C-band wavelength (approximately 5.5 cm) is more affected by soil roughness and vegetation cover than the L-band wavelength (approximately 30 cm) [18,34,35]. In consequence, S1-based SMC estimates are expected to be less accurate than SMOS and SMAP SMC estimates. Still, some authors have taken advantage of S1’s high spatial resolution (i.e., 10 m) to improve low-resolution SMAP SMC estimates (i.e., 10 km) [9,36,37,38]. To date, these spatial resolution improvements have been limited to 1 km [36,38], which is not adapted to monitor SMC at the agriculture plot level to support farmers in their decision-making.

In the above-described context, this study aims at integrating several C-band (S1) and L-band (SMAP) features with satellite-based precipitation estimates (IMERG) in a machine-learning model. In doing so, this approach allows capitalizing on the L-band (SMAP) ability to monitor SMC in time and on the C-band (S1) ability to monitor SMC variation in space to provide high-resolution SMC estimates (i.e., 20 m) for all sky conditions (cloudy or not). Due to its socio-economic context and the importance of agriculture in maintaining economic activity, the Bolivian Altiplano region is selected as the study area to support agriculture planning.

2. Materials

2.1. Study Area

The research was conducted in two plots of approximately 6500 and 2540 m² located at the Patacamaya experimental station from Universidad Mayor de San Andrés (UMSA) (Figure 1), located in the central Bolivian Altiplano region, at an average altitude of 3800 m.a.s.l. with a flat topography [39]. The region is semiarid with (i) an average rainfall of less than 400 mm∙year⁻¹, concentrated in the austral summer [40], (ii) an average temperature that varies from 4 °C to 8 °C [41], and (iii) a high evapotranspiration rate of approximately 1300 mm∙year⁻¹. Agriculture is the main economic activity, which increasingly depends on irrigation to cope with the water deficit [42]. According to World Reference Base for Soil (WRB) System guidelines, the soils are classified as Association of Leptosols—Durisols—Regosols [43].

2.2. Reference Soil Moisture Data

A total of 166 soil samples were collected within the agricultural plots during twelve field visits starting on 4 September 2023 and ending on 14 January 2024. The visits (n = 12) were realized with a 12-day frequency to match with the S1 satellite observation dates and time (approximately 10 h local time) (Figure 1c). Twelve to fifteen soil samples were collected during each field visits. Each soil sample consists of a composite of 3 subsamples taken diagonally (upper left, center, lower right) from a 10 m × 10 m square area at 10 cm depth. Prior to soil sample collection, SMC was also measured using a TDR-150 Field Scout made by the Spectrum Technologies, Inc. (Aurora, CO, USA). Finally, two SMC measurement were obtained from each composite soil sample, one in the soil laboratory of the Faculty of Chemistry at Universidad Mayor de San Andrés (UMSA) by averaging the three soil samples’ information [10] and the other one from the TDR-150 Field Scout measurements.

2.3. Sentinel-1 Images and Pre-Processing

Sentinel-1 (S1) satellite images provide polarization values in VV (vertical/vertical) and VH (vertical/horizontal) in interferometry mode, with a center frequency of 5.405 GHz. S1 images are available as a Single Look Complex (SLC) product, which provides both phase and amplitude information, and as a post-processed Ground Range Detected (GRD) product, which provides direct surface backscatter in the form of intensity and amplitude images. Due to its dual-polarization antenna that can discriminate subtle changes in SMC, the GRD product has emerged as an essential tool for mapping soil moisture and is therefore being considered for the current study. S1 images are available from both ascending and descending orbits. To maintain consistency in S1 observations across space and time, only the descending orbit is used. All S1 images acquired during field soil sample collection were preprocessed according to four successive steps: (i) border noise removal, (ii) radiometric calibration, (iii) topographic correction, and (iv) focal mean filter to reduce the speckle effect. S1 preprocessing (i.e., downloading, preprocessing, and compositing) was carried out via GEE [44] on the free cloud platform Google Colab (retrieved from https://colab.research.google.com/, accessed on 20 October 2024). Finally, for each soil sampling date a composite image is obtained by averaging the value of two S1 images: the one acquired during the sampling date and the one obtained 12 days before.

