Space–Air–Ground Synergistic Approaches for Field Water Status Precision Monitoring: A Review

Highlights

What are the main findings?

• There are two core issues in space–air–ground field water status monitoring: unresolved scale mismatch and error propagation in multi-source data fusion, and a lack of reliable quantitative retrieval for root zone soil moisture and paddy field water depth.

What are the implications of the main findings?

• This study provides a methodological reference framework for field water status monitoring tailored to different agricultural scenarios.
• Two major methodological challenges, scale transformation and error quantification, and two critical research gaps, root zone soil moisture retrieval and paddy field water depth retrieval, are identified.

Abstract

Field water status is a critical variable for agricultural water management. In recent years, the development of space–air–ground multi-platform collaborative observation and data fusion technologies has provided new options for precision monitoring. However, challenges in applicability, robustness, and transferability persist. This study employs bibliometric analysis to systematically synthesize the literature, revealing that research has evolved from single-point observations to multi-platform synergy. Satellite, unmanned aerial vehicle (UAV), and ground-based monitoring are analyzed, as well as challenges in multi-source data fusion, including scale mismatch, error propagation, and uncertainty quantification. Finally, applicability and other barriers are evaluated across three typical agricultural scenarios: large-scale surface soil moisture monitoring, crop root zone soil moisture retrieval, and paddy field water depth estimation. The results indicate that space–air–ground collaborative observation constitutes a mature framework, with satellite and ground-based monitoring as core components and UAV technology as a supplement. However, scale transformation and error propagation mechanisms in multi-source data fusion remain unresolved. Currently available vertical water information is limited, and quantitative retrieval has yet to achieve the reliability required for operational applications. This limitation is particularly evident in paddy field water depth retrieval and root zone soil moisture retrieval. This review provides a theoretical reference for precision field water status monitoring and identifies future research priorities, including the integration of physical mechanisms with machine learning (ML) in multi-source data fusion, as well as error quantification and paddy field water depth retrieval.

1. Introduction

Field water status is a core state variable in the agricultural water cycle, directly governing crop growth, yield formation, and water use efficiency [1,2]. Under the mounting pressures of global water scarcity and climate-driven hydrological extremes, precise field water management is essential for food security and sustainable agriculture [3]. Globally, agriculture accounts for approximately 70% of freshwater withdrawals, yet irrigation efficiency remains below 50% (FAO) [4]. Achieving accurate, large-scale, and dynamic field water status monitoring is a pressing scientific and practical challenge [5].

Traditional monitoring has relied on point-scale in situ methods such as oven drying, time domain reflectometry (TDR), and frequency domain reflectometry (FDR). Although these techniques provide high-accuracy ground truth data for calibration and validation [6], their utility for regional-scale agricultural water management is constrained by poor spatial coverage, high deployment costs, and limited ability to capture spatial heterogeneity [7]. Remote sensing has addressed these limitations. Satellite platforms carrying optical, thermal infrared, and microwave sensors deliver periodic global observations [8], with the latter valued for its all-weather capability and high sensitivity to soil moisture [9]. Concurrently, the rapid development of unmanned aerial vehicle (UAV) remote sensing has bridged the scale gap between satellite and ground-based observations, enabling fine-grained monitoring at centimeter-level resolution in fields [10].

Integrated space–air–ground multi-platform collaborative observation systems encompass satellite, UAV, and ground-based observations. Despite technological progress, however, no single platform can simultaneously satisfy agricultural demands for high accuracy, high spatiotemporal resolution, large-scale coverage, and adequate detection depth, given differences in physical principles and observation modes [11]. Satellite microwave data, for example, offer high temporal resolution but suffer from coarse spatial resolution that precludes the characterization of field-scale spatial heterogeneity [12]. Optical data provide high spatial resolution but are susceptible to cloud cover and limited to sensing the shallow surface layer [13]. UAV data deliver exceptional spatial detail, but they are constrained by limited coverage, high acquisition costs, and inconsistent processing standards [14]. Consequently, the effective fusion of multi-source data to exploit the complementary strengths of different platforms is essential for enhancing field water monitoring [15].

Significant progress has been made in multi-source data fusion in recent years [16]. Researchers have developed various frameworks, including pixel-level, feature-level, and decision-level fusion [17], and explored synergistic approaches such as optical–microwave synergy [18], active–passive microwave combinations [19], and the integration of data assimilation with machine learning [20]. These methods aim to combine observational data from disparate sources and scales into coherent products characterized by spatiotemporal continuity, improved accuracy, and physical consistency [21]. However, challenges in the fusion process, including scale mismatch, error propagation, and uncertainty quantification, have yet to be fully resolved, limiting fused products’ reliability and generalizability [6].

