Soil Moisture Prediction Using Remote Sensing and Machine Learning Algorithms: A Review on Progress, Challenges, and Opportunities

Abstract

Machine learning (ML) has gained significant attention for unraveling the complex, nonlinear relationships between soil moisture (SM) and various predictive variables, including remote sensing (RS; reflectance, brightness temperature, backscatter coefficients) and biophysical (topographic, soil, vegetation, and weather) variables. We reviewed the literature to extract and synthesize ML algorithms, reliable input features, and challenges in SM estimation using RS data. We analyzed results from 144 articles published from 2010 to 2024. Random forest (40 out of 67 studies), support vector regressor (13 out of 39 studies), and artificial neural networks (12 out of 27 studies) often outperformed other algorithms to estimate SM using RS datasets. Multi-source RS data often outperformed single-source data in SM estimation. Satellite-derived features, such as vegetation indices and backscattering coefficients, provided critical information on surface SM (SSM) variability to estimate SSM. For root zone SM estimation, soil properties and SSM generally were more reliable predictors than surface information derived solely from RS. Two recent advances—the use of semi-empirical models and L-band SAR to mitigate vegetation effects, and transfer learning to improve model transferability—have shown promise in addressing key challenges in SM estimation.

1. Introduction

Soil moisture (SM) accounts for just 0.05% of global fresh water [1], but it plays an outsized role in our world through its influence on such critical biosphere functions as river flows and flood timing [2], agricultural water management [3,4], terrestrial photosynthesis [5], and recharging of groundwater aquifers [6,7]. Accurate information on the spatial and temporal occurrence of SM is essential for understanding its influence on a wide range of critical zone processes.

SM refers to the quantity of water present within the unsaturated zone of the soil [8]. SM levels are impacted by a range of environmental factors, including precipitation [9,10], land-use patterns [11,12,13,14], soil type and depth [15,16,17,18], topography [19,20,21], and local atmospheric factors including radiation, temperature, humidity, and pressure [22]. The impact of these factors on SM variation varies based on the spatial scale of the study area, the temporal resolution, and the prevailing environmental conditions [23,24,25,26].

Various methods have been proposed for monitoring SM, which can be divided into two categories: pointwise and regional measurements (Figure 1). Pointwise measurements can be further differentiated into direct and indirect measurements. Direct measurement, such as the gravimetric method, which is independent of soil type and salinity, is standard and the most reliable method for estimating SM. However, its application is limited due to the impracticality of obtaining continuous spatial and temporal coverage and its destructive nature [27,28,29,30]. To overcome these challenges, numerous indirect SM measurement techniques have evolved, including neutron scattering, time domain reflectometer (TDR), frequency domain reflectometer (FDR), gamma attenuation, thermal dissipation, heat flux, capacitance, electrical resistance (gypsum block), tensiometer, and pressure plate. Despite their benefits, these methods are often expensive, require calibration for different soil types, pose health risks (e.g., radiation exposure, electrical hazards), and may not be suitable for fast-draining soils [31,32,33,34,35,36]. In response to these limitations, remote sensing (RS)-based SM estimation has emerged as a viable alternative, offering regular updates on a global scale at a lower cost [37,38]. A comprehensive review of traditional methods of SM estimation can be found in Vereecken et al. [39] and Zhang and Zhou [40].

Since the 1970s, various RS techniques have been introduced to analyze SM, utilizing different segments of the electromagnetic spectrum ranging from optical to microwave [34,38,41,42,43,44,45,46]. Each type of sensor comes with its own set of advantages and disadvantages, offering a range of capabilities in the measurement of SM. The wealth of information from optical, thermal, and passive and active microwave sensors has led to the development of numerous techniques aimed at estimating SM. Figure 1 provides a detailed representation of SM retrieval methods using these sensors’ outputs.

A decade ago, early RS research provided promising insights in estimating SM using optical and thermal infrared RS [40]. Although optical and thermal RS technologies offer high spatial resolution, widespread satellite availability, and mature technology, they often face challenges in retrieving SM due to vegetation interference, cloud contamination, night effects, poor temporal resolution, and atmospheric effects [47,48,49,50]. To overcome these issues, microwave RS techniques have been adopted, offering advantages such as all-weather observation capabilities and robust physical foundations [34,51,52].

Since 1978, passive and active microwave satellite sensors have been delivering valuable global-scale surface SM (SSM) data [53,54,55]. Some of these satellites do not provide SM directly and require the application of various techniques to convert backscatter coefficients, incident angles, polarization, temporal changes in backscatter, and interferometric coherence into SM estimates. Various methods have been developed to estimate SM using passive and active microwave techniques. Passive microwave techniques provide promising SM estimation results over bare soil surfaces with high temporal resolution and are unaffected by cloud cover or diurnal conditions. On the other hand, active microwave techniques offer finer spatial resolution and are also unaffected by clouds and diurnal conditions. Nonetheless, they exhibit coarse temporal resolution, and the accuracy of SM estimation is affected by vegetation cover and surface roughness. Detailed discussions of SM estimation methods using passive and active microwaves, along with their advantages and disadvantages, can be found in the reviews conducted by Petropoulos et al. [34] and Karthikeyan et al. [56]. The commonly used microwave satellite for SM estimation is illustrated in Figure 2.

Although RS offers significant opportunities for SM monitoring, SM retrieval methods still face several challenges and limitations. The aforementioned methods are complex, requiring consideration of local meteorological conditions, a detailed understanding of microwave interactions with both bare and vegetated soils, and numerous ground parameters. Consequently, there is an evident need for a shift towards a more simplistic method using fewer inputs and less parameterization. Machine learning (ML) algorithms offer promising solutions to the challenges listed above. This review thoroughly examines the role of ML and RS in estimating SM, detailing the progress, challenges, and opportunities in this field.

2. Research Methodology

The overall review process is presented in Figure 3. The following research questions guided the review: How do different ML models perform across various geographies and with diverse input features? Which features are commonly used for estimating SM (SSM vs. root zone SM (RZSM), field scale vs. regional/global)? Which approach works better for SM estimation, using single or multiple RS data sources? How does vegetation cover impact model performance? How do models perform when applied to regions different from those where they were trained? What in situ measurement methods have been used globally to validate ML algorithms, and is there a preference pattern?

The search was performed in five databases (Web of Science, Google Scholar, Science Direct, IEEE Xplore, and Scopus). We used (“Soil Moisture” OR “Surface Soil Moisture” OR “Root Zone Soil Moisture”, AND (“Machine Learning” OR “Artificial Intelligence” OR “Deep Learning”) AND “Remote Sensing”) as search input. Irrelevant studies were removed using the exclusion criteria provided in

Table 1. Exclusion criteria for paper selection.

Criterion Number	Exclusion Criteria
1	Publications not written in English
2	Study outside the scope of research questions (not related to soil moisture estimation using remote sensing and machine learning)
3	Publication is a book, e-book, poster, survey
4	Published before 2010

. The remaining 144 publications were extracted and synthesized.

3. Results and Discussion

ML has been increasingly utilized for SM estimation through RS since 2010 using various approaches and algorithms (Figure 4). We collected 144 papers related to SM estimation based on ML and RS from 2010 to 2024. Over that period, publications with ML utilization increased with the most studies from Asia (68 studies, concentrated in China, Iran, and India), followed by North America (24 studies), Europe (21), Australia (5), Africa (3), and South America (1) (Supplementary Figure S1). Since 2019, there has been a growing emphasis on root zone SM (RZSM) estimation.

3.1. Suitability of ML Models in SM Estimation

Among the various ML models applied in SM estimation using RS data (Figure 5), random forest (RF) was the most frequently used model (67 studies, 47%), followed by support vector regression (SVR, 39 studies, 27%) and artificial neural networks (ANNs, 27 studies, 19%). Similarly, RF was most often reported as the best model (40 studies, 28%) followed by SVR (13 studies, 9%) and ANNs (12 studies, 8%). Other classical ML models (20 out of 55 studies), other neural networks (20 out of 36), and other neuro-fuzzy models (6 out of 10), all listed in the figure caption, were often reported as the best-performing models in the respective studies. In some studies, multiple models performed equally well, leading to ties. In such cases, all tied models were counted as best-performing models. Additionally, some studies relied on a single model from their preliminary analysis. The performance metrics of each model can be found in Supplementary Table S1. RF was the best-performing and most widely used model for SM estimation. Among neural network (NN) models, ANNs emerged as the most effective algorithm, achieving substantial recognition for their accuracy and reliability. Studies have used various performance metrics (R²: coefficient of determination, RMSE: root mean squared error, and MAE: mean squared error) to evaluate the ML models in SM estimation. This review synthesizes model performance metrics (e.g., R², RMSE, MAE) as reported in the original studies without recalculating them under a unified framework. As such, differences in R²/RMSE calculation methods may introduce inconsistencies when comparing results across studies. This methodological heterogeneity should be considered when interpreting the reported model performance.

Classical ML models (RF, SVR, and XGBoost) are most effective when working with structured, tabular datasets derived from RS indices and ground-based measurements [57,58,59,60]. Their simplicity, interpretability, and robustness to noise make them well-suited for scenarios where the spatial and temporal complexity of the data is relatively low. ANNs, while also operating on tabular inputs, are better equipped to capture nonlinear interactions among diverse input features, such as those from combined climatic, soil, and vegetation data sources [61,62]. CNNs, on the other hand, are particularly advantageous for image-type datasets (such as multispectral or radar imagery) due to their ability to extract spatial features automatically through convolutional layers [63,64,65]. This makes CNNs well-suited for tasks where spatial patterns (e.g., texture, land cover variation) are important. LSTM networks outperform other models in time-series applications by capturing long-range temporal dependencies, making them ideal for predicting SM based on historical sequences of weather or vegetation indices [66,67,68].

Traditional single ML models often struggle to capture the complex variability of SM. One major limitation of single ML models is the bias-variance tradeoff. For example, linear regression-based models tend to have high bias, potentially oversimplifying the relationship between predictors and the target variable. In contrast, decision tree-based and NN models are often high-variance, making them prone to overfitting. To address these limitations, recent studies have explored hybrid models, which combine multiple ML models to take advantage of their respective strengths while mitigating their weaknesses.

Lv et al. [69] compared the performance of individual models, such as RF, SVR, and Backpropagation Neural Network (BPNN), against a hybrid approach that combined these models using ordinary least squares techniques. Their findings indicated that the combined models significantly outperformed the individual models in SM estimation. Similarly, Moosavi et al. [70] conducted a study in Iran and found that a hybrid model integrating convolutional neural network and long short-term memory (CNN-LSTM) yielded superior results compared to individual models. Numerous studies across various regions have consistently demonstrated that combining models substantially improves SM estimation accuracy [71,72,73,74,75,76]. Hence, future research should prioritize the implementation of hybrid models to enhance SM estimation accuracy rather than relying solely on individual models.

3.2. Satellite/UAV/Ground-Based Observation for SM Estimation

Out of 144 studies, the most widely used methods were Sentinel-1 (36 studies) followed by Moderate Resolution Imaging Spectroradiometer (MODIS; 35 studies), Sentinel-2 (27 studies), Soil Moisture Active Passive (SMAP; 26 studies), Landsat (20 studies), and Unmanned Aerial Vehicle (UAV; 19 studies) (Figure 6). The details of the spatial and temporal resolution of all RS products used in SM estimation can be found in Supplementary Table S2. These sensors differ considerably in spatial resolution, revisit frequency, and soil depth sensitivity, which directly affect their estimation accuracy and scale of application. Sentinel-1 uses C-band Synthetic Aperture Radar (SAR), allowing it to penetrate cloud cover and operate in all weather conditions [77]. SAR backscatter is highly sensitive to soil dielectric properties, which are directly influenced by SM content [78]. Backscattering coefficient with different polarization modes, incident angles, and Polarimetric Decompositions are commonly used features from SAR [79,80,81,82]. Although SMAP was launched in 2015, it has been widely used in SM estimation due to its ability to directly provide SSM using L-band microwave radiometry, which is less affected by atmospheric conditions and vegetation cover compared to optical and thermal sensors [54]. However, its coarse spatial resolution (~36 km) limits its application in fine-scale mapping, although its high revisit frequency (2–3 days) and surface sensing (~5 cm) make it valuable for large-area monitoring. Many studies have used SMAP-derived SSM for RZSM estimation using ML models and reported enhancement of SM prediction using SSM from SMAP [83,84,85]. Optical sensors, MODIS, Sentinel-2, and Landsat provide information on vegetation conditions, land-surface temperature, and evapotranspiration, making them suitable for SM estimation. MODIS offers daily observations but at coarser resolution (250–1000 m), while Sentinel-2 and Landsat provide finer resolution (10–30 m) but with longer revisit cycles (5 to 16 days). However, their inability to penetrate clouds or vegetation canopies make them less reliable for direct SM estimation, necessitating integration with microwave sensors or models for improved accuracy [40,47,48,49].

UAVs are widely used to estimate SM at field scale due to its high spatial and temporal resolution, flexibility, and real-time or near-real-time data collection for site-specific monitoring. Multispectral, thermal, and RGB cameras have been widely used to derive different spectral indices from UAVs to estimate SM [86,87,88,89]. However, UAVs are limited by short flight times, dependence on weather conditions, and challenges in processing large volumes of data for accurate SM estimation [90]. These differences in spatial resolution, revisit frequency, and soil depth sensitivity underline the importance of selecting appropriate RS products based on the spatial scale, crop cover, and temporal dynamics of the study area.

3.3. ML Model Performance on Single Source and Multi-Source Data

Relying on a single data source to estimate SM often introduces limitations, as factors such as vegetation cover, topography, soil properties, and meteorological conditions influence SM variability. Researchers increasingly integrate multi-source data, combining different RS variables with geophysical and meteorological data to improve ML model performance in SM estimation.