2.4. Soil Moisture Active Passive (SMAP)

The SMAP mission is an orbiting satellite launched in January 2015 to measure soil moisture globally using an L-band radiometer. SMAP Level-4 (L4) soil moisture product provides global observations of SMC at different depths (e.g., 0–5 cm, 0–100 cm), along with other research outputs, such as soil temperature and evapotranspiration at high temporal resolution (every 3 h) and low spatial resolution (11 km) [45]. SMAP L4 ensures continuous data availability, even during instrument outages, by relying on land model simulations when necessary [46]. For this study we used SMAP L4 SMC of the top (0–5 cm) and root zone (0–100 cm) layers, both available on GEE.

2.5. Integrated Multi-Satellite Retrievals for GPM (IMERG)

The Global Precipitation Measurement (GPM) mission was launched in February 2014 as the successor to the Tropical Rainfall Measuring Mission (TRMM) [47]. Compared to the TRMM, the GPM enhances spatial resolution (from 0.25° to 0.1°) and temporal resolution (from 3 h to 30 min), expands coverage (from 50°N–50°S to 60°N–60°S), and improves sensitivity for detecting light and solid precipitation [48]. The Integrated Multi-satellitE Retrievals for GPM (IMERG) is the precipitation product derived from the GPM mission. IMERG reliability was previously assessed across the Altiplano region, and results show that IMERG precipitation estimates are reliable for daily total amount monitoring and hydrological modelling [41,49]. For this study, we used the last released version (v.07) available at the hourly time-step in GEE [50].

2.6. Machine Learning Models

Four supervised machine learning (ML) models based on decision trees were evaluated: (i) Random Forest (RF) [51], (ii) Extreme Gradient Boosting (XGB) [52], (iii) Decision Tree (DT) [53], and (iv) Gradient Boosting (GB) [54]. Decision trees are non-parametric supervised learning models commonly used for regression and classification tasks. These models represent decisions and their potential outcomes in a hierarchical structure analogous to a tree, allowing the derivation of simple decision rules from dataset features to predict a target variable (e.g., soil moisture) [55].

In the DT model, a single decision tree is built by recursively dividing the dataset into smaller subsets, using features that maximize information gain or reduce impurity [53,55]. The RF model employs an ensemble of decision trees, each trained on different subsets of the dataset using bagging (bootstrap aggregation). The final prediction is derived as the average of all tree predictions, enhancing model robustness and reducing overfitting [51]. The GB model builds decision trees sequentially, with each tree aiming to correct the errors of its predecessor by minimizing a defined loss function. This iterative process leverages the residuals of the prior tree to guide subsequent tree construction [54]. XGB is an advanced implementation of GB that incorporates parallel computing and regularization techniques, improving computational efficiency and mitigating overfitting [52].

All models were implemented using default hyper-parameters provided by a Python library [52,56], as these default configurations yielded satisfactory performance in the ML models [56]. Default hyper-parameter values are presented in

Table 1. Hyper-parameters considered for the machine learning models’ set-up.

Hyper-Parameter	Set-Up Model				Explanation
	RF *	GB *	DT *	XGB **
n_estimators	100	100	-	100	The number of trees in the forest/boosting stages.
max_features	None	None	None	-	Features considered for splitting. None = n_feature.
max_depth	None	3	None	6	Nodes expand until leaves are pure or contain fewer than min_samples_split samples.
min_samples_split	2	2	2	-	Minimum samples required to split a node.
min_samples_leaf	1	1	1	-	Minimum samples required in a leaf node.
learning_rate	-	0.1	-	0.3	Step size shrinkage to prevent overfitting.
subsample	-	1	-	1	Fraction of training samples used per tree.
criterion/objective	squarederror	Friedman_mse	squarederror	squarederror	The function to measure the quality of a split.
eval_metric	-	-	-	rmse	The function of monitoring the performance model.