Against this background, existing reviews have summarized remote-sensing-based field water status monitoring from various perspectives. Some have focused on soil moisture retrieval methods for a single platform, comprehensively outlining principles and progress across optical, thermal infrared, and microwave technologies [22,23,24]. Others have focused on theoretical frameworks for multi-source data fusion or water monitoring for specific regions [25,26,27]. While this work has laid a solid foundation, most reviews tend to concentrate on specific methods or data sources, lacking an integrated perspective that considers the interplay between data types, platform characteristics, and agricultural application scenarios. Furthermore, they predominantly address soil moisture in upland crops, often overlooking the unique challenges presented by paddy fields [28]. Accordingly, this study reviews space–air–ground integrated field water status monitoring and data fusion methods by focusing on the following three research questions. (1) What are the technical principles, advantages, and inherent limitations of satellite, UAV, and ground-based platforms? How can these three platforms achieve synergistic benefits through collaborative observation? (2) What are the primary multi-source data fusion methods? What solutions have existing studies proposed to tackle issues such as scale mismatch, error propagation, and uncertainty quantification? (3) What are the research progress, applicable scopes, and bottleneck issues across three typical agricultural application scenarios: large-scale surface soil moisture monitoring, crop root zone soil moisture retrieval, and quantitative estimation of paddy field water depth?

The remainder of this paper is organized as follows. Section 2 introduces the methodology, as well as bibliometric analysis results to reveal research trends and hotspots. Section 3 discusses space–air–ground integrated field water status monitoring technical systems and multi-source data fusion methods. Section 4 reviews progress in the three aforementioned agricultural application scenarios. Finally, Section 5 summarizes current research challenges and prospects for future development.

2. Data Sources and Bibliometric Analysis

2.1. Data Retrieval and Screening

To establish a comprehensive corpus that represents space–air–ground research, a systematic literature search was performed in the Web of Science (WoS) Core Collection on 27 January 2026, covering peer-reviewed publications from 1995 to 2025. This strategy consisted of two complementary phases (

Table 1. Literature search strategy and screening results.

Search Objective	Query Structure	Core Search Query (Topic Field in WoS)	Initial Results	Screened Results
Space-based (satellite)	(“soil moisture” OR “surface soil moisture” OR “root zone moisture” OR “water content” OR “water depth” OR “flooding” OR “ponding water”) AND (……) AND (“agricultur*” OR “farmland” OR “cropland” OR “irrigation” OR “drainage” OR “water management” OR “paddy” OR “rice field”)	(same prefix) AND (“remote sensing”) AND (same suffix)	3623	3023
Air-based (UAV)		(same prefix) AND (“UAV” OR “drone” OR “unmanned aerial vehicle”) AND (same suffix)	413	343
Ground-based (in situ)		(same prefix) AND (“in situ” OR “sensor network” OR “ground-based” OR “point measurement”) AND (same suffix)	3085	2537
Integrated search		(same prefix) AND (“remote sensing” OR “UAV” OR “drone” OR “unmanned aerial vehicle” OR “in situ” OR “sensor network” OR “ground-based” OR “point measurement”) AND (same suffix)	6017	4951

). Only articles, review articles, and publications in English were retained for analysis.

In both phases, a core set of moisture-related terms, including “soil moisture,” “surface soil moisture,” “root zone moisture,” “water content,” “water depth,” “flooding,” and “ponding water,” was combined with agricultural context terms, such as “agricultur*,” “farmland,” “cropland,” “irrigation,” “drainage,” “water management,” “paddy,” and “rice field,” to ensure relevance to agricultural applications. Phase 1 encompassed platform-specific searches to capture the literature on space-based (satellite), air-based (UAV), and ground-based (in situ) monitoring. Phase 2 included the overarching literature on synergistic approaches through an integrated search for publications on data fusion and multi-platform integration that might have been missed in the first phase. After merging all retrieved records and removing duplicates, the final bibliometric corpus was formed. This two-phase design captured the literature on each technology and ensured comprehensiveness.

2.2. Bibliometric Analysis Results

To visualize the research landscape, keyword co-occurrence networks were generated using VOSviewer (version 1.6.20), and annual publication trends were analyzed (Figure 1). The findings revealed that this field developed through three distinct phases. In the initial stage (1995–2005), scholarly output was modest, averaging fewer than 44 articles per year, and research was concentrated on theoretical exploration and preliminary single-platform applications. UAV-related studies were virtually absent, with only three articles published during the entire period. The steady growth phase (2006–2015) saw annual publications rise to 65–160, with satellite and ground-based monitoring as dual pillars (~95 articles/year each) and early conceptions of multi-platform collaboration emerging. The exponential growth phase (2016–2025) witnessed a sharp increase to 671 articles in 2025. UAV technology saw explosive growth at an annual rate of 35.7% (99 articles in 2025), emerging as a critical link that bridged the scale gap. An inflection point occurred around 2019, after which growth steepened markedly.