Cheng et al. [91] estimated SM in China using multispectral, thermal, and RGB datasets. They found that incorporating thermal and RGB data into multispectral datasets increased the coefficient of determination (R²) from 0.69 to 0.78. Adab et al. [92] estimated SM in Iran using Landsat reflectance values and ML models across diverse land-cover types. They concluded that SM is not solely a function of reflectance values but is also significantly influenced by land-cover characteristics. Eroglu et al. [93] used CYGNSS (Cyclone Global Navigation Satellite System) reflectivity to estimate SM in the USA and Australia, showing that reflectivity is influenced not only by SM but also by geophysical factors such as vegetation canopy, topography, and soil texture, indicating the need for multi-source data in SM estimation. Guo et al. [94] estimated SM in China by combining ultra-wideband radar echoes with multispectral vegetation indices, showing that incorporating multispectral RS data significantly enhanced the convolutional neural network (CNN) model’s performance, increasing R² from 0.74 to 0.92. Zhang et al. [59] estimated SM using raw Landsat bands combined with multi-source data, including SSM from ERA5, precipitation, digital elevation model, and soil texture. Their findings showed that incorporating these additional datasets improved SM estimation, increasing R² from 0.79 to 0.95. Zhu et al. [62] estimated SM across the Chinese Loess Plateau and found that incorporating soil physical and hydraulic properties alongside RS variables significantly improved ML model performance, increasing R² from 0.10 to 0.62. Several studies conducted in different regions of the world have shown that incorporating multi-source data enhances the performance of ML models in SM estimation [95,96,97,98]. Hence, future studies could use the benefits of multi-source data to enhance SM prediction accuracy using ML models.

3.4. Important Features

The choice of input features in ML models plays a crucial role in determining the accuracy of SM estimation. Commonly used input features include backscattering coefficients, incidence angles, and brightness temperatures from microwave sensors; vegetation and thermal indices from optical sensors; climatic variables; soil properties; and topographic variables (Table 2 and Table 3). The selection of these variables largely depends on the scale of the study (global/regional or field scale) and the soil depth being considered (surface or root zone). We examined which input features are frequently used in SSM estimation, RZSM estimation, and across different scales (global/regional and field scale). Figure 7 illustrates the top 10 features used in SM estimation using ML models across different scales and soil depths.

Table 2. Surface Soil Moisture (SSM) estimation studies: detailed overview of study areas, machine learning techniques, data sources, and input features.