Where * scikit learn library, ** xgb library, and “-” is used when the considered models do not include this hyper-parameter.

.

3. Methods

3.1. Laboratory SMC Measurement

SMC was computed using both the gravimetric and volumetric methods. To do so, each soil sample was collected with a metal cylinder directly pushed all the way in from the surface before being removed. Then, the soil sample trapped in the metal cylinder was extracted into a polypropylene plastic container (150 mL), hermetically sealed, and directly brought to the laboratory for analysis. Each soil sample was (i) weighed to obtain humid mass (M_h), (ii) dried during 24 h at 105 °C, (iii) left to cool in a desiccant capsule to avoid absorption of moisture from the environment, and (iv) weighed again to obtain the dry mass (M_d). Finally, gravimetric humidity was computed (Equation (1)) and converted to volumetric humidity (Equation (2)) to compare with TDR Field Scout SMC measurements. Hereafter, the volumetric humidity is considered as the reference SMC values.

$${\mathsf{\theta }}_{\mathrm{g}}={\displaystyle \frac{{\mathrm{M}}_{\mathrm{h}}-{\mathrm{M}}_{\mathrm{d}}}{{\mathrm{M}}_{\mathrm{d}}}}\times 100$$

(1)

$${\mathsf{\theta }}_{\mathrm{v}}={\mathsf{\theta }}_{\mathrm{g}}\times {\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{a}}}{{\mathsf{\rho }}_{\mathrm{w}}}}$$

(2)

where θ_g and θ_v represent the gravimetric and volumetric humidity, respectively, and ρ_a and ρ_w represents the bulk density of soil (1.48 g·cm⁻³ [57]) and the water density (1 g·cm⁻³), respectively.

3.2. TDR Measurement Assessment

SMC obtained through the TDR-150 Field Scout was compared to the SMC obtained in laboratory to assess TDR-150 Field Scout SMC reliability. The comparison is based on R² and RMSE.

3.3. Learning Database Elaboration

Eight polarization indices (PIs) and 36 texture indices (TIs) were derived from all S1 single scenes and composite images (

Table 2. Sentinel-1 features. Note that all indices were computed from both the S1 images acquired during field soil sampling (i.e., single scene) and the corresponding composite image.

N	Index/Acronym	Formula * or Description	Reference
	Sentinel-1 polarization (dB)	VV, VH
1	Polarization index 1 (PI1)	$\mathrm{P}\mathrm{I}1=\mathrm{V}\mathrm{V}+\mathrm{V}\mathrm{H}$	[62]
2	Polarization index 2 (PI2)	$\mathrm{P}\mathrm{I}2={\mathrm{V}\mathrm{V}}^2+\mathrm{V}\mathrm{H}$
3	Polarization index 3 (PI3)	$\mathrm{P}\mathrm{I}3={\mathrm{V}\mathrm{H}}^2-\mathrm{V}\mathrm{V}$
4	Polarization index 4 (PI4)	$\mathrm{P}\mathrm{I}4={\mathrm{V}\mathrm{V}}^2+{\mathrm{V}\mathrm{H}}^2$
5	Polarization index 5 (PI5)	$\mathrm{P}\mathrm{I}5={\displaystyle \frac{{\mathrm{V}\mathrm{V}}^2+{\mathrm{V}\mathrm{H}}^2}{\mathrm{V}\mathrm{H}}}$
6	Polarization index 6 (PI6)	$\mathrm{P}\mathrm{I}6=\mathrm{L}\mathrm{o}\mathrm{g}\left(\mathrm{V}\mathrm{V}\right)\times 10$
7	Polarization index 7 (PI7)	$\mathrm{P}\mathrm{I}7=\mathrm{L}\mathrm{o}\mathrm{g}\left(\mathrm{V}\mathrm{H}\right)\times 10$
8	Polarization index 8 (PI8)	$\mathrm{P}\mathrm{I}8=(\mathrm{L}\mathrm{o}\mathrm{g}\left(\mathrm{V}\mathrm{V}\right)+\mathrm{L}\mathrm{o}\mathrm{g}(\mathrm{V}\mathrm{H}\left)\right)\times 10$
9	Textural index (extracted from VV and VH band) Angular second moment (asm) Contrast (contrast) Correlation (corr) Variance (var) Inverse difference moment (idm) Sum average (savg) Sum variance (savr) Sum entropy (sent) Entropy (ent) Difference variance (dvar) Difference entropy (dent) Information measure of correlation 1 (imcorr1) Information measure of correlation 2 (imcorr2) Maximum correlation coefficient (maxcorr) Dissimilarity (diss) Inertia (inertia) Shade (shade) Cluster prominence (prom)	VV asm, VH asm VV contrast, VH contrast VV corr, VH corr VV var, VH var VV idm, VH idm VV savg, VH savg VV svar, VH svar VV sent, VH sent VV ent, VH ent VV dvar, VH dvar VV dent, VH dent VV imcorr1, VH imcorr1 VV imcorr2, VH imcorr2 VV maxcorr, VH maxcorr VV diss, VH diss VV inertia, VH inertia VV shade, VH shade VV prom, VH prom	[58,59,63]