Keyword networks also revealed distinct yet converging platform-specific emphases across themes (Figure 2). The satellite-specific network (Figure 2b) is dominated by “remote sensing,” “soil moisture,” and “model,” with high-frequency terms including Sentinel-1/2, MODIS, SMOS, SMAP, data assimilation, retrieval, validation, NDVI, agricultural drought, and evapotranspiration. These nodes cluster around satellite data sources, soil moisture retrieval algorithms, agricultural applications, and environmental response, reflecting a focus on multi-sensor synergy for global and regional agricultural monitoring. The UAV network (Figure 2c) is characterized by “UAV,” “vegetation index,” and “soil moisture,” with prominent terms such as multispectral imagery, machine learning, canopy temperature, water stress, leaf area index, and crop yield, concentrating on field-scale crop water stress diagnosis and precision agriculture. The ground-based network (Figure 2d) features “soil moisture,” “remote sensing,” “model,” and “water content” as central nodes, with in situ observation, validation, SMOS/SMAP, sensor network, data assimilation, downscaling, and temporal stability forming the main clusters, underscoring its core role in providing ground truth constraints for satellite product calibration and validation. Across all platform-specific networks, cross-cutting themes such as machine learning and precision agriculture appear prominently, signaling that data-driven multi-platform collaborative observation has become the mainstream research direction.

Although our search strategy included terms related to paddy field water depth (water depth, flooding, ponding water), these terms were not prominent nodes in the keyword co-occurrence network. This absence confirms that field water status research remains overwhelmingly dominated by upland soil moisture, while paddy field water depth, a distinct and challenging scenario, has yet to accumulate sufficient research mass to form a high-frequency keyword cluster. Overall, based on the retrieval results described in Section 2.1, the research landscape exhibits a dual-core structure: satellite remote sensing and ground-based monitoring serve as the established pillars (accounting for 61.1% and 51.2% of total publications), while UAV technology (6.9%) represents a dynamic new growth pole. The technological trajectory has evolved from early single-point exploration to an intelligent, multi-platform integration paradigm driven by machine learning and collaborative observation.

3. Space–Air–Ground Integrated Technical System for Field Water Status Monitoring and Data Fusion Methods

3.1. Data Sources for Field Water Status Monitoring

3.1.1. Ground-Based Monitoring for Field Water Status Data Acquisition

In situ ground-based monitoring provides absolute reference values (ground truth) for calibrating and validating remote sensing products and land surface process models [29,30,31]. The evolution of this field has been driven by a core challenge: achieving an optimal balance between measurement accuracy, spatial representativeness, temporal continuity, and implementation cost. To address this, ground-based monitoring technologies have been integrated into multi-level, multi-principle, collaborative observation systems [32]. These technologies are based on gravimetric methods (mass determination), volumetric methods (assessing water content from the medium’s physical properties), and potentiometric methods (soil water potential measurement). These methods include direct and indirect water content measurements [33].

Among the direct methods, oven drying is internationally recognized as the “gold standard.” It calculates soil water content by measuring the mass difference in soil samples before and after drying, with errors typically controlled within 2% [32]. However, inherent limitations, including destructive sampling, labor-intensive operations, and inability to provide continuous observations, confine its role to a calibration benchmark for other indirect sensing technologies [34]. This method directly yields gravimetric water content (GWC, mass of water per mass of dry soil), whereas most other in situ electronic sensors and remote sensing retrievals provide volumetric water content (VWC, volume of water per volume of soil). Conversion between GWC and VWC requires knowledge of soil bulk density. Consequently, contemporary automated field monitoring primarily relies on indirect methods that measure variables with a well-established physical correlation to soil moisture [35,36]. Among these, the volumetric method, which aims to directly obtain volumetric soil water content, is the mainstream approach for current point-scale field monitoring [11].

The volumetric method encompasses several technical pathways, with their core characteristics summarized in

Table 2. Comparison of ground-based monitoring technologies.

Category	Technology	Depth (cm)	Spatial Scale	Temporal Resolution	Accuracy	Main Advantages	Key Limitations
Direct Methods	Oven drying	Arbitrary	Point	Hour to day (discrete)	±1–2 vol% (absolute)	Absolute benchmark	Destructive, non-real-time, labor-intensive
Volumetric Methods (Indirect)	TDR	0–30	Point	Second–minute (continuous)	±1–3 vol% (RMSE)	High accuracy, fast response	High cost, sensitive to installation conditions
	FDR	0–30	Point	Second–minute (continuous)	±2–5 vol% (RMSE)	Low cost, easy to network	Susceptible to salinity, requires calibration
	ADR	0–30	Point	Second–minute (continuous)	±2–5 vol% (RMSE)	Moderate cost, easy to integrate	Requires calibration, variable performance
	Time domain transmissometry (TDT)	0–30	Point	Second–minute (continuous)	±1–3 vol% (RMSE)	Performance close to TDR, high stability	Requires calibration, variable implementation performance
	Electrical capacitance tomography (ECT)	10–50	Small-scale 3D	Second–minute (continuous)	±2–6 vol% (RMSE)	3D imaging, dynamic observation	Limited resolution, difficult to apply at field scale
	Neutron scattering	0–60	Point profile	Hour	±2–5 vol% (RMSE)	Reliable profile measurement, high accuracy	Radioactive risk, cumbersome operation
	Gamma attenuation	5–50	Point 1D profile	Minute–hour	±2–5 vol% (RMSE)	Non-destructive, continuous profile	Background interference, requires calibration
	Nuclear Magnetic Resonance	0–30	Point	Laboratory level	±1–2 vol% (absolute)	Distinguishes water states, strong physical basis	Extremely high cost, not field-deployable
	CRNS	0–70	Mesoscale (130–240 m radius)	Minute–hour	±3–5 vol% (RMSE)	Continuous mesoscale monitoring	Greatly affected by vegetation and environmental factors
Potentiometric Methods (Indirect)	Tensiometer	5–40	Single point	Minute–hour	±1–5 kPa (water potential)	Measures plant-available water, high sensitivity	Significant maintenance required, fails in arid areas
	Thermal dissipation method	5–30	Single point	Minute–hour	±1–10 kPa (water potential)	Fast response, suitable for irrigation monitoring	Requires calibration, non-linear deviation
	Gypsum resistance sensor	5–50	Single point	Hour–day	±5–20 kPa (water potential)	Extremely low cost, easy to use	Low accuracy, slow response, susceptible to salinity
Network Platforms	ISMN	0–100	Station network	30 min–h	±1–5 vol% (site-dependent RMSE)	Standardized ground truth	Limitations to station representativeness
	China National Validation Field Network	0–100	Representative area	10 min–h	±1–5 vol% (site-dependent RMSE)	Supports domestic satellite validation	Significant station maintenance and deployment costs