Author	Study Area/Scale	ML Technique	Satellites/UAV	Input Features	In Situ SM Measured Method	In Situ SM Depth (cm)	Best ML Model
Ahmad et al. [99]	USA, Regional	SVM, MLR, ANN	TRMM, AVHRR	Incident angle, σ°, NDVI	Electrical resistance	10	SVM
Araya et al. [100]	USA, Regional	ANN, SVR, RVR, RF, BRT	UAV	Precipitation *, red band *, ET *, TPI *, curvature, NIR, green band, flow accumulation, slope, aspect, direction	TDR	4	BRT
Guan et al. [57]	USA, Field	RF	UAV	NDVI, NDRE *, GNDVI, CIG, VARI *, raw multispectral bands, crop type, drainage condition	GPR	18	RF
Khedri et al. [101]	USA, Regional	SVR	AIRSAR	Coherent and non-Coherent decomposition of PolSAR images	-	-	-
Liang et al. [102]	USA, Regional	GA-BPNN	GNSS-IR	Relative phase	-	SSM	-
Nabi et al. [103]	USA, Global	CNN	CYGNSS, SMAP, MODIS	Analog power, effective scattering area, BRCS, incident angle, peak reflectivity, NDVI, VWC, slope, water percentage, elevation, clay, silt.	-	SSM	-
Ren et al. [104]	USA, Regional	LS-SVM	GNSS-IR	Relative phase of multi-satellite	-	-	-
Senyurek et al. [105]	USA, Field	RF	UAV	Carrier-to-noise density ratio (C/N0), elevation, azimuth angle, NDVI, VWC	HOBO SM probe	SSM	-
Torres-Rua et al. [106]	USA, Field	RVM	Landsat 7	Blue, green, red, NIR *, SWIR1, SWIR2, 7, Red/NIR *, NDVI, GNDVI, BNDVI, NBR, brightness, greenness, wetness, haze, NDWI, LAI *, α, ε, energy balance products from METRIC *	Decagon GS3 sensor	SSM	-
Xu et al. [107]	USA, Regional	GRNN	SMAP	SMAP brightness temperatures	-	10	-
Akhavan et al. [79]	Canada, Field	GRNN, NN, SVR	Sentinel-1	VV, VH, entropy, alpha, GLCM	-	5	GRNN
Dabboor et al. [108]	Canada, Regional	ANN, DT, SVM, GPR, Ensemble Learning	RADARSAT	Radar backscattering coefficients (RH and RV), incident angle	Stevens HydraProbe	5	GPR
Jiaxin et al. [109]	Canada. Field	ERT, XGBoost, GPR, GRNN	Sentinel-1, Sentinel-2, Landsat 8, PLSA	VV, VH, VH/VV, RIA, PRVI, DpRVI, H, α, A, blue, green, red, NIR, SWIR1 *, SWIR2 *, TAv, TAh, MPDI *, STT *, SOM *, SR	-	5	GPR
Chen et al. [110]	Canada, Field	SVM, RF, GBM	RADARSAT-2, Sentinel-2	VH, VV *, HH *, HV, Polarimetric parameters	Theta-probe SM sensor	5	RF
Lee et al. [111]	Canada, Regional	DNN	Sentinel-1, Sentinel-2	VV, incidence angle *, elevation *, Sentinel-2 bands 2–8, band 8A, band 11, band 12, NDVI, EVI, SAVI, MSI, and NDWI, crop type *, month *.	Stevens HydraProbe	5	-
Liu et al. [112]	Canada, Field	RF, SVR, DNN, GRNN	Landsat 8, Sentinel-2, Sentinel -1, Gaofen-1 RADARSAT-2	SRWI *, NDVI, DVI, EVI *, MSAVI, CRI1, MRENDVI, MRERSR *, S2REP, TVI, IRECI, MTIC, MTCC, CIred edge, NDVIre3 *, MCARI, MCARI2, TCARI, SWIRI2 *, MSI, NDMI, NMCI, VV *, VH *, VV-VH *, VV/VH, incident angle	Probe-based and the core-based	5	DNN
Zhang et al. [113]	Canada, Field	Quantile regression forest	UAV SAR	SERD, DERD, normalized, reference and actual backscattering, HH, VV, HV, coefficients, surface, double-bounce, volume	FDR	6	-
Filgueiras et al. [114]	Brazil, Field	LR, RF, PLS, PCA, GBR, Cubist	MODIS	Irrigation depth, Kc, ETo, solar radiation, NDVI, simple ratio	Soil water balance equation	SSM	RF
Cheng et al. [91]	China, Field	PLSR, KNN, RF, BPNN	UAV	CC, MSR, NDVI, OSAVI, RV1, RV12, SAVI, SIPI, TVI, EVI, MCARI, NPCI, GI, GNDVI, SRPI, NPCI, NDVIgb, PSRI, VARI, CIVE, VA, NRCT, TVDI, VA, CO, EN, COR, ME, HO, DS, SM.	TDR	10, 20	RF
Ge et al. [115]	China, Field	RF, ELM	UAV	NDVI *, NDVI705 *, RVI, NDCI, GNDVI, OSAVI, NDRE, mNDVI705, VOG1, VOG3, VOG2, CARI, MTVI1, TVI, DVI, RDVI, SPVI, WI/NDVI, EVI, NVI, MSAVI *, WI, REP, PRI, MTVI2, TCARI2, TCARI/OSAVI, MCARI/OSAVI, TCARI1, MCARI	Oven drying	10	RF
Ge et al. [116]	China, Field	XGBoost	UAV	Difference index (DI), ratio index (RI), normalized difference index (NDI)	Gravimetric method	10
Guo et al. [94]	China, Field	SVM, GRNN, CNNR	UAV	NDVI, DVI, MSAVI, rectified average value (Av), kurtosis (Ku), root mean square (Rm), peak factor (C), pulse (I)	TR-6D Soil Thermometer	6	CNNR
Guo et al. [117]	China, Field	MLR, BPNN, SVM	Sentinel-1, Sentinel-2	VV, VH, salinity index, NDVI, EVI, MSAVI, NDVIre, BI, intensity index, PDI, SMMI	Oven drying	20	SVM
Han et al. [118]	China, Regional	CART	MODIS	LST, ET, NDVI, precipitation, soil texture, elevation, Soil AWC, FC	-	10
Hou et al. [119]	China, Field	RF, SVR, ANN	Sentinel-1	VV, VH, incident angle, polarimetric parameters	TDR	7	RF
Hu et al. [120]	China, Field	LR, BPNN, GA-BPNN	GNSS-R	Signal to noise ratio	Hygrometer	7.5	GA-BPNN
Li and Yan [121]	China, Regional	RF, XGBoost, LightGBM, CatBoost, DNN, CNN, GRU, Stacking	Sentine-1, Sentinel-2	VV, VH, NDVI, GNDVI, MNDWI, NBRI, RVI, SATVI, SAVI, B2, B3, B4, B8, B11, B12, elevation *, latitude *, longitude *, month_cos, month_sin	Decagon 5TM	5	Stacking
Li et al. [76]	China, Regional	RF, GBR, RF-GBR	Sentinel-2	Band 1–8	Oven drying	20–30	RF-GBR
Li et al. [122]	China, Regional	Deep Forest, RF, GRNN, GBRT, SVM, KNN	Sentinel-1, Sentinel-2	VV *, VH, NDVI *, NDWI, clay *, sand, BD, land cover, fractional cover, longitude *, elevation *	Decagon 5TM ECH2O	5	Deep Forest
Zhang et al. [123]	China, Regional	BPNN, XGBoost, LSR, RF	MODIS, SMAP, ERA-Interim, ERA5-Land	NDVI, LST *, Land cover, reflectance, albedo, latitude, longitude, SMAP_TBH *, SMAP_TBV *, ERA_SR *, ERA_Eva, ERA_Runoff, ERA_TP, elevation, T_Clay, T_Sand, T_Silt, SOC, DOY	-	SSM	BPNN
Liu et al. [124]	China, Regional	RF, Extra Tree, LR	GF-1, Landsat 8, GF-4	COSRI *, DVI, EVI, GDVI, GLI, FNDVI, GOSAVI, GRVI, GSAVI, IPVI, MSAVI2, NDVI, NNIR, NR, OSAVI, RI, RVI, TVI, VARI, WDRI	Decagon 5TM SM sensor	3, 10, 20	Extra Tree
Zhang et al. [125]	China, Field	Copula Quantile Regression	RADARSAT-2	HH, VV, HV, VH, entropy, alpha, anisotropy, adaptive non-negative eigenvalue decomposition (odd, Dbl, Vol)	SPADE and Hydra Probe	4	-
Lv et al. [69]	China, Regional	RF, SVM, BPNN, Stacking	Landsat-9	NDVI, RVI, DVI, SAVI, EVI, GVI, SI-T, S1, S2, S3, NDSI, bright, green, wet, albedo, NDWI, NDBI, elevation	Oven drying	10	Stacking
Wang et al. [126]	China, Regional	PR, Ridge regression, LASSO, Elastic net, RF	Fengyun-3C, MODIS	FY-3C SM *, NDVI *, months, latitude *, longitude, elevation, clay, sand, silt	-	10	RF
Mu et al. [127]	China, Regional	LR, BPNN	GF-2, GF-3, GF-5	Bands normalization of GF-2, GF-3, and GF-5	Oven drying	10	BPNN
Wang et al. [71]	China, Regional	CART, RF, GBDT, ERT, Stacking	Landsat, Sentinel -1	VV *, VH, NDVI, EVI, SAVI, RVI, IPVI, LST, albedo, elevation, slope, SLIA, TVDI *, VSDI, VSWI, NIR, SWIR1, SWIR2, surface temperature *	Oven drying	10	Stacking
Qiao et al. [128]	China, Field	RBF, Multivariate RBF	Ground penetration radar	GPR signals	-	SSM
Wang et al. [64]	China, Field	RF, GRNN, CNN, SSA-CNN	Sentinel-1, Sentinel-2	Incident angle, VH, VV, scattering entropy, inverse entropy, α, eigenvalues (λ1 *, λ2), Zs), NDVI *, NDWI, MSI *, FVI *	TDR	3.8	SSA-CNN
Taghavi-Bayat et al. [129]	China, Regional	DRF	Sentinel-1	S-1:TFGVV *, TFGVH, LIA, SAR-VI: PRVI, DpRVI *, SAR-VI: DPSVI, mDPSVI *, RVI *, elevation, slope, TRI, TI, flow direction, curvature, SM, soil temperature, diurnal temperature range, minimum temperature, maximum temperature, daily rainfall events, days of year *	-	5	-
Tao et al. [72]	China, Regional	CatBoost, RF, GBDT, Stacking	MODIS	VCI, NDWI, NDVI, EVI, TCI *, CWSI *, VSWI, elevation *	-	SSM	Stacking
Jiang et al. [130]	China, Regional	RBFNN, LSTM, RNN, PCA-LSTM	-	RH, Ta, soil temperature, U, wind direction, rainfall, light intensity	-	SSM	PCA-LSTM
Xiaoxia et al. [131]	China, Regional	BPNN	Fengyun-3B, MODIS	LST, NDVI, albedo, NDVI, elevation	-	SSM	-
Xiang-yu et al. [132]	China, Field	GBR, RF, XGBoost	UAV	DI, NDI, RI, PI	-	10	XGBoost
Xu et al. [133]	China, Regional	GABP, SVR, RF	Sentinel-1, Sentinel-2, TerraSAR	Incident angle, VV, VH, HH, NDVI, NDWI	TDR	SSM	RF
Jiang et al. [134]	China, Regional	PCR, PLSR, BPNN	ASD Field Spec FR spectrometer	EVI, TVI, DSI, NDMI, SARVI	-	20	BPNN
Yang et al. [135]	China, Regional	DBN	Sentinel-1, Landsat 8	VV, VH, incident angle, EVI, OSAVI, elevation, slope, aspect	-	10	-
Yang et al. [136]	China, Regional	RF, ERT, XGBoost, LightGBM	MODIS, SMAP, TRIMS	NDVI, NDWI, DDI, SM, LST	-	10	LightGBM
Zhang et al. [75]	China, Regional	DI-Conv GRU, LSTM, Interp-ConvGRU, DI-LSTM	SMAP	SM from SMAP, precipitation, Ta, Rs, RH, U, sand, silt, clay content, BD, land cover, elevation	-	5	DI_Conv GRU
Zhang et al. [137]	China, Regional	RF, SVM, GA-BP	Sentinel-1, Sentinel-2	NDVI *, NDWI, RVI, MSI, WBI, FVI, Zs, average scattering angle *, θ *, cos(θ) *, sin(θ) *, VV, VH, inverse entropy, scattering entropy	TDR	SSM	RF
Zhang et al. [138]	China, Field	PLSR, KNN, RF	UAV	SAVI, NDVI, TCARI, VARI, OSAVI, SIPI, MCSARI, NDVIgb, PSRI, CIVE, MCARI/OSAVI, TCARI/OSAVI, NRCT, TVDI, GLCM features	Oven drying	10, 20	RF
Zhang et al. [139]	China, Regional	DFNN	VIIRS-RDR	Sixteen moderate resolution bands (M1–M16)	Gravimetric method	10
Zhan et al. [140]	China, Regional	MLR, AdaBoost, RF, GBM	Sentinel-2	MTCI, MNDVI, PSRI, RENDVI, S2REP, IRECI, OSAVI, SAVI red, MSR, soil textures	Oven drying	10	GBM
Chaudhary et al. [141]	India, Field	SVM, RF, MLP, RBF, WM, SBC, ANFIS, HyFIS, DENFIS	Sentinel-1	VH, VV	Steven’s HydraGo	SSM	SBC
Das et al. [142]	India, Field	Cubist, GBM, RF, Stacking	Landsat, Sentinel -1	shadowMin, shadowMax, NDVI, TVI, TTVI, SR, EVI2, CTVI, GEMI, RVI, MSAVI2, SAVI, NRVI, SLAVI, MSAVI, WDVI, DVI, B5_sre, Sigma0_VH, NDWI2, B6_sre, B4_sre, B2_sre, B3_sre, GNDVI, NDWI, SATVI, B7_sre *, NBRI *, LST *, MNDWI *, Sigma0_VV *	ML3 ThetaProbe	5	Stacking
Datta et al. [143]	India, Regional	LR, MLR, RF, KNN, SVM	Sentinel-1, Sentinel-2	VV *, VH	Oven drying	5	RF
Khose and Mailapalli [144]	India, Field	LR, KNN, RF, DT, SVR	UAV	NDVI, NDWI, TNDVI, Simple Ratio or Ratio Vegetation Index (RVI), SAVI, PDI.	Gravimetric method	1, 5	LR, SVR
Nijaguna et al. [73]	India, Regional	DMN-Bi-GRU	Sentinel-1, Sentinel-2	NDVI, GLAI, GNDVI, WDRVI	-	SSM	DMN-Bi-GRU
Pal and Maity [145]	India, Regional	RF, SVN, GP	Radar Imaging Satellite 1 (RISAT1)	HH *, HV, VH, VV, sand, silt, clay	Gravimetric method	5	GP
Singh and Gaurav [61]	India, Regional	ANN, GRNN, RBN, ERBN, GPR, SVR, RF, Boosting, RNN, RNNBDT, AutoML	Sentinel-, Sentinel-2	VV, VH, VH/VV, VH-VV, angle, NDVI, elevation, longitude, latitude	ML3 theta probe	5	ANN
Singh et al. [66]	India, Regional	LR, SVR, DT, RF, LSTM	SMAP	SSM, LST, soil texture, land cover, wind direction, U, surface pressure, dew point, Ta, precipitation	-	5	LSTM
Lee et al. [96]	South Korea, Regional	DNN	MODIS	OLR, Isolation accumulated precipitation (12 h, 24 h), NDVI, Ta, RH, Land cover, elevation, slope	SM probes	10	-
Lee et al. [146]	South Korea, Regional	RF, AutoML	IMERG, GK2A, VIIRS, LDAPS, SRTM	10- and 20-day cumulative standardized precipitation indexes (SPI10 * and SPI20 *), NDVI, DSR, Ta, LST, soil temperature, RH, LE, slope *, elevation *, TRI, aspect.	TDR	10	AutoML
Adab et al. [92]	Iran, Regional	RF, SVM, ANN, EN	Landsat 7, 8	LST *, blue *, green, red, NIR *, SWIR1, SWIR2	Capacitance Probe	10	RF
Asadollah et al. [74]	Iran, Regional	GBM, SVR, GBM-SVR	GLDAS, AMSR2, SMAP	AMSR2-C1, AMSR2-C2, ASMR2-X, SMAPL3, SMAPL4, GLDAS *	TDR	6	GBM-SVR
Bandak et al. [147]	Iran, Regional	RF, SVR, XGBoost, Extra tree	Sentinel-2, Landsat 8	Salinity index -S1, Salinity index -S2, Salinity index -S3, Salinity index -S4, Salinity index -S5, NDSI *, NDVI, SAVI, VSSI, NDWI *, Extended EVI, MNDWI *	Oven drying	10	Extra tree
Fathololoumi et al. [148]	Iran, Regional	RF, Cubist	Landsat 8	Elevation, aspect-sin, aspect-cos, catchment area, convergence index, maximum difference of mean elevation, general curvature, hill shade, MSP, MRRTF, LS, total curvature, albedo *, emissivity *, LST *, Incidence	Portable SM meter	SSM	RF, Cubist
Hemmati and Sahebi [63]	Iran, Regional	RF, SVR, MLP, CNN	Sentinel-1, Sentinel-2	VV, VH, incident angle, NDVI	Thetaprobe sensor	5	CNN
Karamvand et al. [149]	Iran, Regional	ELM	SMAP, MODIS	Precipitation, NDVI, LST	-	5	-
Moosavi et al. [150]	Iran, Regional	PSO-ANFIS, PSO-SVR	MODIS, Landsat 8	Evaporative fraction, TVDI, VTCI, STVDI, TVX	-	SSM	PSO-SVR
Moosavi et al. [151]	Iran, Regional	PSO-CMAC, PSO-GMDH	Sentinel-2	NDVI, NDWI, all bands from Sentinel-2, slope, aspect, elevation	TDR	5	PSO-CMAC
Moosavi et al. [70]	Iran, Regional	LSTM, CNN, CNN-LSTM	Sentinel-2	NDVI, NDWI, NDMI, BSI, elevation, slope, aspect, TWI, land use, soil properties	TDR	8	CNN-LSTM
Nouraki et al. [152]	Iran, Regional	MLR, CART, M5P, GBM, RF	Landsat	Blue, green, red, NIR, SWIR1, SWIR2, LST, NDVI, NSMI, NTR, SWI, sand, silt, clay, BD	-	10	RF, GBM
Tahmouresi et al. [153]	Iran, Regional	RF, GBM, XGBoost	SMAP, AMSR2, MODIS	Elevation, slope, sand, silt, LST, SSM, NDVI, precipitation	TDR	5	XGBoost, GBM
Shokati et al. [154]	Iran, Field	M5 tree, RF, SMOreg, MLP	UAV	DI, RI, NDI, PI	Gravimetric method	5	RF
Pasolli et al. [58]	Switzerland, Field	SVR, MLP	Radiometer–scatterometer	VV, VH, emissivity polarizations, incident angle	Oven drying	SSM	SVR
Kseneman et al. [155]	Slovenia, Regional	FFBP Neural Network	TerraSAR-X	Backscatter coefficients	TRIME-PICO64	5	-
Pongrac and Gleich [156]	Slovenia, Regional	CNN	ALOS-2	-	-	SSM	-
Jia et al. [157]	Italy, Regional	RF, SVM	GNSSR	Reflectivity, elevation angle	-	SSM	RF
Notarnicola et al. [158]	Italy, Regional	SVR	ASAR WS sensor, MODIS, GEOtop	VV, VH, slope, elevation, aspect, land cover, soil type	-	SSM	-
Pasolli et al. [159]	Italy, Regional	SVR	RADARSAT-2, Envisat ASAR, MODIS	HH, HV, HV/VV, altitude *, incident angle, NDVI *, land use	TDR	5	SVR
Santi et al. [160]	Italy, Regional	ANN	Sentinel-1, SMAP	Brightness temperature, acquisition geometry, land use, NDVI	-	SSM	-
Portal et al. [161]	Spain, Regional	ANN	Sentinel-2, MODIS, ERA5, ESA CCI	Precipitation, reflectance, LST, elevation, slope, SM, land cover, NDVI, NDRE, EVI, GNDVI, SAVI, NDMI, MSI, NBRI, BSI, NDWI	-	10	-
Chen et al. [162]	Spain, Field	1D-CNN	Spectrophotometer	Spectral measurement from different soil samples	Oven drying	SSM	-
Abebrese et al. [163]	Czech Republic, Field	XGBoost	Sentinel-2	NDVI, NDRI, NDWI, DVI, RVI, MSI, elevation	EC-5 Sensor	5–10	-
Lendzioch et al. [164]	Czech Republic, Field	RF	UAV	RGI, VVI, VDVI, VARI, TGI, SI, SHP, SCI, SAT, NGRDI, NDTI, NDI, HI, GRVI, GLI, GLAI, ExG, ERGBVE, CI, BI, RI, HUE, NDVI, Slope, ProfCurv, PlnCurv, WEI, TRI (1, 4), VRM (1, 4), TPI, TWI, Temperature, GWL, SM.	Oven drying	SSM	-
Hajdu et al. [80]	New Zealand, Regional	RF	Sentinel-1	Backscatter, incident angle, NDVI, slope, aspect, roughness index, wetness index	Capacitance-based, AquaCheck sub-surface probes	10	-
Holtgrave et al. [165]	Germany, Regional	SVR	Sentinel-1, Landsat	VV, VH *, incident angle, NDVI *, height *, slope, and aspect	FDR	SSM	-
Morellos et al. [166]	Germany, Field	LS-SVM, Cubist	Spectrophotometer	VIS-NIR reflection	Oven drying	20	LS-SVM
Schröter et al. [167]	Germany, Regional	Fuzzy c-means sampling and estimation approach (FCM SEA)	LIDAR	Elevation	-	10	-
Schröter et al. [168]	Germany, Field	Fuzzy c-means sampling and estimation approach (FCM SEA)	RapidEye sensor	reNDVI, elevation, LAI	TDR	10	-
Özerdem et al. [169]	Turkey, Regional	GRNN	RADARSAT-2	VV, VH, HV, HH, entropy, anisotropy, alpha angle, volume scattering, odd bounce, double bounce	Gravimetric method	3–5	-
Şekertekin, Marangoz [170]	Turkey, Regional	MLR	Sentinel-1	VV, VH, incident angle	Oven drying	SSM
Usta [98]	Turkey, Regional	ANN	Landsat 8	EVI, SAVI, NDMI, MNDVI, MSI, SMI, TIRS 1, TIRS2, slope *, TWI *, ALT, TPI *, TRASP, EAST *, NORTH, SITE, CTI	-	0–10, 10–20	-
Huang et al. [171]	Austria, Regional	Deep belief network	Landsat, MODIS, SMAP, ERA-5	LST, NDVI, albedo, precipitation, SMAP SM, ERA-5 SM	-	5	-
Ayehu et al. [172]	Ethiopia, Regional	SCA, SVM	Sentinel-1, MODSI, SRTM	VV, VH, NDVI, elevation	-	SSM	SCA
Ayehu et al. [173]	Ethiopia, Field	LR, ANN	Sentinel-1, Landsat 7 and 8	VV, surface heights (hrms), surface correlation length (leff)	FDR	5	ANN
Attarzadeh et al. [174]	Kenya, Regional	SVR	Sentinel-, Sentinel-2	VV, GLCM Contrast, GLCM Correlation, GLCM Dissimilarity, GLCM Entropy, GLCM Homogeneity, GLCM Mean, GLCM Standard Dev, GLCM Angular, SAVI, NDVI, TSAVI, MSAVI, MSAVI2, DVI *, RVI, PVI, IPVI, WDVI, TNDVI, GNDVI *, GEMI *, ARVI, NDI45, MTCI, MCARI, REIP, S2REP, IRECI, PSSRa, LAI, FCOVER *, NDWI, NDWI2, BI, BI2, RI, CI	TDR	5	-
Dutta and Terhorst [175]	Australia, Regional	S-ANFIS, MLP, PNN, RBFN	CosmOz Cosmic Ray Sensor	Neutron counts from cosmic ray probe, rainfall, Ta, RH	-	0–5, 10–15, 25–30	S-ANFIS
Celik et al. [67]	Global	LSTM	Sentinel-1, Sentinel-, SMAP	VV, VH, SSM, soil textures *, elevation, slope, aspect, hill shade, Ta, ET, precipitation *	-	10	-
Nativel et al. [82]	Global	ANN	Sentinel-1, Sentinel-2	VV *, VH *, the classical change detection SSM index ISSM, incidence angle *, NDVI *, VH/VV ratio *	-	10	-
Jia et al. [176]	Global	XGBoost, ANN, SVM, RF	CYGNSS, SMAP	Reflectivity from CYGNSS, vegetation opacity, and roughness coefficients from SMAP	-	SSM	XGBoost
Rodríguez-Fernández et al. [177]	Global	Neural Network	SMOS	RH, Ta, ASCAT SM, LST, snow cover, snow temperature, brightness temperature	-	5
Senyurek et al. [178]	Global	RF	CYGNSS, MODIS	Reflective power, incident angle, trailing edge slope, NDVI, vegetation water content, elevation, slope, soil properties	-	SSM	-
Zhang et al. [59]	Global	RF, XGBoost	SMAP, ERA5, Landsat	Landsat 8 (5 VNIR bands, 2 SWIR bands, 2 TIR bands, solar elevation, angle), elevation *, slope, aspect *, sand *, silt, clay *, precipitation, surface skin temperature, ERA-5 SM *	-	SSM	XGBoost
Hamidisepehr et al. [179]	Global	20 different ML models available in MATLAB	Spectrometers	1024 wavelengths into 20 components using Partial Least Square regression.	-	0.64, 1.92, 3.20, 4.48, 5.76	Cubic SVM
Jia et al. [180]	Global	RF	GNSS-R	Signal-to-noise ratio, elevation angle	-	SSM	RF
Lei et al. [60]	Global	RF	CYGNSS, MODIS	CYGNSS reflectivity, trailing edge slope, specular point incidence angle, NDVI, vegetation water content, elevation, clay, silt	-	SSM	RF
Lei et al. [181]	Global	RF	CYGNSS, SMAP	Reflectivity, TES, incidence angle, elevation, NDVI, VWC, silt, clay	-	10
Nabi et al. [182]	Global	CNN	CYGNSS, SMAP, MODIS	Analog power, effective scattering area, BRCS VWC, NDVI, elevation, slope, water percentage, clay, silt	-	10
Roberts et al. [183]	Global	CNN	CYGNSS, SMAP	DDM, VWC, elevation, surface water occurrence	-	SSM	-
Wang et al. [184]	Global	ANN	Fengyun, SMAP	Brightness temperature, microwave vegetation index	-	5	-
Zhang et al. [185]	Global	RF	MODIS	LST *, LST difference, NDVI, EVI, API *, clay, sand, silt, latitude *, longitude *	-	SSM
Zhu et al. [186]	Global	Transfer Learning	SMAP, Sentinel-1, MODIS	VV, VH, incident angles, sand, clay, BD, elevation, slope, aspect, NDVI, Ta, precipitation, day of year	-	5	-
Li et al. [187]	China and USA	Deep and parallel model	Sentine-1, Sentinel-2	VV, VV+VH, incident angle, NDVI, elevation, land covers, soil properties	-	SSM	-
Fang et al. [188]	United States and Mexico	Fully CNN (FCNN)	MODIS, SMAP	NDVI, LST, SSM	-	10
Hegazi et al. [81]	Australia and Austria	CNN	Sentinel-1	Sigma0_VH and Sigma0_VV or Gamma0_VH and Gamma0_VV or Beta0_VH and Beta0_VV	Campbell Scientific water content reflectometers	20	-
Hegazi et al. [65]	Australia and Austria	CNN	Sentinel-2	B1-B12 bands (B1—Coastal aerosol, B2—Blue, B3—Green, B4—Red, B5—Vegetation red edge *, B6—Vegetation red edge, B7—Vegetation red edge, B8—NIR, B8A—Narrow NIR, B9—Water vapor, B10—SWIR—Cirrus, B11—SWIR *, B12—SWIR)	Campbell Scientific water content reflectometers and TDR	30	-
Eroglu et al. [93]	USA and Australia	ANN	CYGNSS	Surface reflectivity, TES, and SP incidence angle, NDVI, VWC, elevation, slope, h-parameters.	-	SSM	-
Tsagkatakis et al. [189]	USA and Switzerland	CNN	SMAP	SSM	-	SSM	-