* Where VV = vertical transmission/vertical reception; VH = vertical transmission/horizontal reception.

). Note that S1 TIs were derived from the Gray Level Co-occurrence Matrix (GLCM) available in GEE [58]. All these indices were selected because they have shown strong correlation with soil moisture [37,59,60,61]. SMAP top layer (0–5 cm) and root zone (0–100 cm) SMC obtained from the nearest time of the field measurement (13–14 h UTC-0), along with IMERG total precipitation amount for the 5-day period preceding the soil moisture measurements date, were considered.

Following this process, the learning database included 166 soil moisture observations with corresponding S1 features obtained from the single scenes (n = 46) and the composite images (n = 46), SMAP SMC (n = 2), and IMERG precipitation amount (n = 1) (Table 2 and

Table 3. Hydrological features.

N	Feature	Reference
1	Total precipitation (5 days)	[50]
2	Top layer SMC (0–5 cm)	[46]
3	Root zone SMC (0–100 cm)

).

3.4. Machine Learning Modelling Set-Up

First, SMAP and IMERG features (i.e., top layer SMC, root zone SMC, and total precipitation) were considered as independent features to assess their potential for SMC mapping across agriculture field (scenario-1). Then, S1 features were aggregated to the independent features to assess the potential improvement in SMC estimates brought by S1 features. To assess the potential of S1 features obtained from single scenes and composite images, different scenarios were considered. First, only S1 features obtained from the single scenes were considered (scenario-2); second, only S1 features obtained from the composite images were considered (scenario-3); and finally, S1 features obtained from both the single scenes and the composite images were considered (scenario-4) (Figure 2).

Multi-collinearity is a common issue in machine learning that reduces the robustness of models because of redundancy in the chosen independent features. To mitigate this, the Recursive Feature Elimination Cross Validation (RFEcv) algorithm was employed to identify the subset of independent features that provides the best performance [64]. As a model-specific feature selection technique (a wrapping method), RFEcv is executed independently for each model. The model undergoes iterative runs, eliminating one redundant feature at a time until the model’s performance experiences the smallest decline. In this process, the 10-fold cross-validation is used, with the Root Mean Square Error (RMSE, Equation (1)) serving as the objective function. Afterward, the Variance Inflation Factor (VIF) is applied to the subset of independent features selected by RFEcv to further reduce multi-collinearity. Only independent features with a VIF lower than 10 are retained [65]. This two-step feature selection process (RFEcv, VIF) was selected, as it considerably improved machine learning outputs [35]. It is worth mentioning that, due to the low number of independent features in scenario-1 (n = 3), no multi-collinearity issue was expected for this scenario, and therefore, the above-described two-step feature selection was only applied to scenario-2, -3, and -4.

Finally, each machine-learning model (RF, GT, GB, and XGB) was trained using 70% (n = 116) of the total learning database observations (n = 166) and validated using the remaining 30% of the observations (n = 50) not used during the training. The 70–30% splitting was based on a random selection among all available observations (n = 166). To ensure consistent comparisons among all the considered model and scenario combinations, the same splitting was used for all considered set-ups, so that each model and scenario combination was trained and validated with the exact same observations.