. Electromagnetic dielectric techniques (e.g., time domain reflectometry (TDR), frequency domain reflectometry (FDR), amplitude domain reflectometry (ADR)) retrieve soil water content by measuring the soil dielectric constant [37,38]. They offer high accuracy (errors as low as 2–3%) and enable automated continuous recording. Their fundamental limitation, however, lies in the poor spatial representativeness of point measurements. Capturing field heterogeneity requires dense deployment, leading to high costs [39,40]. Nuclear techniques (e.g., neutron scattering, gamma attenuation) exploit hydrogen atoms’ moderation effect on neutrons and gamma rays’ attenuation characteristics, enabling fast responses and non-destructive measurements [41]. Due to equipment costs and radioactive safety management requirements, their application is typically restricted to constructing small-scale, high-precision validation networks [42]. Cosmic-Ray Neutron Sensing (CRNS) is a revolutionary mesoscale observation technique that can retrieve average water content over a radius of hundreds of meters (approximately 10–100 hectares) by measuring the surface fast neutron flux [43]. As shown in Table 2, its key advantage lies in its ability to characterize regional water status with very few stations. However, its detection depth is relatively fixed (approximately 0–70 cm) [44], and it is sensitive to non-soil hydrogen sources, such as vegetation biomass and snow water [45].

Potentiometric methods (e.g., tensiometers and thermal dissipation sensors) directly measure soil water potential [46], intuitively reflecting the ease with which crop roots can extract water. This physiological information is highly valuable for irrigation decision making and plant water stress diagnosis. Nevertheless, its measurement range is highly localized, response times are long, and inferring the spatial pattern of soil moisture typically requires complex networked deployment [47].

At global and regional scales, multiple ground-based sensing networks have been established. The International Soil Moisture Network (ISMN) has become the world’s leading reference database for in situ soil moisture, integrating data from 82 networks and 3200 stations worldwide [48]. It provides high-quality ground truth to support remote sensing validation and fusion [49]. In China, the National Civil Space Infrastructure Validation Field Network and the National Validation Station Network for the Gaofen (GF) Special Project have jointly established a validation system comprising over 60 stations capable of validating more than 40 types of remote sensing products [50].

In summary, ground-based monitoring networks’ density, data quality standards, and scale-representative design determine field water status remote sensing accuracy at regional and global scales. Future progress will depend on developing low-cost, low-maintenance sensor nodes with wireless transmission capabilities along with collaborative observation protocols synchronized with satellite overpasses to enhance space–air–ground system efficiency.

3.1.2. Remote-Sensing-Based Field Water Status Data Acquisition

In the space–air–ground integrated monitoring system, remote sensing technology is used to dynamically perceive water status at the regional and global scales [35]. This system uses optical, thermal infrared, and microwave remote sensing to detect electromagnetic waves’ interactions with the land’s surface [8]. These technologies do not operate in isolation; instead, they collaborate with unmanned aerial vehicles (UAVs) and near-ground platforms to form a multi-level observation network spanning centimeters to kilometers [51]. As summarized in Figure 3, the platforms and sensors exhibit significant differences in spatial resolution, temporal resolution, detection depth, and sensitivity to water content. Satellite platforms provide large-scale, periodic coverage, making them well-suited for macro-monitoring and trend analysis. UAV platforms, with their high flexibility and spatiotemporal resolution, fill the scale gap between satellite and ground-based observations, enabling fine-grained monitoring at the field scale down to centimeter-level resolution.