In the input feature column, asterisk (*) indicates the most important features for surface soil moisture estimation. All the abbreviations can be found in the Supplementary Table S3.

Table 2. Surface Soil Moisture (SSM) estimation studies: detailed overview of study areas, machine learning techniques, data sources, and input features.

Author	Study Area/Scale	ML Technique	Satellites/UAV	Input Features	In Situ SM Measured Method	In Situ SM Depth (cm)	Best ML Model
Ahmad et al. [99]	USA, Regional	SVM, MLR, ANN	TRMM, AVHRR	Incident angle, σ°, NDVI	Electrical resistance	10	SVM
Araya et al. [100]	USA, Regional	ANN, SVR, RVR, RF, BRT	UAV	Precipitation *, red band *, ET *, TPI *, curvature, NIR, green band, flow accumulation, slope, aspect, direction	TDR	4	BRT
Guan et al. [57]	USA, Field	RF	UAV	NDVI, NDRE *, GNDVI, CIG, VARI *, raw multispectral bands, crop type, drainage condition	GPR	18	RF
Khedri et al. [101]	USA, Regional	SVR	AIRSAR	Coherent and non-Coherent decomposition of PolSAR images	-	-	-
Liang et al. [102]	USA, Regional	GA-BPNN	GNSS-IR	Relative phase	-	SSM	-
Nabi et al. [103]	USA, Global	CNN	CYGNSS, SMAP, MODIS	Analog power, effective scattering area, BRCS, incident angle, peak reflectivity, NDVI, VWC, slope, water percentage, elevation, clay, silt.	-	SSM	-
Ren et al. [104]	USA, Regional	LS-SVM	GNSS-IR	Relative phase of multi-satellite	-	-	-
Senyurek et al. [105]	USA, Field	RF	UAV	Carrier-to-noise density ratio (C/N0), elevation, azimuth angle, NDVI, VWC	HOBO SM probe	SSM	-
Torres-Rua et al. [106]	USA, Field	RVM	Landsat 7	Blue, green, red, NIR *, SWIR1, SWIR2, 7, Red/NIR *, NDVI, GNDVI, BNDVI, NBR, brightness, greenness, wetness, haze, NDWI, LAI *, α, ε, energy balance products from METRIC *	Decagon GS3 sensor	SSM	-
Xu et al. [107]	USA, Regional	GRNN	SMAP	SMAP brightness temperatures	-	10	-
Akhavan et al. [79]	Canada, Field	GRNN, NN, SVR	Sentinel-1	VV, VH, entropy, alpha, GLCM	-	5	GRNN
Dabboor et al. [108]	Canada, Regional	ANN, DT, SVM, GPR, Ensemble Learning	RADARSAT	Radar backscattering coefficients (RH and RV), incident angle	Stevens HydraProbe	5	GPR
Jiaxin et al. [109]	Canada. Field	ERT, XGBoost, GPR, GRNN	Sentinel-1, Sentinel-2, Landsat 8, PLSA	VV, VH, VH/VV, RIA, PRVI, DpRVI, H, α, A, blue, green, red, NIR, SWIR1 *, SWIR2 *, TAv, TAh, MPDI *, STT *, SOM *, SR	-	5	GPR
Chen et al. [110]	Canada, Field	SVM, RF, GBM	RADARSAT-2, Sentinel-2	VH, VV *, HH *, HV, Polarimetric parameters	Theta-probe SM sensor	5	RF
Lee et al. [111]	Canada, Regional	DNN	Sentinel-1, Sentinel-2	VV, incidence angle *, elevation *, Sentinel-2 bands 2–8, band 8A, band 11, band 12, NDVI, EVI, SAVI, MSI, and NDWI, crop type *, month *.	Stevens HydraProbe	5	-
Liu et al. [112]	Canada, Field	RF, SVR, DNN, GRNN	Landsat 8, Sentinel-2, Sentinel -1, Gaofen-1 RADARSAT-2	SRWI *, NDVI, DVI, EVI *, MSAVI, CRI1, MRENDVI, MRERSR *, S2REP, TVI, IRECI, MTIC, MTCC, CIred edge, NDVIre3 *, MCARI, MCARI2, TCARI, SWIRI2 *, MSI, NDMI, NMCI, VV *, VH *, VV-VH *, VV/VH, incident angle	Probe-based and the core-based	5	DNN
Zhang et al. [113]	Canada, Field	Quantile regression forest	UAV SAR	SERD, DERD, normalized, reference and actual backscattering, HH, VV, HV, coefficients, surface, double-bounce, volume	FDR	6	-
Filgueiras et al. [114]	Brazil, Field	LR, RF, PLS, PCA, GBR, Cubist	MODIS	Irrigation depth, Kc, ETo, solar radiation, NDVI, simple ratio	Soil water balance equation	SSM	RF
Cheng et al. [91]	China, Field	PLSR, KNN, RF, BPNN	UAV	CC, MSR, NDVI, OSAVI, RV1, RV12, SAVI, SIPI, TVI, EVI, MCARI, NPCI, GI, GNDVI, SRPI, NPCI, NDVIgb, PSRI, VARI, CIVE, VA, NRCT, TVDI, VA, CO, EN, COR, ME, HO, DS, SM.	TDR	10, 20	RF
Ge et al. [115]	China, Field	RF, ELM	UAV	NDVI *, NDVI705 *, RVI, NDCI, GNDVI, OSAVI, NDRE, mNDVI705, VOG1, VOG3, VOG2, CARI, MTVI1, TVI, DVI, RDVI, SPVI, WI/NDVI, EVI, NVI, MSAVI *, WI, REP, PRI, MTVI2, TCARI2, TCARI/OSAVI, MCARI/OSAVI, TCARI1, MCARI	Oven drying	10	RF
Ge et al. [116]	China, Field	XGBoost	UAV	Difference index (DI), ratio index (RI), normalized difference index (NDI)	Gravimetric method	10
Guo et al. [94]	China, Field	SVM, GRNN, CNNR	UAV	NDVI, DVI, MSAVI, rectified average value (Av), kurtosis (Ku), root mean square (Rm), peak factor (C), pulse (I)	TR-6D Soil Thermometer	6	CNNR
Guo et al. [117]	China, Field	MLR, BPNN, SVM	Sentinel-1, Sentinel-2	VV, VH, salinity index, NDVI, EVI, MSAVI, NDVIre, BI, intensity index, PDI, SMMI	Oven drying	20	SVM
Han et al. [118]	China, Regional	CART	MODIS	LST, ET, NDVI, precipitation, soil texture, elevation, Soil AWC, FC	-	10
Hou et al. [119]	China, Field	RF, SVR, ANN	Sentinel-1	VV, VH, incident angle, polarimetric parameters	TDR	7	RF
Hu et al. [120]	China, Field	LR, BPNN, GA-BPNN	GNSS-R	Signal to noise ratio	Hygrometer	7.5	GA-BPNN
Li and Yan [121]	China, Regional	RF, XGBoost, LightGBM, CatBoost, DNN, CNN, GRU, Stacking	Sentine-1, Sentinel-2	VV, VH, NDVI, GNDVI, MNDWI, NBRI, RVI, SATVI, SAVI, B2, B3, B4, B8, B11, B12, elevation *, latitude *, longitude *, month_cos, month_sin	Decagon 5TM	5	Stacking
Li et al. [76]	China, Regional	RF, GBR, RF-GBR	Sentinel-2	Band 1–8	Oven drying	20–30	RF-GBR
Li et al. [122]	China, Regional	Deep Forest, RF, GRNN, GBRT, SVM, KNN	Sentinel-1, Sentinel-2	VV *, VH, NDVI *, NDWI, clay *, sand, BD, land cover, fractional cover, longitude *, elevation *	Decagon 5TM ECH2O	5	Deep Forest
Zhang et al. [123]	China, Regional	BPNN, XGBoost, LSR, RF	MODIS, SMAP, ERA-Interim, ERA5-Land	NDVI, LST *, Land cover, reflectance, albedo, latitude, longitude, SMAP_TBH *, SMAP_TBV *, ERA_SR *, ERA_Eva, ERA_Runoff, ERA_TP, elevation, T_Clay, T_Sand, T_Silt, SOC, DOY	-	SSM	BPNN
Liu et al. [124]	China, Regional	RF, Extra Tree, LR	GF-1, Landsat 8, GF-4	COSRI *, DVI, EVI, GDVI, GLI, FNDVI, GOSAVI, GRVI, GSAVI, IPVI, MSAVI2, NDVI, NNIR, NR, OSAVI, RI, RVI, TVI, VARI, WDRI	Decagon 5TM SM sensor	3, 10, 20	Extra Tree
Zhang et al. [125]	China, Field	Copula Quantile Regression	RADARSAT-2	HH, VV, HV, VH, entropy, alpha, anisotropy, adaptive non-negative eigenvalue decomposition (odd, Dbl, Vol)	SPADE and Hydra Probe	4	-
Lv et al. [69]	China, Regional	RF, SVM, BPNN, Stacking	Landsat-9	NDVI, RVI, DVI, SAVI, EVI, GVI, SI-T, S1, S2, S3, NDSI, bright, green, wet, albedo, NDWI, NDBI, elevation	Oven drying	10	Stacking
Wang et al. [126]	China, Regional	PR, Ridge regression, LASSO, Elastic net, RF	Fengyun-3C, MODIS	FY-3C SM *, NDVI *, months, latitude *, longitude, elevation, clay, sand, silt	-	10	RF
Mu et al. [127]	China, Regional	LR, BPNN	GF-2, GF-3, GF-5	Bands normalization of GF-2, GF-3, and GF-5	Oven drying	10	BPNN
Wang et al. [71]	China, Regional	CART, RF, GBDT, ERT, Stacking	Landsat, Sentinel -1	VV *, VH, NDVI, EVI, SAVI, RVI, IPVI, LST, albedo, elevation, slope, SLIA, TVDI *, VSDI, VSWI, NIR, SWIR1, SWIR2, surface temperature *	Oven drying	10	Stacking
Qiao et al. [128]	China, Field	RBF, Multivariate RBF	Ground penetration radar	GPR signals	-	SSM
Wang et al. [64]	China, Field	RF, GRNN, CNN, SSA-CNN	Sentinel-1, Sentinel-2	Incident angle, VH, VV, scattering entropy, inverse entropy, α, eigenvalues (λ1 *, λ2), Zs), NDVI *, NDWI, MSI *, FVI *	TDR	3.8	SSA-CNN
Taghavi-Bayat et al. [129]	China, Regional	DRF	Sentinel-1	S-1:TFGVV *, TFGVH, LIA, SAR-VI: PRVI, DpRVI *, SAR-VI: DPSVI, mDPSVI *, RVI *, elevation, slope, TRI, TI, flow direction, curvature, SM, soil temperature, diurnal temperature range, minimum temperature, maximum temperature, daily rainfall events, days of year *	-	5	-
Tao et al. [72]	China, Regional	CatBoost, RF, GBDT, Stacking	MODIS	VCI, NDWI, NDVI, EVI, TCI *, CWSI *, VSWI, elevation *	-	SSM	Stacking
Jiang et al. [130]	China, Regional	RBFNN, LSTM, RNN, PCA-LSTM	-	RH, Ta, soil temperature, U, wind direction, rainfall, light intensity	-	SSM	PCA-LSTM
Xiaoxia et al. [131]	China, Regional	BPNN	Fengyun-3B, MODIS	LST, NDVI, albedo, NDVI, elevation	-	SSM	-
Xiang-yu et al. [132]	China, Field	GBR, RF, XGBoost	UAV	DI, NDI, RI, PI	-	10	XGBoost
Xu et al. [133]	China, Regional	GABP, SVR, RF	Sentinel-1, Sentinel-2, TerraSAR	Incident angle, VV, VH, HH, NDVI, NDWI	TDR	SSM	RF
Jiang et al. [134]	China, Regional	PCR, PLSR, BPNN	ASD Field Spec FR spectrometer	EVI, TVI, DSI, NDMI, SARVI	-	20	BPNN
Yang et al. [135]	China, Regional	DBN	Sentinel-1, Landsat 8	VV, VH, incident angle, EVI, OSAVI, elevation, slope, aspect	-	10	-
Yang et al. [136]	China, Regional	RF, ERT, XGBoost, LightGBM	MODIS, SMAP, TRIMS	NDVI, NDWI, DDI, SM, LST	-	10	LightGBM
Zhang et al. [75]	China, Regional	DI-Conv GRU, LSTM, Interp-ConvGRU, DI-LSTM	SMAP	SM from SMAP, precipitation, Ta, Rs, RH, U, sand, silt, clay content, BD, land cover, elevation	-	5	DI_Conv GRU
Zhang et al. [137]	China, Regional	RF, SVM, GA-BP	Sentinel-1, Sentinel-2	NDVI *, NDWI, RVI, MSI, WBI, FVI, Zs, average scattering angle *, θ *, cos(θ) *, sin(θ) *, VV, VH, inverse entropy, scattering entropy	TDR	SSM	RF
Zhang et al. [138]	China, Field	PLSR, KNN, RF	UAV	SAVI, NDVI, TCARI, VARI, OSAVI, SIPI, MCSARI, NDVIgb, PSRI, CIVE, MCARI/OSAVI, TCARI/OSAVI, NRCT, TVDI, GLCM features	Oven drying	10, 20	RF
Zhang et al. [139]	China, Regional	DFNN	VIIRS-RDR	Sixteen moderate resolution bands (M1–M16)	Gravimetric method	10
Zhan et al. [140]	China, Regional	MLR, AdaBoost, RF, GBM	Sentinel-2	MTCI, MNDVI, PSRI, RENDVI, S2REP, IRECI, OSAVI, SAVI red, MSR, soil textures	Oven drying	10	GBM
Chaudhary et al. [141]	India, Field	SVM, RF, MLP, RBF, WM, SBC, ANFIS, HyFIS, DENFIS	Sentinel-1	VH, VV	Steven’s HydraGo	SSM	SBC
Das et al. [142]	India, Field	Cubist, GBM, RF, Stacking	Landsat, Sentinel -1	shadowMin, shadowMax, NDVI, TVI, TTVI, SR, EVI2, CTVI, GEMI, RVI, MSAVI2, SAVI, NRVI, SLAVI, MSAVI, WDVI, DVI, B5_sre, Sigma0_VH, NDWI2, B6_sre, B4_sre, B2_sre, B3_sre, GNDVI, NDWI, SATVI, B7_sre *, NBRI *, LST *, MNDWI *, Sigma0_VV *	ML3 ThetaProbe	5	Stacking
Datta et al. [143]	India, Regional	LR, MLR, RF, KNN, SVM	Sentinel-1, Sentinel-2	VV *, VH	Oven drying	5	RF
Khose and Mailapalli [144]	India, Field	LR, KNN, RF, DT, SVR	UAV	NDVI, NDWI, TNDVI, Simple Ratio or Ratio Vegetation Index (RVI), SAVI, PDI.	Gravimetric method	1, 5	LR, SVR