The models’ output reliability was assessed using the coefficient of determination (R², Equation (3)) and the Root Mean Square Error (RMSE, Equation (4)).

$${\mathrm{R}}^2={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}-\overline{o}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{O}}_{\mathrm{i}}-\overline{o}\right)}^2}}$$

(3)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}-{\mathrm{O}}_{\mathrm{i}}\right)}^2}$$

(4)

where n represents the number of samples, and P_i and O_i represent the predicted and observed values of SMC at the site i, respectively, and $\overline{o}$ represents the average of the observed values.

4. Results

4.1. TDR-150 Field Scout Reliability

Figure 3 shows the correlation between SMCs obtained (i) with the TDR-150 Field Scout and (ii) in laboratory (gravimetric method). With an R² and an RMSE of 0.98 and 2.8%, respectively, the TDR-150 Field Scout provides reliable SMC estimates across the considered region. It is worth mentioning that laboratory SMC measurements are also subject to uncertainties because of the loss of humidity during the transport from the sampling site to the laboratory. In consequence, SMCs obtained through the TDR tend to be slightly superior to the SMCs obtained in the laboratory (Bias = 2.6%, Figure 3). According to this result, the SMCs obtained through the TDR-150 Field Scout are used as a reference (i.e., target) in this study.

4.2. Feature Selection and Importance

Figure 4 shows the feature importance (FI) of the selected independent features for each model and scenario-2, -3, and -4 (SMAP + IMERG + S1). FI was obtained from the models during the training step. It quantifies the respective contribution of each independent variable in the output prediction. For all scenarios (-2, -3, and -4), “total precipitation” (IMERG) and “top layer soil moisture” (SMAP) were identified as the most relevant features, together accounting for more than 60% of the FI for all considered models and scenarios. Precipitation is of particular importance, because it is the main driver of SMC variations for the studied soils. Moreover, “root zone soil” SMC has a very low FI value (close to 0) for all scenarios and models, as (i) target observations are superficial, so not necessarily correlated to deeper SMC (i.e., “root zone soil” SMC), and (ii) “root zone soil” SMC, through the assimilation of SMAP observation, turns into a land surface model with its own uncertainties. It can be explained with the difference between “root zone soil” depth (0–100 cm) and SMC’s use as a target (0–10 cm). Actually, no clear relationship between top layer (0–10 cm) and in-depth layer (0–100 cm) SMC is observed across semi-arid regions, as the top layer SMC is expected to dry faster than the in-depth layer.

Even if S1 had been previously used to retrieve SMCs across different regions [9,36,37,38], only three (four) were selected for the modelling using scenario-2 (scenario-3 and -4). According to their equations (Table 2), all PIs are quite comparable, and therefore combining them is unlikely to offer any additional benefits.

Similarly, 11 (15) TIs were selected for the modelling using scenario-2 and -3 (scenario-4), with a majority of TIs based on VV. The dominance of VV TIs is in line with previous studies highlighting a higher SMC sensitivity to VV than to VH polarization [31,34,66]. It is worth mentioning that same TIs are selected for all scenarios, with some TIs (VH imcorr1, VH imcorr2, VV corr) considered twice in scenario-4 (i.e., derived from both the single scenes and composite images).

Finally, even if S1 independent features present low FI values (Figure 4), put together, they contribute approximately 7% to 39% in the modelling process (Figure 4). This indicates that S1 improves SMC estimates at higher spatial resolution (i.e., 20 m). Actually, IMERG and SMAP bring relevant information for the SMC temporal dynamic, whereas S1 brings relevant information for the SMC spatial dynamic.

4.3. Soil Moisture Estimates Evaluation

The comparison of scenario-1 with scenario-2, -3, and -4 shows that the integration of S1 features quite substantially improves the SMC estimates. Indeed, for each model, higher R² and lower RMSE are obtained with scenario-2, -3, and -4.