Optical Remote Sensing

Optical remote sensing retrieves water content by analyzing changes in the spectral reflectance of the soil–vegetation system, leveraging water molecules’ absorption characteristics in specific wavelength bands [52]. It includes direct spectral modeling for bare soil or low-vegetation-coverage areas and indirect vegetation index diagnosis for high-vegetation-coverage areas.

In areas with low vegetation coverage, quantitative models can be directly established between soil reflectance spectra and water content [53]. These methods have evolved from early statistical empirical models into physically based approaches. Two representative models are the SOILSPEC transmission model developed by Jacquemoud et al. [54] and the simplified empirical model of exponential negative correlation between reflectance and soil moisture proposed by Lobell et al. [55]. Although these models collectively explain the regulatory effect of soil moisture on spectral reflectance, they lack in-depth physical interpretation [56]. In contrast, the spectral simulation model of wet soil reflectance proposed by Philpot et al. [57] reveals the direct correlation between soil moisture and reflectance from an optical physics perspective. This model decomposes wet soil reflectance into two components: specular reflection from the water film and diffuse reflection from soil particles. It constructs a forward model capable of simulating the reflectance spectra of soils with varying water content with accuracy in the 380–1000 nm range. However, optical signals only capture information from the surface’s top few millimeters and centimeters. They are also highly susceptible to atmospheric interference, particularly cloud cover, which limits observation frequency [58].

When vegetation coverage is high, researchers indirectly diagnose soil moisture status by constructing vegetation indices sensitive to water stress. These indices include classic ones such as the Normalized Difference Vegetation Index (NDVI) [59] and the Enhanced Vegetation Index (EVI) [60], as well as improved indices that incorporate short-wave infrared bands for enhanced sensitivity, such as the Land Surface Water Index (LSWI) [61] and the Visible and Shortwave Infrared Drought Index (VSDI) [62]. These indices estimate soil moisture indirectly by capturing spectral changes associated with crop stomatal closure; however, their accuracy is modulated in complex ways by crop type, growth stage, and environmental stress factors [63].

A notable advantage of optical remote sensing is its high spatial resolution (from meter to sub-meter levels) and rich spectral information, which makes it well-suited for fine-scale field monitoring. Its limitations, however, include high sensitivity to atmospheric conditions, signals that represent only the shallow surface layer, and significant uncertainty in the indirect relationship with soil moisture under dense vegetation cover.

Thermal Infrared Remote Sensing

Thermal infrared remote sensing departs from the spectral reflectance framework and diagnoses water status from a surface energy balance perspective. It is based on water’s high specific heat capacity and latent heat from vaporization [64]. Soil moisture content influences Land Surface Temperature (LST) by regulating the partitioning between latent heat flux (evaporation/transpiration) and sensible heat flux. When water is abundant, more energy is allocated to evaporation, producing a strong surface cooling effect and lower LST; conversely, LST increases under water-limited conditions [65].

Models based on thermal inertia are widely used. They retrieve water content from the relationship between diurnal temperature variation and soil thermal properties, and they perform well in bare soil or sparsely vegetated areas [66,67]. The thermal stress index method is also frequently used, exemplified by the Crop Water Stress Index (CWSI), which directly reflects the degree of root water uptake limitation by quantifying the deviation between canopy temperature and air temperature, providing an indicator for irrigation decision making [68].

The unique value of thermal infrared technology lies in its direct sensing of energy fluxes related to crop water stress, which offers an independent observational dimension distinct from optical and microwave remote sensing [69]. However, it is also affected by cloud cover, and surface temperature is influenced by meteorological conditions (wind speed, humidity, and radiation) and surface heterogeneity. Consequently, accurately translating temperature signals into reliable water content estimates remains a complex challenge [70].

Microwave Remote Sensing

Microwave remote sensing exploits the large contrast in dielectric constant between liquid water and dry soil. Its longer wavelengths can penetrate clouds and, to some extent, vegetation canopies, making it an indispensable tool for all-weather, large-scale water monitoring [71,72]. Interactions between electromagnetic waves and soil–vegetation systems are primarily controlled by wavelengths [73]. Longer wavelengths offer stronger penetration, lower sensitivity to vegetation and surface roughness, and higher sensitivity to soil moisture. Shorter wavelengths exhibit the opposite characteristics and are more susceptible to vegetation and roughness interference [74]. These physical differences govern the applications of each frequency band in soil moisture retrieval.