Nijaguna et al. [73]	India, Regional	DMN-Bi-GRU	Sentinel-1, Sentinel-2	NDVI, GLAI, GNDVI, WDRVI	-	SSM	DMN-Bi-GRU
Pal and Maity [145]	India, Regional	RF, SVN, GP	Radar Imaging Satellite 1 (RISAT1)	HH *, HV, VH, VV, sand, silt, clay	Gravimetric method	5	GP
Singh and Gaurav [61]	India, Regional	ANN, GRNN, RBN, ERBN, GPR, SVR, RF, Boosting, RNN, RNNBDT, AutoML	Sentinel-, Sentinel-2	VV, VH, VH/VV, VH-VV, angle, NDVI, elevation, longitude, latitude	ML3 theta probe	5	ANN
Singh et al. [66]	India, Regional	LR, SVR, DT, RF, LSTM	SMAP	SSM, LST, soil texture, land cover, wind direction, U, surface pressure, dew point, Ta, precipitation	-	5	LSTM
Lee et al. [96]	South Korea, Regional	DNN	MODIS	OLR, Isolation accumulated precipitation (12 h, 24 h), NDVI, Ta, RH, Land cover, elevation, slope	SM probes	10	-
Lee et al. [146]	South Korea, Regional	RF, AutoML	IMERG, GK2A, VIIRS, LDAPS, SRTM	10- and 20-day cumulative standardized precipitation indexes (SPI10 * and SPI20 *), NDVI, DSR, Ta, LST, soil temperature, RH, LE, slope *, elevation *, TRI, aspect.	TDR	10	AutoML
Adab et al. [92]	Iran, Regional	RF, SVM, ANN, EN	Landsat 7, 8	LST *, blue *, green, red, NIR *, SWIR1, SWIR2	Capacitance Probe	10	RF
Asadollah et al. [74]	Iran, Regional	GBM, SVR, GBM-SVR	GLDAS, AMSR2, SMAP	AMSR2-C1, AMSR2-C2, ASMR2-X, SMAPL3, SMAPL4, GLDAS *	TDR	6	GBM-SVR
Bandak et al. [147]	Iran, Regional	RF, SVR, XGBoost, Extra tree	Sentinel-2, Landsat 8	Salinity index -S1, Salinity index -S2, Salinity index -S3, Salinity index -S4, Salinity index -S5, NDSI *, NDVI, SAVI, VSSI, NDWI *, Extended EVI, MNDWI *	Oven drying	10	Extra tree
Fathololoumi et al. [148]	Iran, Regional	RF, Cubist	Landsat 8	Elevation, aspect-sin, aspect-cos, catchment area, convergence index, maximum difference of mean elevation, general curvature, hill shade, MSP, MRRTF, LS, total curvature, albedo *, emissivity *, LST *, Incidence	Portable SM meter	SSM	RF, Cubist
Hemmati and Sahebi [63]	Iran, Regional	RF, SVR, MLP, CNN	Sentinel-1, Sentinel-2	VV, VH, incident angle, NDVI	Thetaprobe sensor	5	CNN
Karamvand et al. [149]	Iran, Regional	ELM	SMAP, MODIS	Precipitation, NDVI, LST	-	5	-
Moosavi et al. [150]	Iran, Regional	PSO-ANFIS, PSO-SVR	MODIS, Landsat 8	Evaporative fraction, TVDI, VTCI, STVDI, TVX	-	SSM	PSO-SVR
Moosavi et al. [151]	Iran, Regional	PSO-CMAC, PSO-GMDH	Sentinel-2	NDVI, NDWI, all bands from Sentinel-2, slope, aspect, elevation	TDR	5	PSO-CMAC
Moosavi et al. [70]	Iran, Regional	LSTM, CNN, CNN-LSTM	Sentinel-2	NDVI, NDWI, NDMI, BSI, elevation, slope, aspect, TWI, land use, soil properties	TDR	8	CNN-LSTM
Nouraki et al. [152]	Iran, Regional	MLR, CART, M5P, GBM, RF	Landsat	Blue, green, red, NIR, SWIR1, SWIR2, LST, NDVI, NSMI, NTR, SWI, sand, silt, clay, BD	-	10	RF, GBM
Tahmouresi et al. [153]	Iran, Regional	RF, GBM, XGBoost	SMAP, AMSR2, MODIS	Elevation, slope, sand, silt, LST, SSM, NDVI, precipitation	TDR	5	XGBoost, GBM
Shokati et al. [154]	Iran, Field	M5 tree, RF, SMOreg, MLP	UAV	DI, RI, NDI, PI	Gravimetric method	5	RF
Pasolli et al. [58]	Switzerland, Field	SVR, MLP	Radiometer–scatterometer	VV, VH, emissivity polarizations, incident angle	Oven drying	SSM	SVR
Kseneman et al. [155]	Slovenia, Regional	FFBP Neural Network	TerraSAR-X	Backscatter coefficients	TRIME-PICO64	5	-
Pongrac and Gleich [156]	Slovenia, Regional	CNN	ALOS-2	-	-	SSM	-
Jia et al. [157]	Italy, Regional	RF, SVM	GNSSR	Reflectivity, elevation angle	-	SSM	RF
Notarnicola et al. [158]	Italy, Regional	SVR	ASAR WS sensor, MODIS, GEOtop	VV, VH, slope, elevation, aspect, land cover, soil type	-	SSM	-
Pasolli et al. [159]	Italy, Regional	SVR	RADARSAT-2, Envisat ASAR, MODIS	HH, HV, HV/VV, altitude *, incident angle, NDVI *, land use	TDR	5	SVR
Santi et al. [160]	Italy, Regional	ANN	Sentinel-1, SMAP	Brightness temperature, acquisition geometry, land use, NDVI	-	SSM	-
Portal et al. [161]	Spain, Regional	ANN	Sentinel-2, MODIS, ERA5, ESA CCI	Precipitation, reflectance, LST, elevation, slope, SM, land cover, NDVI, NDRE, EVI, GNDVI, SAVI, NDMI, MSI, NBRI, BSI, NDWI	-	10	-
Chen et al. [162]	Spain, Field	1D-CNN	Spectrophotometer	Spectral measurement from different soil samples	Oven drying	SSM	-
Abebrese et al. [163]	Czech Republic, Field	XGBoost	Sentinel-2	NDVI, NDRI, NDWI, DVI, RVI, MSI, elevation	EC-5 Sensor	5–10	-
Lendzioch et al. [164]	Czech Republic, Field	RF	UAV	RGI, VVI, VDVI, VARI, TGI, SI, SHP, SCI, SAT, NGRDI, NDTI, NDI, HI, GRVI, GLI, GLAI, ExG, ERGBVE, CI, BI, RI, HUE, NDVI, Slope, ProfCurv, PlnCurv, WEI, TRI (1, 4), VRM (1, 4), TPI, TWI, Temperature, GWL, SM.	Oven drying	SSM	-
Hajdu et al. [80]	New Zealand, Regional	RF	Sentinel-1	Backscatter, incident angle, NDVI, slope, aspect, roughness index, wetness index	Capacitance-based, AquaCheck sub-surface probes	10	-
Holtgrave et al. [165]	Germany, Regional	SVR	Sentinel-1, Landsat	VV, VH *, incident angle, NDVI *, height *, slope, and aspect	FDR	SSM	-
Morellos et al. [166]	Germany, Field	LS-SVM, Cubist	Spectrophotometer	VIS-NIR reflection	Oven drying	20	LS-SVM
Schröter et al. [167]	Germany, Regional	Fuzzy c-means sampling and estimation approach (FCM SEA)	LIDAR	Elevation	-	10	-
Schröter et al. [168]	Germany, Field	Fuzzy c-means sampling and estimation approach (FCM SEA)	RapidEye sensor	reNDVI, elevation, LAI	TDR	10	-
Özerdem et al. [169]	Turkey, Regional	GRNN	RADARSAT-2	VV, VH, HV, HH, entropy, anisotropy, alpha angle, volume scattering, odd bounce, double bounce	Gravimetric method	3–5	-
Şekertekin, Marangoz [170]	Turkey, Regional	MLR	Sentinel-1	VV, VH, incident angle	Oven drying	SSM
Usta [98]	Turkey, Regional	ANN	Landsat 8	EVI, SAVI, NDMI, MNDVI, MSI, SMI, TIRS 1, TIRS2, slope *, TWI *, ALT, TPI *, TRASP, EAST *, NORTH, SITE, CTI	-	0–10, 10–20	-
Huang et al. [171]	Austria, Regional	Deep belief network	Landsat, MODIS, SMAP, ERA-5	LST, NDVI, albedo, precipitation, SMAP SM, ERA-5 SM	-	5	-
Ayehu et al. [172]	Ethiopia, Regional	SCA, SVM	Sentinel-1, MODSI, SRTM	VV, VH, NDVI, elevation	-	SSM	SCA
Ayehu et al. [173]	Ethiopia, Field	LR, ANN	Sentinel-1, Landsat 7 and 8	VV, surface heights (hrms), surface correlation length (leff)	FDR	5	ANN
Attarzadeh et al. [174]	Kenya, Regional	SVR	Sentinel-, Sentinel-2	VV, GLCM Contrast, GLCM Correlation, GLCM Dissimilarity, GLCM Entropy, GLCM Homogeneity, GLCM Mean, GLCM Standard Dev, GLCM Angular, SAVI, NDVI, TSAVI, MSAVI, MSAVI2, DVI *, RVI, PVI, IPVI, WDVI, TNDVI, GNDVI *, GEMI *, ARVI, NDI45, MTCI, MCARI, REIP, S2REP, IRECI, PSSRa, LAI, FCOVER *, NDWI, NDWI2, BI, BI2, RI, CI	TDR	5	-
Dutta and Terhorst [175]	Australia, Regional	S-ANFIS, MLP, PNN, RBFN	CosmOz Cosmic Ray Sensor	Neutron counts from cosmic ray probe, rainfall, Ta, RH	-	0–5, 10–15, 25–30	S-ANFIS
Celik et al. [67]	Global	LSTM	Sentinel-1, Sentinel-, SMAP	VV, VH, SSM, soil textures *, elevation, slope, aspect, hill shade, Ta, ET, precipitation *	-	10	-
Nativel et al. [82]	Global	ANN	Sentinel-1, Sentinel-2	VV *, VH *, the classical change detection SSM index ISSM, incidence angle *, NDVI *, VH/VV ratio *	-	10	-
Jia et al. [176]	Global	XGBoost, ANN, SVM, RF	CYGNSS, SMAP	Reflectivity from CYGNSS, vegetation opacity, and roughness coefficients from SMAP	-	SSM	XGBoost
Rodríguez-Fernández et al. [177]	Global	Neural Network	SMOS	RH, Ta, ASCAT SM, LST, snow cover, snow temperature, brightness temperature	-	5
Senyurek et al. [178]	Global	RF	CYGNSS, MODIS	Reflective power, incident angle, trailing edge slope, NDVI, vegetation water content, elevation, slope, soil properties	-	SSM	-
Zhang et al. [59]	Global	RF, XGBoost	SMAP, ERA5, Landsat	Landsat 8 (5 VNIR bands, 2 SWIR bands, 2 TIR bands, solar elevation, angle), elevation *, slope, aspect *, sand *, silt, clay *, precipitation, surface skin temperature, ERA-5 SM *	-	SSM	XGBoost
Hamidisepehr et al. [179]	Global	20 different ML models available in MATLAB	Spectrometers	1024 wavelengths into 20 components using Partial Least Square regression.	-	0.64, 1.92, 3.20, 4.48, 5.76	Cubic SVM
Jia et al. [180]	Global	RF	GNSS-R	Signal-to-noise ratio, elevation angle	-	SSM	RF
Lei et al. [60]	Global	RF	CYGNSS, MODIS	CYGNSS reflectivity, trailing edge slope, specular point incidence angle, NDVI, vegetation water content, elevation, clay, silt	-	SSM	RF
Lei et al. [181]	Global	RF	CYGNSS, SMAP	Reflectivity, TES, incidence angle, elevation, NDVI, VWC, silt, clay	-	10
Nabi et al. [182]	Global	CNN	CYGNSS, SMAP, MODIS	Analog power, effective scattering area, BRCS VWC, NDVI, elevation, slope, water percentage, clay, silt	-	10
Roberts et al. [183]	Global	CNN	CYGNSS, SMAP	DDM, VWC, elevation, surface water occurrence	-	SSM	-
Wang et al. [184]	Global	ANN	Fengyun, SMAP	Brightness temperature, microwave vegetation index	-	5	-
Zhang et al. [185]	Global	RF	MODIS	LST *, LST difference, NDVI, EVI, API *, clay, sand, silt, latitude *, longitude *	-	SSM
Zhu et al. [186]	Global	Transfer Learning	SMAP, Sentinel-1, MODIS	VV, VH, incident angles, sand, clay, BD, elevation, slope, aspect, NDVI, Ta, precipitation, day of year	-	5	-
Li et al. [187]	China and USA	Deep and parallel model	Sentine-1, Sentinel-2	VV, VV+VH, incident angle, NDVI, elevation, land covers, soil properties	-	SSM	-
Fang et al. [188]	United States and Mexico	Fully CNN (FCNN)	MODIS, SMAP	NDVI, LST, SSM	-	10
Hegazi et al. [81]	Australia and Austria	CNN	Sentinel-1	Sigma0_VH and Sigma0_VV or Gamma0_VH and Gamma0_VV or Beta0_VH and Beta0_VV	Campbell Scientific water content reflectometers	20	-
Hegazi et al. [65]	Australia and Austria	CNN	Sentinel-2	B1-B12 bands (B1—Coastal aerosol, B2—Blue, B3—Green, B4—Red, B5—Vegetation red edge *, B6—Vegetation red edge, B7—Vegetation red edge, B8—NIR, B8A—Narrow NIR, B9—Water vapor, B10—SWIR—Cirrus, B11—SWIR *, B12—SWIR)	Campbell Scientific water content reflectometers and TDR	30	-
Eroglu et al. [93]	USA and Australia	ANN	CYGNSS	Surface reflectivity, TES, and SP incidence angle, NDVI, VWC, elevation, slope, h-parameters.	-	SSM	-
Tsagkatakis et al. [189]	USA and Switzerland	CNN	SMAP	SSM	-	SSM	-