Except for the DT model, the consideration of S1 features (scenario-2, -3, and -4) systematically improves the SMC estimates (

Table 4. Performance of ML models with (SF) and without (TF) feature selection.

Model	Metric	Scenario-1 (Training/Validation)	Scenario-2 (Training/Validation)	Scenario-3 (Training/Validation)	Scenario-4 (Training/Validation)
RF	R2	0.56/0.73	0.90/0.82	0.90/0.79	0.90/0.84
	RMSE (%)	4.96/3.49	2.35/2.86	2.36/3.09	2.35/2.67
XGB	R2	0.57/0.73	0.95/0.79	0.96/0.73	0.96/0.84
	RMSE (%)	4.96/3.47	1.52/3.05	1.49/3.49	1.49/2.67
GB	R2	0.57/0.73	0.95/0.81	0.95/0.79	0.95/0.86
	RMSE (%)	4.96/3.47	1.70/2.92	1.77/3.05	1.63/2.55
DT	R2	0.57/0.73	0.96/0.64	0.96/0.64	0.96/0.83
	RMSE (%)	4.96/3.47	1.49/4	1.49/4	1.49/2.80

Where TF = all features; SF = selected features with RFEcv and VIF.

). Even if composite images are generally recommended to minimize local noise [35,67,68], the results show that more reliable SMC estimates are obtained when considering the single scenes (scenario-2) rather than the composite images (scenario-3) (Table 4). Yet, the combination of the two (scenario-4) provides the most reliable SMC estimates for all considered models showing the complementarity of single scenes and composite images.

When comparing scenario-1 with scenario-4, an R² increase of approximately 15%, 13.5%, 16%, and 12% is observed at the validation step for the RF, XGB, GB, and DT model, respectively. Similarly, an RMSE decrease of 15%, 14%, 18%, and 10% is observed at the validation step for the RF, XGB, GB, and DT model, respectively. These results underline the effectiveness of the S1 integration in the SMC mapping at high spatial resolution. Overall, based on scenario-4, all models provide reliable SMC estimates (R² ≥ 0.83 and RMSE ≤ 3.15%), with GB performing the best (R² = 0.86 and RMSE = 2.55%).

Scenario-1’s good fit (R² ≥ 0.73) is due to an outlier compensation effect (Figure 5a,c,e,g) because the low spatial resolution of SMAP and IMERG (11 × 11 km) does not allow capturing SMC spatial variation occurring at the agriculture field level, as both considered plot fall in the same pixel. This is clearly observable through the horizontal line that corresponds to the same field sample collection date, for which SMAP and IMERG attribute the same value to all collected soil sample, whereas all these samples present very contrasted SMC values (Figure 5a,c,e,g). This aspect is considerably attenuated with the integration of S1 features (scenario-2, -3, and -4) (Figure 5). Actually, the high spatial resolution of S1 features (20 × 20 m) allows capturing SMC variation occurring for the same date at the agriculture field scale level and therefore the removal of the horizontal line (especially between 10% and 17.5% SMC, Figure 5b,d,f,h).

This spatial discrepancy in between scenario-1 and scenario-2, -3, and -4 explains the overfitting issue observed in scenario-1 with lower (higher) R² (RMSE) scores observed during the training step than during the validation step. Actually, the coarse spatial resolution of scenario-1 features (11 × 11 km) cannot capture the SMC spatial variability across the studied area. As a result, for each field visit, a unique combination of “root zone SMC”, “top layer SMC”, and “total precipitation” is attributed to all the SMCs observed that day at the different sampled points. As a result, when testing the models, the targeted SMC to predict has high probability to be included in the SMC range used in the training step, leading to an overfitting issue. This problem is avoided in scenario-2, -3, and -4 due to the consideration of the 20 m spatial resolution feature (S1), insuring a better discretization of the SMC spatial variability.