Among commonly used microwave bands, the P-band represents the longest wavelength currently available in spaceborne Synthetic Aperture Radar (SAR) systems (0.3–1 GHz, wavelength approximately 30–100 cm). It possesses exceptionally strong penetration ability and can theoretically detect deep-layer soil moisture [75]. However, spaceborne P-band data remain scarce, and their spatial resolution is relatively coarse. At present, P-band observations rely primarily on airborne platforms such as Airborne Microwave Observatory of Subcanopy and Subsurface (AirMOSS) [76]. The L-band (1–2 GHz, wavelength approximately 15–30 cm) is widely regarded as the optimal band for soil moisture detection. It can penetrate approximately 5 cm below the soil’s surface with a high signal-to-noise ratio, and both the Soil Moisture Active Passive (SMAP) and Soil Moisture and Ocean Salinity (SMOS) operate in this band [77]. However, the spatial resolution of passive L-band data is coarse (36–40 km), which limits its applicability at the field scale [78]. Compared with the L-band, the C-band (4–8 GHz, wavelength approximately 5.6 cm) trades penetration ability for higher spatial resolution. For instance, Sentinel-1 can achieve a resolution of 5 m × 20 m, but the C-band penetrates only approximately 2 cm and is significantly affected by surface roughness and vegetation cover, often requiring correction methods such as the water cloud model (WCM) [79]. As the wavelength further shortens to the X-band (8–12 GHz, wavelength approximately 3 cm), the penetration depth drops to less than 1 cm. The X-band is highly sensitive to vegetation and roughness and is therefore suitable only for surface moisture monitoring over bare soil or sparse vegetation [80].

The complementarity between multiple frequency bands is a prerequisite for designing fusion schemes and interpreting error sources. Successful joint retrieval using SMAP L-band passive data and Advanced Scatterometer (ASCAT) C-band active data has been reported, and the physical basis of this strategy lies in the complementarity between these two bands [81]. Conversely, after the failure of the SMAP active sensor, attempts to substitute the L-band active component with the C-band Sentinel-1 encountered difficulties due to differences between the two frequency bands [82]. Thus, a thorough understanding of how wavelength governs microwave signal behavior is a prerequisite for designing sound fusion strategies and accurately interpreting error sources.

According to their observation mode, microwave remote sensing can be further classified into active, passive, and Global Navigation Satellite System–Reflectometry (GNSS-R) technologies [83].

Active microwave remote sensing (e.g., SAR, Scatterometer) actively transmits and receives surface backscattering signals [84]. Signal intensity is directly related to the soil dielectric constant and can provide high spatial resolution at the meter to decameter level [85]. However, the backscattering coefficient (σ⁰) is also influenced by surface roughness and vegetation structure [86]. Accordingly, the main thrust of technological development has been decoupling the water content signal from these confounding factors, progressing from physically based scattering models [87] to empirical/semi-empirical correction models (e.g., the water cloud model) [88] and, more recently, machine learning approaches that learn complex nonlinear relationships from multi-polarization and multi-temporal data [89]. Currently, integrating physical mechanisms with data-driven models represents the cutting edge, aiming to improve both retrieval accuracy and physical interpretability [90].

Passive microwave remote sensing retrieves soil moisture by receiving naturally emitted microwave brightness temperatures from the land’s surface [91]. Its advantages include a clear physical basis, high temporal resolution (1–3 days), and sensitivity to soil moisture [77]. Missions such as Soil Moisture Active Passive (SMAP) and Soil Moisture and Ocean Salinity (SMOS) have generated operational global products [92]. Its critical limitation is its coarse spatial resolution (typically >25 km), which hinders its applicability to farm-scale management [93]. Consequently, research has focused on spatial downscaling integrating high-resolution optical, SAR, or topographic data to generate finer-scale products [78].

GNSS-R technology, an emerging bistatic radar approach, exploits reflected L-band signals from navigation satellites (e.g., GPS and BeiDou) for surface sensing [94]. It offers free signal sources and dense spatiotemporal sampling [95]. By processing raw waveform data with deep learning, researchers have generated products with kilometer-scale resolution and accuracy comparable to SMAP, complementing passive microwave sensing’s spatiotemporal continuity. Roberts et al. [96] demonstrated that these methods can increase retrieval correlations from approximately 75% using traditional approaches to over 90%. Wernicke et al. [97], by fusing CYclone Global Navigation Satellite System (CYGNSS) reflectivity with SMAP brightness temperature data and employing spatial interpolation techniques to address GNSS-R coverage gaps, produced high-quality products with 3 km spatial resolution and 2–3-day revisit intervals. However, signal attenuation remains a significant challenge in dense forest areas [98]. Additionally, CubeSat constellations are emerging as a new data dimension, offering optical and SAR data with high spatiotemporal resolution [99].

In summary, each technology has inherent strengths and limitations at its respective scale and information dimension, underscoring the inevitability of synergistic fusion. Ground-based monitoring provides calibration benchmarks at centimeter-level accuracy but suffers from poor spatial representativeness and high deployment costs. Passive microwave sensing offers strong temporal continuity but coarse spatial resolution, making it incapable of resolving field-scale heterogeneity. Active microwave and optical data deliver higher spatial resolution but are constrained by complex scattering mechanisms or cloud and rainfall interference, with revisit intervals extending from days to weeks. UAV-based observation is flexible and spatially refined, but its coverage and cost impose practical constraints. More fundamentally, mainstream remote sensing technologies’ detection depth is generally limited to the top few centimeters of the soil’s surface, whereas the water that drives crop transpiration is stored in the root zone. Consequently, no single technical path can simultaneously satisfy the need for high accuracy, spatiotemporal resolution, reliability, and detection depth. Constructing a space–air–ground integrated intelligent sensing system using systematic, collaborative, multi-source data fusion has become a viable path to overcome existing technical bottlenecks.