In the input feature column, asterisk (*) indicates the most important features for surface soil moisture estimation. All the abbreviations can be found in the Supplementary Table S3.

Figure 7. Top 10 input features used in machine learning-based soil moisture estimation. This is derived from Table 2 and Table 3. (a) Surface soil moisture, (b) root zone soil moisture, (c) field-scale studies, and (d) global/regional-scale studies. We defined the field scale as studies focused on specific locations or individual fields, the regional scale as studies covering particular countries, states, or regions, and the global scale as studies examining multiple countries or worldwide.

Figure 7. Top 10 input features used in machine learning-based soil moisture estimation. This is derived from Table 2 and Table 3. (a) Surface soil moisture, (b) root zone soil moisture, (c) field-scale studies, and (d) global/regional-scale studies. We defined the field scale as studies focused on specific locations or individual fields, the regional scale as studies covering particular countries, states, or regions, and the global scale as studies examining multiple countries or worldwide.

3.4.1. SSM and RZSM Prediction

The 144 papers reviewed focused on either SSM estimation (119 studies, Table 2) or RZSM estimation (25 studies, Table 3). SSM studies most frequently used the normalized difference vegetation index (NDVI) feature (60 studies), followed by elevation (40 studies) and backscattering coefficients (VV in 32 studies, VH in 27 studies) (Figure 7a). Notably, soil properties did not rank among the top 10 features. RZSM studies most frequently used SSM and NDVI (13 studies each), followed by precipitation, land surface temperature (LST), enhanced vegetation index (EVI), and soil properties (sand, silt, and bulk density) (Figure 7b).

Few studies have systematically analyzed feature importance to identify the most sensitive variables within a given set of input features. We report the most influential features by marking them with an asterisk (*) at the end of features in the input features column of Table 2 and Table 3. For SSM, vegetation indices and backscattering coefficients were found to be the most significant predictors. Chen et al. [110] estimated SSM using RADARSAT-2 in agricultural fields (corn, soybean, and winter wheat) in Canada and observed that backscattering coefficients (VV and VH) played a critical role. Backscattering coefficients establish a direct relationship between radar signals and SM content. As SM increases, the dielectric constant rises, enhancing signal reflection and backscatter intensity. Conversely, dry soil lowers the dielectric constant, reducing backscatter. This sensitivity makes backscattering coefficients reliable indicators of SM variability [190]. Similarly, Holtgrave et al. [165] assessed SSM in vegetated floodplains in Germany using Sentinel-1 and Landsat data, identifying backscattering coefficients and NDVI as the most important features in ML models. Other studies conducted across various regions have also reported that vegetation indices, backscattering coefficients, and topographical features are key predictors for SSM estimation [112,145,159].

In contrast, for RZSM, soil properties and SSM play a more significant role than backscattering coefficients and vegetation indices. Karthikeyan and Mishra [95] estimated multi-layer SM (up to 100 cm) across CONUS and found that soil properties were crucial for predicting deeper-layer SM. Ma et al. [97] estimated RZSM (up to 80 cm) and identified soil properties (sand and clay content), topographical variables (slope and aspect), and LST as key influencing factors. Similarly, Babaeian et al. [86] estimated RZSM in wheat fields in Arizona, USA, and reported that SSM and soil properties were the primary determinants. Studies conducted in various regions have consistently indicated that SSM, soil properties, and topographical features are more influential than vegetation indices and backscattering coefficients in RZSM estimation [83,191,192]. Satellite-based features, such as vegetation indices and backscattering coefficients especially from X and C-bands SAR, could not provide information on deeper SM due to their inability to penetrate to a deeper profile [34,40,193]. SSM serves as an essential proxy for deeper SM content due to its strong hydrological connectivity with subsurface layers [194,195], while soil physical and hydraulic properties control water retention, infiltration, and redistribution processes, significantly influencing SM availability in the root zone [86,97,196].

Table 3. Root Zone Surface Soil Moisture (RZSM) estimation studies: detailed overview of study areas, machine learning techniques, data sources, and input features.

Author	Study Area/Scale	ML Technique	Satellites/UAV	Input Features	In Situ SM Measured Method	In Situ SM Depth (cm)	Best ML Model
Babaeian et al. [86]	USA, Field	AutoML (NN, GBM, GLM, DRF)	UAV	NTR, NDVI, texture, BD, OC, θfc, θpwp, Ksat, ϕ, SSM	TDR	2, 10, 50	NN
Karthikeyan and Mishra [95]	USA, Regional	XGBoost	MODIS, SMAP	SM and SSM from SMAP, sand, silt, clay, BD, NDVI, EVI, LST, precipitation, elevation	-	5, 10, 20, 50, 100	-
Kisekka et al. [196]	USA, Field	RF	Landsat 7 and 8	NDVI *, elevation *, BD *, soil temperature *, Rs, RH, ET, U, precipitation, evaporation fraction *	TDR	30, 60, 90	-
Peng et al. [86]	USA, Regional	Quantile RF	Sentinel-1, SMAP, Landsat-8, MODIS	VV, VH, near infrared, thermal bands, LST, SSM, elevation, land cover, terrain attributes	TEROS 12, Meter Group	50, 100	-
Sahaar and Niemann [83]	USA, Regional	ANN, RF, XGBoost, CatBoost, LightGBM	SMAP, GPM, ECOSTRESS, MODIS	SSM, RZSM, PSM, NCLD, NDVI, EVI, LAI, fPAR, sand, silt, clay *, OM, BD, EC, pH, precipitation, LST, ET, AI *, elevation *, slope, aspect, hill shade, TPI	-	5, 10, 20, 50, 100	XGBoost
Xia et al. [197]	USA, Regional	Quantile RF	NLDAS, MODIS, PRISM, SRTM	NLDAS based SM *, LST, NDWI, EVI, GPP, BD *, SOC *, precipitation, Ta, VPD, elevation, slope *, aspect *, vertical, horizontal, and mean curvatures, TWI, surface roughness, land cover, tree cover percentage	-	5, 10, 20, 50, 100	-
Zeng et al. [191]	USA, Regional	RF	SMOS, MODIS, TRMM	SMOS SM *, vertically polarized brightness, horizontally polarized brightness *, ET, PET, NR, NDVI, LST, LST gap, TVDI, precipitation, soil properties, elevation, slope	-	5, 25, 60	-
A et al. [85]	China, Regional	ConvLSTM	MODIS, GLDAS	Precipitation, Ta, BD, SSM, ETa, NDVI, LAI, SSM *	Gravimetric Method	5, 10, 20, 40, 60, 80	-
A et al. [198]	China, Regional	RF	MODIS, ASTER	LAI, PET, LST	-	3, 5, 10, 20, 50
Luo et al. [199]	China, Field	RLR, RLSSVM, RELM	GF-1	soil brightness index *, four salinity indices (SI, SI1, SI2, SI3) *, EVI, ARVI, SAVI2 *, TSAVI, PDI	Oven drying	20, 40, 60	RELM
Ma et al. [97]	China, Regional	RF	MODIS	NDVI, ET, albedo, LST *, sand *, silt *, clay, elevation, slope *, aspect *, precipitation, SMC	-	5–80	-
Zeng et al. [192]	China, regional	RF, LightGBM, XGBoost	ERA5, ESA-CCI, CSSPv2, GLDASv2.1	Elevation *, latitude *, longitude *, clay, sand, OM, Ta, specific humidity, U, surface pressure, precipitation, solar radiation, SM of ERA5, GLDASv2.1, ESA-CCI	FDR	0–100	-
Zhu et al. [87]	China, Field	RF, SVM, ELM	UAV	NDVI, NDRE, SCCI, RVI, NDCI, GNDVI, OSAVI1, OSAVI2, VOG1, CARI, CARI717, CARI740, MTVI1, TVI, DVI, RDVI, SPVI, EVI1, EVI2, EVI3, MSAVI1, MSAVI2, REP1, PRI, MTVI2, TCARI1, TCARI2, TCARI1, OSAVI1, OSAVI2, MCARI, TCARI, MCARI, MSAVI1, OSAVI1, OSAVI2, MCARI, TCARI1, SAVI, VARI, GLI, IPVI, NNIR, GCI, RECI, WDRVI.	FDR-TRIME	10, 20, 30, 40, 60	RF
Li et al. [200]	China, Regional	RF	ERA5-Land, MODIS	Precipitation, Ta, evaporation *, LAI, ERA5-Land SM *, rock fragment, porosity, sand, silt, clay *, land cover type, elevation *	-	0–100
Tang et al. [201]	China, Regional	RF, SVM, BPNN, ELM	ASD Field-Spec 3 portable field hyper spectral spectrometer	Spectral reflectance	Oven drying	0–20, 20–40, 40–60	RF
Yin et al. [88]	China, Field	PLSR, SVR, RF, DNN	UAV	Raw bands, GRVI, NDVI, GNDVI, NDRE, SIPI, SAVI, OSAVI, MSR, MCARI, TCARI, EVI, NDREI, CH, VF, CWSI, TVDI, NCRT, GLCM	Oven drying	30	DNN
Zhu et al. [202]	China, Field	RF, GBR, XGBoost, LightGBM, CataBoost	UAV	SAVI, SMMI, MSAVI, MSMMI, EVI, MTVI	FDR-TRIME	0–10, 10–20, 20–30, 30–40, 40–50, 50–60	XGBoost
Zhu et al. [89]	China, Field	RF, GBR, XGBoost, LightGBM, CatBoost	UAV	NDVI, EVI, CWSI, TDVI, TDDI	FDR-TRIME	0–10, 10–20, 20–30, 30–40, 40–50, 50–60	CatBoost
Zhu et al. [62]	China, Regional	MLR, RF, ANN	MODIS	SWC, Ts *, albedo *, ET *, NDVI *, clay, sand, SOC, BD, hydraulic conductivity, slope	neutron-probe	0–5 m	ANN
Manninen et al. [203]	Finland, Regional	GBM	Sentinel-1	VV, VH, altitude, aspect, local slope angle, incident angle, cos(θ), LAI, time difference between the SM measurement and SAR image acquisition, azimuth difference of radar looking direction and terrain slope	WET-2 and PR2 Profile Probe	SSM, 10, 20, 30, 40 cm	-
Joshi et al. [204]	Australia, Regional	LightGBM	MODIS	Ta, Rs, RH, rainfall, latitude, longitude, MODIS bands, LST, soil properties	-	1–6 m	-
Efremova et al. [205]	Australia, Regional	RF, LR, MLP, SVM, Ridge Regression, Kernel Ridge Regression	Sentinel-1, Sentinel-2	Sentinel-1 and Sentinel-2 images	Embedded SM sensors	10–120	RF
Fuentes et al. [206]	Australia, Regional	MLP	Sentinel-1, MODIS, SMAP	VV, relative SSM, LST, NDVI, land cover, clay, SOC, AWC	TDR	0–30, 30 to 60
Madhukumar et al. [68]	Australia, Regional	LSTM, Hybrid-GRU, Multi-head TNN, Hybrid TNN	SMAP, MODIS	SSM, soil moisture indices, weather variables	-	80	-
Souissi et al. [207]	Global	ANN	MODIS	NDVI *, SST, SWI *, SSM	-	30, 55	-