4.4. Soil Moisture Mapping

Figure 6 shows SMC maps derived from scenario-1 and -4 for 4 September 2023. It is worth mentioning that, as SMC maps derived from scenario-1 are similar for all considered models, Figure 6 only shows SMC maps obtained with the RF model. Similarly, as scenario-4 is more reliable than scenario-2 and -3 (Table 4, Figure 5), Figure 6 only shows maps obtained with scenario-4.

Due to SMAP’s and IMERG’s low spatial resolution (11 × 11 km), SMC maps based on scenario-1 show a uniformity in the study area (Figure 6a). Conversely, SMC maps based on scenario-4 present SMC variability in space that confirms the benefit of S1 features’ high spatial resolution (20 × 20 m) for the redistribution of SMAP SMC estimates (i.e., top layer soil moisture) at the sub-grid scale.

Although the models generated comparable metrics during the validation step (Table 4), significant differences are observed in the spatial distribution of the predicted SMCs. XGB and DT produced a more homogeneous SMC mapping in the agricultural plots, with values ranging between 6 and 8%. In contrast, RF shows a more marked variability in SMC, identifying sectors with up to 11% moisture. GB, on the other hand, reports a more balanced SMC distribution and greater agreement with reference SMC (Figure 6).

5. Discussion

Despite S1’s opportunity to substantially increase SMC estimates spatial resolution and accuracy (Figure 5 and Figure 6, Table 4), its 12-day revisit time does not allow daily monitoring. In this context, several studies combine daily MODIS optical/infrared and/or land surface temperature (LST) estimates to achieve continuous SMC estimates [18,25,26,27]. However, these MODIS datasets are available at coarse spatial resolution (i.e., 500 m) and for clear sky conditions. Therefore, high spatial (i.e., 20 m) and temporal (i.e., daily) resolution SMC mapping from satellite still represents a challenge to investigate.

Despite its low spatial resolution (11 km), IMERG precipitation estimates play a fundamental role in SMC mapping (Figure 3). Therefore, satellite-based precipitation estimates available at higher spatial resolution should considerably improve SMC mapping. As available satellite-based precipitation estimates across the region are only available at kilometric spatial resolution [69,70], a spatial downscaling procedure should be considered to obtaina precipitation estimates at high spatial resolution. In this line, the method proposed by [47] that integrates MODIS cloud optical and microphysical properties to improve IMERG spatial resolution from 11 km to 1 km could be used as a guideline. On the other hand, evapotranspiration (ET) is highly correlated with SMC in areas with poor vegetation cover [71]. Therefore, the integration of satellite-based ET estimates in the presented modelling approach should improve SMC estimates. Yet, satellite-based ET estimates rely on optical/IR data (i.e., SEBAL [72,73], SSEBop [74], METRIC [75]) that are only available for clear sky conditions, preventing daily monitoring.

Topography plays a crucial role in SMC dynamics [76], as it provides essential information for identifying depressions and areas prone to water accumulation, which in turn influence soil moisture levels. Numerous studies have demonstrated the benefits of incorporating topographic data as an independent variable in machine learning (ML) models for soil moisture mapping [77,78,79]. A study conducted in the UK found that using a high-resolution digital elevation model (DEM) derived from drone imagery significantly improved the reliability of ML models for predicting soil organic carbon in plowed fields [80]. Considering that microtopography in plowed areas is recognized as a key factor influencing the spatial distribution of soil moisture [81,82], incorporating such a detailed DEM could substantially enhance the performance and reliability of SMC predictions. It is worth mentioning here that freely available DEMs (i.e., SRTM) are not suitable for such a purpose. Indeed, these DEMs were retrieved from a single date, so the topography at that time might differ from the present one, especially across agriculture plots where plough practices considerably change microtopographic features.

In this study, the RFE wrapper features selection has been selected based on previous studies reporting on its efficiency [35,64]. However, other wrapper features selection such as the Genetic Algorithm (GA) or Sequential Feature Selector (SFS) have also proven their efficiency. Actually, the reliability of machine learning models (e.g., RF and Support Vector Machine (SVM)) significantly increases when using the GA or SFS with GA, being more suitable than SFS for SVM models [67,83]. Similarly, by default, hyper-parameter values have been used in this study for all the considered models (i.e., RF, XGB, GB, DT). However, hyper-parameters tuning techniques such as the grid-search function could be used to increase model performance and prevent model overfitting or underfitting [83,84,85,86,87]. In this context, future studies should consider combining hyper-parameter tuning with different feature-selection processes (i.e., RFE, GA, and SFS) to ensure an as efficient as possible modelling set-up.