3.2. Multi-Source Data Fusion Methods: Hierarchies, Methodologies, and Scale Issues

Multi-source data fusion is essential for overcoming the limitations of individual space–air–ground platforms and enabling collaborative retrieval of field water status with high accuracy, spatiotemporal resolution, and reliability. The core challenge lies in harmonizing spatiotemporal matching, physical consistency, and information complementarity [100,101,102]. This paper proposes a methodological decision making framework for multi-source data fusion (Figure 4). The framework is structured around three hierarchical layers, with the fusion level serving as the top-level guide. The input data types encompass three broad categories: multi-platform and multi-band remote sensing data, auxiliary and environmental forcing data, and multi-model and product integration data. Pixel-level, feature-level, or decision-level fusion is selected according to the monitoring objective and data characteristics. Technical pathways form the middle-level support. The observation synergy layer reduces multi-platform data mismatches at the source through spatial site optimization and temporal synchronization design. The algorithmic fusion layer addresses residual spatiotemporal inconsistencies via scale transformation and temporal resampling. The sensor synergy layer selects the optimal cross-sensor combination strategy based on the application scenario. Uncertainty quantification provides confidence constraints for the fused results through layer-by-layer assessment and feedback for aleatoric and epistemic uncertainties.

3.2.1. Fusion Theoretical Framework and Hierarchy Classification

Based on the level of information abstraction, multi-source data fusion is categorized into pixel-level, feature-level, and decision-level fusion. Feature-level and decision-level methods are most widely adopted in soil moisture retrieval [100,103]. Pixel-level fusion operates at the raw data stage, enhancing spatial details through geometric registration and numerical synthesis [104]. Feature-level fusion constructs input vectors from variables with clear physical meaning, significantly reducing dimensionality while preserving key information [105]. Decision-level fusion integrates outputs from multiple sources and offers strong fault tolerance, making it particularly suitable for global soil moisture assessment and cross-regional drought comparison [106].

The source data underpinning fusion can be broadly grouped into three categories. (1) Multi-platform and multi-band data provide complementarity in spatial and temporal resolution, with their signals exhibiting differential sensitivity to surface dielectric properties, vegetation cover, and energy balance processes [101]. (2) Auxiliary and environmental forcing data, such as reanalysis products, DEMs, soil texture and organic matter maps, and vegetation parameters, offer physical constraints and contextual support for interpreting remote sensing signals [107]. (3) Multi-model and product integration data enable spatiotemporally continuous and vertically extended soil moisture estimates by assimilating remote sensing retrievals into land surface process models or reanalysis products [108,109].

Achieving effective fusion across diverse data sources is limited by four recurring challenges. (1) Spatiotemporal scale matching: significant disparities in resolution and revisit frequency across platforms must be reconciled through downscaling or spatiotemporal interpolation [110]. (2) Physical mechanism synergy: data derived from different observation principles must be unified; for example, optically derived leaf area index must be converted into vegetation attenuation coefficients in microwave signals via the water cloud model [111]. (3) Error propagation and uncertainty quantification: inherent errors from individual data sources may compound during fusion, necessitating the identification of error sources through ablation experiments or triple collocation analysis [112,113]. (4) Dynamic information weighting: the relative value of different data sources varies with vegetation phenology, moisture status, and weather conditions, requiring the fusion system to adaptively evaluate and balance the contribution of each input [114].

3.2.2. Main Fusion Technical Pathways

The space–air–ground integrated framework aims to establish a closed-loop system encompassing observation, calibration, modeling, and assimilation. Its realization requires active coordination at the observation design stage to improve data quality and consistency and fusion algorithms capable of addressing spatiotemporal mismatches that cannot be eliminated through observation design alone.

Spatiotemporal Coordination at the Observation Level

Spatial scale mismatch is a key challenge in multi-source soil moisture fusion [115]. Satellite pixels at the kilometer scale and point measurements at the centimeter scale differ by several orders of magnitude, while soil moisture itself exhibits pronounced spatial heterogeneity [22]. Spatial coordination aims to enhance the representativeness of ground observations for satellite pixels through scientifically designed site deployment. Zhu et al. [116] evaluated 23 global soil moisture products using 992 ISMN sites and found that switching to highly representative sites in tropical regions increased the mean correlation coefficient by 0.41, indicating that current benchmark networks’ inadequate representativeness may underestimate errors or mask regional biases. Ground networks should therefore employ spatial variogram analysis and high-resolution auxiliary data to characterize heterogeneity in advance and determine monitoring sites’ optimal density and placement [117].