In the input feature column, asterisk (*) indicates the most important features for root zone soil moisture estimation. OM: Organic Matter, BD: Bulk Density, AWC: Available Water Capacity, K_sat: Saturated Hydraulic Conductivity, SOC: Soil Organic Carbon, SMC: Soil Moisture Content, EC: Electrical Conductivity, NDVI: Normalized Difference Vegetation Index, EVI: Enhanced Vegetation Index, LAI: Leaf Area Index, ET_a: Actual Evapotranspiration, PET: Potential Evapotranspiration, R_s: Solar Radiation, RH: Relative Humidity, T: Temperature, U: Wind Speed, VPD: Vapor Pressure Deficit, LST: Land Surface Temperature, TWI: Topographic Wetness Index, TPI: Topographic Position Index, SSM: Surface Soil Moisture. All the abbreviations can be found in the Supplementary Table S3.

Table 3. Root Zone Surface Soil Moisture (RZSM) estimation studies: detailed overview of study areas, machine learning techniques, data sources, and input features.

Author	Study Area/Scale	ML Technique	Satellites/UAV	Input Features	In Situ SM Measured Method	In Situ SM Depth (cm)	Best ML Model
Babaeian et al. [86]	USA, Field	AutoML (NN, GBM, GLM, DRF)	UAV	NTR, NDVI, texture, BD, OC, θfc, θpwp, Ksat, ϕ, SSM	TDR	2, 10, 50	NN
Karthikeyan and Mishra [95]	USA, Regional	XGBoost	MODIS, SMAP	SM and SSM from SMAP, sand, silt, clay, BD, NDVI, EVI, LST, precipitation, elevation	-	5, 10, 20, 50, 100	-
Kisekka et al. [196]	USA, Field	RF	Landsat 7 and 8	NDVI *, elevation *, BD *, soil temperature *, Rs, RH, ET, U, precipitation, evaporation fraction *	TDR	30, 60, 90	-
Peng et al. [86]	USA, Regional	Quantile RF	Sentinel-1, SMAP, Landsat-8, MODIS	VV, VH, near infrared, thermal bands, LST, SSM, elevation, land cover, terrain attributes	TEROS 12, Meter Group	50, 100	-
Sahaar and Niemann [83]	USA, Regional	ANN, RF, XGBoost, CatBoost, LightGBM	SMAP, GPM, ECOSTRESS, MODIS	SSM, RZSM, PSM, NCLD, NDVI, EVI, LAI, fPAR, sand, silt, clay *, OM, BD, EC, pH, precipitation, LST, ET, AI *, elevation *, slope, aspect, hill shade, TPI	-	5, 10, 20, 50, 100	XGBoost
Xia et al. [197]	USA, Regional	Quantile RF	NLDAS, MODIS, PRISM, SRTM	NLDAS based SM *, LST, NDWI, EVI, GPP, BD *, SOC *, precipitation, Ta, VPD, elevation, slope *, aspect *, vertical, horizontal, and mean curvatures, TWI, surface roughness, land cover, tree cover percentage	-	5, 10, 20, 50, 100	-
Zeng et al. [191]	USA, Regional	RF	SMOS, MODIS, TRMM	SMOS SM *, vertically polarized brightness, horizontally polarized brightness *, ET, PET, NR, NDVI, LST, LST gap, TVDI, precipitation, soil properties, elevation, slope	-	5, 25, 60	-
A et al. [85]	China, Regional	ConvLSTM	MODIS, GLDAS	Precipitation, Ta, BD, SSM, ETa, NDVI, LAI, SSM *	Gravimetric Method	5, 10, 20, 40, 60, 80	-
A et al. [198]	China, Regional	RF	MODIS, ASTER	LAI, PET, LST	-	3, 5, 10, 20, 50
Luo et al. [199]	China, Field	RLR, RLSSVM, RELM	GF-1	soil brightness index *, four salinity indices (SI, SI1, SI2, SI3) *, EVI, ARVI, SAVI2 *, TSAVI, PDI	Oven drying	20, 40, 60	RELM
Ma et al. [97]	China, Regional	RF	MODIS	NDVI, ET, albedo, LST *, sand *, silt *, clay, elevation, slope *, aspect *, precipitation, SMC	-	5–80	-
Zeng et al. [192]	China, regional	RF, LightGBM, XGBoost	ERA5, ESA-CCI, CSSPv2, GLDASv2.1	Elevation *, latitude *, longitude *, clay, sand, OM, Ta, specific humidity, U, surface pressure, precipitation, solar radiation, SM of ERA5, GLDASv2.1, ESA-CCI	FDR	0–100	-
Zhu et al. [87]	China, Field	RF, SVM, ELM	UAV	NDVI, NDRE, SCCI, RVI, NDCI, GNDVI, OSAVI1, OSAVI2, VOG1, CARI, CARI717, CARI740, MTVI1, TVI, DVI, RDVI, SPVI, EVI1, EVI2, EVI3, MSAVI1, MSAVI2, REP1, PRI, MTVI2, TCARI1, TCARI2, TCARI1, OSAVI1, OSAVI2, MCARI, TCARI, MCARI, MSAVI1, OSAVI1, OSAVI2, MCARI, TCARI1, SAVI, VARI, GLI, IPVI, NNIR, GCI, RECI, WDRVI.	FDR-TRIME	10, 20, 30, 40, 60	RF
Li et al. [200]	China, Regional	RF	ERA5-Land, MODIS	Precipitation, Ta, evaporation *, LAI, ERA5-Land SM *, rock fragment, porosity, sand, silt, clay *, land cover type, elevation *	-	0–100
Tang et al. [201]	China, Regional	RF, SVM, BPNN, ELM	ASD Field-Spec 3 portable field hyper spectral spectrometer	Spectral reflectance	Oven drying	0–20, 20–40, 40–60	RF
Yin et al. [88]	China, Field	PLSR, SVR, RF, DNN	UAV	Raw bands, GRVI, NDVI, GNDVI, NDRE, SIPI, SAVI, OSAVI, MSR, MCARI, TCARI, EVI, NDREI, CH, VF, CWSI, TVDI, NCRT, GLCM	Oven drying	30	DNN
Zhu et al. [202]	China, Field	RF, GBR, XGBoost, LightGBM, CataBoost	UAV	SAVI, SMMI, MSAVI, MSMMI, EVI, MTVI	FDR-TRIME	0–10, 10–20, 20–30, 30–40, 40–50, 50–60	XGBoost
Zhu et al. [89]	China, Field	RF, GBR, XGBoost, LightGBM, CatBoost	UAV	NDVI, EVI, CWSI, TDVI, TDDI	FDR-TRIME	0–10, 10–20, 20–30, 30–40, 40–50, 50–60	CatBoost
Zhu et al. [62]	China, Regional	MLR, RF, ANN	MODIS	SWC, Ts *, albedo *, ET *, NDVI *, clay, sand, SOC, BD, hydraulic conductivity, slope	neutron-probe	0–5 m	ANN
Manninen et al. [203]	Finland, Regional	GBM	Sentinel-1	VV, VH, altitude, aspect, local slope angle, incident angle, cos(θ), LAI, time difference between the SM measurement and SAR image acquisition, azimuth difference of radar looking direction and terrain slope	WET-2 and PR2 Profile Probe	SSM, 10, 20, 30, 40 cm	-
Joshi et al. [204]	Australia, Regional	LightGBM	MODIS	Ta, Rs, RH, rainfall, latitude, longitude, MODIS bands, LST, soil properties	-	1–6 m	-
Efremova et al. [205]	Australia, Regional	RF, LR, MLP, SVM, Ridge Regression, Kernel Ridge Regression	Sentinel-1, Sentinel-2	Sentinel-1 and Sentinel-2 images	Embedded SM sensors	10–120	RF
Fuentes et al. [206]	Australia, Regional	MLP	Sentinel-1, MODIS, SMAP	VV, relative SSM, LST, NDVI, land cover, clay, SOC, AWC	TDR	0–30, 30 to 60
Madhukumar et al. [68]	Australia, Regional	LSTM, Hybrid-GRU, Multi-head TNN, Hybrid TNN	SMAP, MODIS	SSM, soil moisture indices, weather variables	-	80	-
Souissi et al. [207]	Global	ANN	MODIS	NDVI *, SST, SWI *, SSM	-	30, 55	-

In the input feature column, asterisk (*) indicates the most important features for root zone soil moisture estimation. OM: Organic Matter, BD: Bulk Density, AWC: Available Water Capacity, K_sat: Saturated Hydraulic Conductivity, SOC: Soil Organic Carbon, SMC: Soil Moisture Content, EC: Electrical Conductivity, NDVI: Normalized Difference Vegetation Index, EVI: Enhanced Vegetation Index, LAI: Leaf Area Index, ET_a: Actual Evapotranspiration, PET: Potential Evapotranspiration, R_s: Solar Radiation, RH: Relative Humidity, T: Temperature, U: Wind Speed, VPD: Vapor Pressure Deficit, LST: Land Surface Temperature, TWI: Topographic Wetness Index, TPI: Topographic Position Index, SSM: Surface Soil Moisture. All the abbreviations can be found in the Supplementary Table S3.

3.4.2. Input Features for Field Scale vs. Regional/Global Scale

Choice of input features varied depending on the scale of study, which we classified as either field scale (studies focused on specific locations or individual fields; 39 studies), or regional/global scale (studies covering particular countries, states, regions, or multiple countries; 105 studies) (Table 2 and Table 3). Figure 7c presents the top 10 features most frequently used in SM estimation at the field scale. NDVI was the most commonly used feature (21 studies), followed by VV (11 studies), VH (8 studies), and EVI (8 studies). At the field scale, vegetation indices and backscattering coefficients were widely used. In contrast, for global/regional studies, topographical and soil properties were more commonly used alongside vegetation indices and backscattering coefficients (Figure 7d). For global/regional studies, NDVI was the most common feature (55 studies), followed by elevation (46 studies), and slope (28 studies).

3.5. Impact of Vegetation Cover on Microwave-Based Soil Moisture Estimation

Microwave remote sensing, particularly in the active (e.g., SAR) and passive (e.g., radiometers) domains, has been widely used to estimate SM due to its direct sensitivity to the dielectric properties of soil, which are strongly influenced by moisture content [208]. Microwave sensors are highly sensitive to surface roughness, vegetation cover, and topographic variations, which affect backscatter and emissivity. In flat and sparsely vegetated areas, microwave signals can penetrate the soil surface and provide accurate estimates of SSM [209]. However, in hilly or mountainous terrain, topographic effects lead to distortions such as foreshortening, layover, and shadowing in SAR data, degrading the accuracy of SM retrievals [210]. Vegetation cover reduces the performance of ML models in SAR-based SM estimation by attenuating and scattering SAR signals, making it difficult to isolate the SM contribution [211]. Dense vegetation increases volume scattering, leading to higher backscatter values that do not directly correlate with SM. Several studies have shown the impact of vegetation cover on model performance.