Finally, this study relies on the consideration of single models. However, recently, machine learning model stacking strategies have been successfully applied to downscale SMAP SMC estimates to 1 km [38] and to estimated SMC at 30 m spatial resolution [88]. Both studies show that the stacking modelling approach outperforms the single modelling approach [38,88]. In this context, a stacking modelling approach integrating the above-mentioned datasets (i.e., high spatial resolution satellite-based precipitation, evapotranspiration, and digital elevation models) constitutes a very promising approach to substantially improve SMC mapping.

It is worth mentioning that S1 polarizations are not only sensitive to SMC but also to soil roughness and vegetation cover [89,90]. As the models were calibrated and validated across two agriculture plots with intrinsic soil roughness and vegetation cover, these models can only be used to estimate SMC across areas with similar features (i.e., soil roughness, vegetation cover). Indeed, different soil roughness and/or vegetation cover will result in different S1 polarization, even for a similar SMC range as the one considered in this study. As these models were not trained for such combinations, they will inevitably fail to retrieve SMC in different areas. To minimize such inconsistency, additional observations from areas with different soil roughness and vegetation cover should be added to the learning database.

6. Conclusions

This study assesses the integration of several remote sensing datasets (i.e., SMAP, IMERG and S1) in advanced machine learning techniques for Soil Moisture Content (SMC) mapping at high spatial resolution (20 m). For this task, different machine learning models (Random Forest—RF, Extreme Gradient Boosting—XGB, Gradient Boosting—GB, and Decision Tree—DT) are considered. The key findings can be summarized as follows:

-

Top-layer SMC derived from L-band sensors (i.e., SMAP) and precipitation (i.e., IMERG) are predominant factors in SMC monitoring.

-

Even if accurate SMC (R² ≥ 0.73; RMSE ≤ 3.15%) can be obtained considering only precipitation and top-layer SMC, the coarse spatial resolution (i.e., 10 km) of these datasets cannot capture SMC spatial variation at the agriculture plot scale.

-

Adding S1-based features to IMERG precipitation and SMAP top-layer SMC substantially improved SMC estimates derived from all considered models, with an R² (RMSE) increase (decrease) ranging from 12% to 16% (10% to 18%) depending on the considered model (with scenario-4).

-

S1-based features allow the spatial downscaling of SMC estimates obtained through SMAP and IMERG from 10 km to 20 m. In this process, S1 features derived from single scenes alone lead to more reliable SMC estimates than the consideration of S1 features derived from composite images alone, and the combination of the two provides the most reliable SMC estimates for all considered models.

-

Among the considered models, the GB model (with scenario-4) achieved the highest reliability, with an R² and RMSE of 0.86 and 2.55%, respectively.

The significance of this research lies in its innovative application of remote sensing data and machine learning to address the challenges of SMC estimation at the agriculture plot level to provide a valuable tool for precision agriculture. By effectively downscaling SMAP data using Sentinel-1 (from 10 km to 20 m), our approach not only enhances spatial resolution but also improves the reliability of SMC estimates. The practical applications of these findings may support stakeholders and policymakers to make informed decisions about water resource management (i.e., irrigation) towards sustainable agriculture management strategies. The proposed framework can serve to assess the impacts of climate variability on SMC dynamics to provide complementary insights to support climate-resilient agricultural practices. Even if this study can serve as a blueprint for similar applications in different regions, further research should focus on refining this method by considering (i) the same independent features (i.e., precipitation, top layer SMC) at a higher spatial resolution, (ii) additional independent features, controlling SMC variability in space and time (e.g., evapotranspiration, soil properties), and (iii) more complex machine-learning approaches (e.g., stacking).