Ahmad et al. [99] estimated SM in the lower Colorado River Basin, USA, where the site ranged from low to dense vegetation cover. Their study used backscatter and incidence angle, along with NDVI. The results showed that an increase in vegetation density adversely affects the performance of ML models in SM estimation. The RMSE increased by approximately 75% when moving from low vegetation to dense vegetation conditions. Celik et al. [67] estimated SM across 103 sites across the globe with diverse land covers, including croplands, grasslands, shrublands, forests, and mosaic landscapes (a mixture of trees, shrubs, herbaceous vegetation, and croplands). Their study utilized data from Sentinel-1, Sentinel-2, SMAP, as well as topographical and climatic variables. The results showed that ML models performed the worst under mosaic land cover (MAE of 0.05 cm³/cm³), while the best performance was observed in shrublands (0.02 cm³/cm³). The MAE in mosaic land is nearly 150% higher than MAE in shrublands. The shrubland class is sparsely vegetated, allowing SAR signals to effectively interact with the land surface. In contrast, forests and mosaic land covers posed challenges, as SAR signals struggle to penetrate dense vegetation and become scattered, making SM estimation in these areas more difficult. Similarly, Özerdem et al. [169] used RADARSAT-2 features and ML models to estimate SM across various land cover types in Turkey. The study found that the models performed well in sparsely vegetated areas (R² of 0.80), but their performance declined in densely vegetated areas (R² of 0.74). This decrease in performance was attributed to the scattering effect of the SAR signal caused by dense vegetation, which reduces the correlation between backscattering coefficients and SM content. Wang et al. [64] estimated SM at different phenological stages of wheat (emergence, tillering, overwintering, and standing) in China using ML models. Backscattering coefficients from Sentinel-1 and vegetation indices from Sentinel-2 were used as input features for the models. The results revealed that the performance of the ML models declined as the NDVI of wheat increased. The best performance of the ML models was observed during the emergence stage of wheat (low NDVI), while the worst performance was observed during the standing stage (high NDVI). This decrease in performance can be attributed to the dense vegetation at higher NDVI values, which limits the ability of SAR signals to penetrate the canopy and retrieve accurate SM information. Similarly, Peng et al. [84] estimated SM in the USA across different land covers (cropland, pasture, grassland, shrub, forest, barren, wetland, and developed) and found that ML models performed well in barren land but showed decreased performance in crops, grasslands, and forest areas. Compared to barren land, the RMSE for SSM increased by approximately 69% in cropland, 73% in grassland, and 85% in forest areas, indicating a substantial decline in model performance under denser vegetation cover. Li and Yan [121] estimated SM in China using Sentinel-1 and Sentinel-2 and found that low performance of ML models was observed in forest and croplands.

To improve the performance of ML models, it is recommended to account for the vegetation effect on SAR signals. Yang et al. [135] conducted a study on SM estimation in the Tibetan Plateau, where the vegetation effect was removed using the Water Cloud Model (WCM). The results showed that removing the vegetation effect increased R² between SM and SAR backscattering coefficients from 0.38 to 0.50. This enhanced correlation between backscattering coefficients and SM contributes to reducing the prediction error of ML models. To minimize the impact of vegetation on SAR signals and enhance the performance of ML models, future studies could integrate empirical and semi-empirical models (Figure 1) alongside ML techniques. Additionally, L-band SAR (e.g., NASA’s NISAR, ALOS-2 PALSAR) offers a promising alternative, as it has a longer wavelength and greater penetration capability through vegetation cover compared to C-band SAR (e.g., Sentinel-1 and RADARSAT) [212].

3.6. ML Model Transferability Challenges and Solutions

Transfer of ML models trained in one geographic location (which may have unique climate, land cover, soil types, and environmental conditions) to new locations offers the ability to reuse learned patterns, minimizing the need for extensive retraining, and enabling practical application in regions where local data may be scarce or expensive to collect. Several studies have evaluated the transferability of ML models for SM estimation and observed that models often face challenges in accurately capturing SM dynamics when applied to new locations with differing environmental conditions.

Holtgrave et al. [165] trained SVR models using Sentinel-1 derived features at two separate sites in Germany (Elbe and Peene) and tested their transferability by applying the model trained at one site to the other site. When the model trained at Peene was tested at Elbe, it produced a low R² of 0.41 and a high root mean square error (RMSE) of 47%. Similarly, the model trained at Elbe and tested at Peene performed poorly, with R² of 0.12 and RMSE of 50%. However, when models were trained and tested within the same site, they performed significantly better, achieving R² of 0.86 with RMSE of 9.6% at Elbe and R² of 0.78 with RMSE of 11.6% at Peene. Ahmad et al. [99] trained ML models using data from six sites in the Lower Colorado River Basin, USA, and tested their performance on four separate sites that were excluded from the training process. The results showed that the models struggled to accurately estimate SM at the test sites, performing significantly worse compared to models trained using data from all sites combined. Souissi et al. [213] estimated SM on a global scale using data from multiple networks within the International Soil Moisture Network (ISMN). The study found that models trained on data from a single network performed poorly when applied to different geographic locations. This result illustrates transferability issues for ML models across regions.

To overcome the model transferability issues, a few studies used transfer learning techniques. For example, Jiaxin et al. [109] estimated SM in Canada under various land-cover conditions and soil roughness using ML models. The models performed poorly when trained and tested on different phases, with RMSE exceeding 0.06 cm³/cm³. To address this issue, the study employed various transfer learning techniques, including target domain adaptation, scattering models, and clustering models, achieving improved RMSE ranging from 0.032 to 0.045 cm³/cm³. These findings suggest that future studies should incorporate transfer learning to enhance model transferability. Hemmati and Sahebi [63] applied the concept of transfer learning to estimate SSM and address the challenge of training deep learning (DL) models at the local level with limited data. They trained a DL model using ISMN data followed by a fine-tuned process using locally collected datasets. Their results showed that transfer learning improved R² by 34% and reduced the RMSE by 18% compared to models without transfer learning. Fine-tuning is generally applied in a context where local SM samples are limited but pre-trained knowledge from a larger domain (ISMN data) is available. This technique is particularly applicable when the source and target domains share similar features (e.g., input variables like SAR and NDVI), but the target domain has too few samples to train a deep model from scratch. It is most effective when the model’s initial training can generalize well and only needs minor adjustments to adapt to new conditions.

Another way to improve SM estimation accuracy is to train ML models using data from homogeneous climatic, soil, and land-cover conditions. ML models may struggle when trained on data from diverse climates, soil, and land-cover types, leading to increased uncertainty in predictions. Moosavi et al. [151] estimated SM across various land-cover types in Iran and found that training ML models on homogeneous regions—characterized by similar land cover, elevation, and physiographic factors—substantially improved model performance compared to training the models on the entire dataset at once. ML models often struggle to identify clear patterns when trained on large-scale datasets (having heterogeneous climate, soil, and land covers) as a whole. To address this, Karthikeyan and Mishra [95] estimated multi-layer SM at a 1 km spatial resolution across the contiguous United States (CONUS) by training models on 11 distinct homogeneous clusters based on climate, land cover, vegetation, and soil types. The results showed that the models effectively captured SM variability across all sites, with most stations achieving RMSE below 0.04 cm³/cm³. This approach indicates the effectiveness of training ML models on homogeneous regions to enhance their predictive performance.

3.7. In Situ SM Measuring Techniques and Uncertainty

Each sensor used for SM measurement has inherent uncertainties that can influence the accuracy of the collected data. These uncertainties, when propagated through the training and validation processes of ML models, can impact the reliability of the model outputs. Therefore, it is crucial to identify measurement techniques that minimize uncertainty while also evaluating their efficiency in terms of time and cost.

We found that 80 out of 144 papers explicitly mentioned the in situ SM measurement techniques used to train and validate ML models with RS data. Figure 8 illustrates the frequency of each technique employed in these 80 studies. The oven-drying/gravimetric method was the most frequently used SM measurement technique (30% of the studies), followed by TDR (22.5%), FDR (5%), Theta Probe (5%), and Stevens HydraProbe (5%). These measurement techniques encounter challenges such as inconsistent performance across varying soil types, discrepancies between laboratory and field calibrations compared to manufacturers’ specifications, and limitations in the accuracy of calibration equations [214]. While the gravimetric (oven-drying) method is the most accurate, its destructive nature and inability to provide continuous measurements limit its utility for long-term monitoring [30,215]. Among modern techniques, neutron probes offer high accuracy with a short response time (1–2 min); however, they are less sensitive to SSM (<30 cm), have limited portability, and incur high setup costs [216,217]. Commercially available TDR, FDR, and capacitance sensors provide an accuracy range of ±1–3% [218]. The advantages, disadvantages, and uncertainties associated with various SM measurement techniques are comprehensively documented in Dobriyal et al. [219], Robinson et al. [220], and Su et al. [214]. Future studies should select SM measurement techniques based on their accuracy, response time, and associated costs to optimize the training and validation of ML models.

3.8. Uncertainty and Future Research Directions in SM Estimation

Uncertainty in SM estimation using ML and RS arises from multiple, often interrelated sources, including input data quality, spatial and temporal scale mismatches, and methodological limitations. A fundamental source of uncertainty is the limited generalizability of ML models. Models trained under specific biophysical and climatic conditions may perform poorly when applied to regions with different vegetation types, soil textures, or land management practices [99,165,213]. This limitation is particularly pronounced in models that rely heavily on vegetation indices or microwave backscatter signals, as these inputs are sensitive to surface roughness, canopy structure, and seasonal dynamics, all of which vary across space and time. The quality of RS inputs further contributes to uncertainty. Satellite-derived products can be affected by atmospheric interference, sensor calibration errors, and cloud contamination, particularly in optical imagery [221,222]. Another critical source of uncertainty lies in the in situ measurements used for model training and validation. Different SM sensors (e.g., TDR, FDR, gravimetric) have varying levels of precision, spatial representation, and depth sensitivity, and these differences introduce inconsistencies in the ground measurement data [214,219,220]. When such measurement errors propagate through the modeling process, they can significantly influence the predictive reliability of ML algorithms. Additionally, differences in spatial and temporal resolution across remote sensing datasets introduce uncertainty in SM estimation by causing mismatches between input features and ground truth data, leading to aggregation errors and misaligned observations that affect model accuracy and reliability [223,224]. These combined sources of uncertainty underscore the need for careful model design, transparent data reporting, and improved integration strategies to enhance the robustness and transferability of SM estimation frameworks.

While substantial progress has been made in SM prediction using RS and ML, several critical avenues remain to be explored. Future studies should prioritize the development and application of hybrid models that combine multiple ML or deep learning architectures to reduce bias–variance tradeoffs and improve predictive accuracy across different soil depths and land covers. There is also a growing need for advanced multi-sensor data fusion strategies that leverage the complementary strengths of optical, microwave, thermal, and UAV-based datasets. Transfer learning and domain adaptation techniques offer promising solutions to address the widely observed issue of poor model generalization across geographic regions and climatic zones. In parallel, physics-informed machine learning frameworks, which integrate hydrological constraints and soil physics knowledge into data-driven models, can improve model interpretability and reliability, especially in data-scarce regions. Furthermore, explainability and uncertainty quantification remain underexplored in SM modeling; tools such as SHapley Additive exPlanations, Bayesian frameworks, and dropout-based uncertainty estimation can enhance model transparency and practical usability. Finally, standardization in reporting ground-truth measurements is essential; many reviewed studies lacked details on in situ SM collection methods, limiting reproducibility and cross-study comparisons. Establishing common protocols and expanding open-access benchmark datasets will be crucial to support the next generation of transferable, interpretable, and operationally reliable SM prediction models.

4. Conclusions

This analysis aims to systematically evaluate a range of ML studies applied to the estimation of SM across diverse land covers and geographic regions. Use of ML approaches offers a promising alternative to conventional empirical and physics-driven models, which require extensive biophysical parameterization. Our analysis extends from traditional ML techniques, including support vector machine (SVM) and tree-based models, to the incorporation of advanced multi-classifiers, ensemble frameworks, and state-of-the-art DL models. These recent developments have significantly improved the accuracy of SM predictions at various depths through the application of RS technologies.

Many ML algorithms (Figure 5) were employed to estimate SM using RS datasets. Based on the model performance mentioned in the studies, RF (40 out of 67 times), SVR (13 out of 39 studies), and ANN (12 out of 27 times) often outperformed other algorithms. The selection of input features in ML models depends on whether SSM or RZSM is the focus. All 144 studies reviewed here used a range of input features, including vegetation indices derived from optical sensors, SAR features from microwave sensors, soil physical and hydraulic properties, and topographical features. Some features were utilized more than others. For SSM estimation (out of 119 studies), NDVI (60 studies), elevation (40 studies), and backscattering coefficients (VV: 32 studies and VH: 27 studies) were the most common predictors. For RZSM estimation (out of 25 studies), SSM (13 studies) and NDVI (13 studies) were the most common predictors.

Per this synthesis (Section 3.4.1), for SSM estimation, satellite-derived features such as vegetation indices and backscattering coefficients provide critical information on SSM variability. In contrast, satellite observations alone were insufficient to capture RZSM variability. Soil physical and hydraulic properties, along with SSM, were identified as key predictors of RZSM. Integrating multi-source data consistently enhanced the performance of ML models in SM estimation. Multi-source data includes satellite and other remote sensing products (Sentinel-1, MODIS, Sentinel-2, Landsat, SMAP, and UAV) to derive input features to estimate SM.

Though identifying the best in situ SM method for ground-truth data is out of the scope of this review, we note that only 80 studies out of 144 reported the methods used to measure in situ SM. We advocate that reporting of the method used to collect SM training data is essential to assess RS method performance. Gravimetric (24 out of 80 studies) and TDR (18 out of 80 studies) methods were most frequently used (refer to Section 3.7).

Despite multiple advantages of SM estimation using ML and RS techniques, transferability of models trained at one site to new sites remains a major challenge (Section 3.6). Studies (Section 3.5) also identified that vegetation cover affects SAR-based SM estimation, particularly in the X- and C-bands, and improved performance was observed when removing vegetation effects on SAR signals using semi-empirical models. Hybrid ML models (combining two or more ML models: Section 3.1) were investigated and found to improve SM estimation accuracy, reducing the bias–variance tradeoff. Future research should prioritize exploring and validating these potential approaches (use of semi-empirical models, hybrid ML, and Transformer) to effectively address the challenges identified and overcome the existing limitations.