Soil moisture can undergo significant changes within hours to days following irrigation or precipitation, making cross-platform temporal consistency a prerequisite for effective fusion [118]. The most rigorous approach, quasi-synchronous acquisition, requires UAV flights and ground sampling to be tightly aligned with satellite overpass windows to eliminate temporal mismatch errors at the source [119]. However, its operational application at large scales is constrained by fixed satellite windows, airspace and weather restrictions, and the high cost of ground sampling [120]. Shokati et al. [121] implemented a full suite of quasi-synchronous acquisition protocols and emphasized that this design is critical for removing the bias introduced by rapid soil moisture dynamics. Alvarez-Vanhard et al. [122] noted that UAVs, given their flexible acquisition and immunity to cloud cover, can fill observational gaps left by optical satellites, which is particularly valuable in cloud-prone regions such as the tropics. Moreover, during irrigation, rainstorms, or flooding, field water status can change rapidly. Fixed revisit intervals are unable to capture such short-term dynamics, and observation frequency should be increased following these events to enable continuous tracking [123]. This type of dynamically intensified observation can shift temporal coordination from passive correction to proactive response [124]. However, research on adaptive observation triggered by environmental events remains limited, presenting an important direction for future design.

Methods for Spatial Scale Transformation

When observation design cannot fully eliminate spatial mismatches, scale transformation must be performed at the algorithmic level, which primarily encompasses two categories of methods: upscaling and downscaling.

Upscaling converts point measurements into area-averaged estimates and should incorporate spatial weighting and auxiliary information to reduce representativeness errors [125]. Geostatistical methods, such as kriging methods, are theoretically mature and provide estimation variance, but they require second-order stationarity and perform poorly under non-stationary trends [126]. Physically constrained methods introduce macro-level constraints, such as water or energy balance, to ensure consistency, but they rely on additional observations, including evapotranspiration and runoff, which limits their feasibility [127]. Machine learning and deep learning methods can capture complex nonlinear relationships, but their generalization performance depends heavily on the representativeness of training data and auxiliary variables [128]. Furthermore, upscaled results generally lack independent ground truth validation. Cross-validation or high-resolution reference maps are often used as substitutes, but both implicitly assume that the reference data are closer to the true values [129]. Establishing an independent validation framework through multi-source data synergy is therefore an important direction for future research. Notably, these three categories of methods are not mutually exclusive. In practice, they are often flexibly combined according to site density, the availability of auxiliary data, and the degree of surface heterogeneity.

Downscaling refines coarse-resolution products based on spatial deconvolution. It faces more severe information loss than upscaling [130]. Physics-based methods use land surface process models or radiative transfer models to simulate high-resolution moisture distributions, with data assimilation constraining simulation results. These methods are physically rigorous but computationally expensive and highly sensitive to model parameters and driving data accuracy [131]. Data-driven methods take high-resolution observations as training references and apply machine learning to spatially enhance low-resolution products [132,133]. They offer high computational efficiency and can reconstruct spatial detail, but they depend on the quality and representativeness of training samples [130]. Moreover, as a form of statistical extrapolation, they lack a physical description of vertical water transport processes [134].

In recent years, deep learning methods have been widely applied in data-driven downscaling. Convolutional Neural Networks (CNNs) excel at extracting spatial features from high-resolution auxiliary data [135]. Sequential models such as Long Short-Term Memory (LSTM) capture long-range temporal dependencies [136]. Generative Adversarial Networks (GANs) and diffusion models have shown particular strength in super-resolution reconstruction [137]. Transformers handle complex multimodal correlations through self-attention mechanisms [138], and architectures such as 3D U-Net and PHASE further enable joint spatiotemporal modeling [139]. Purely data-driven methods, however, may violate physical constraints and produce unreliable extrapolations. This has spurred the development of physics-guided machine learning downscaling, which embeds physical constraints into data-driven models to enhance physical consistency and generalization capability [140]. Zhang et al. [134] introduced physical constraints on the Tibetan Plateau, reducing the RMSE by 6.9–13.2% and improving the regression slope by 76.8%. To address surface heterogeneity, cluster-based zonal modeling using k-means can effectively support a single global model in adapting to local heterogeneity [141]. At the post-processing stage, residual correction via the Poisson equation, which harmonizes boundary gradients between filled and original regions, has become the standard for generating high-resolution seamless soil moisture products [142].

Methods for Temporal Consistency Processing

When quasi-synchronous acquisition cannot be achieved in the observation design, temporal resampling must be performed at the algorithmic level [143].

Physics-based forward/backward simulation centers on soil hydrodynamics and uses models such as the Richards equation to simulate soil moisture evolution between observation time points [144]. One can start from an earlier observation and propagate forward to the target time using meteorological forcing data or, alternatively, infer backward from a later high-precision observation [145]. These methods offer clear physical meaning and strong hydrological consistency, but their accuracy depends on soil hydraulic parameters and meteorological forcing data, limiting their applicability in data-sparse regions [146].

Time series interpolation exploits the autocorrelation characteristics of soil moisture time series [147]. Techniques such as spline interpolation, Kalman smoothing, and locally weighted regression are employed to resample irregular observations onto a unified temporal grid. These methods are computationally efficient and straightforward, performing stably during dry periods when moisture changes are gradual. However, they lack physical constraints and can introduce significant bias during rapid moisture changes following rainfall or irrigation, making them suitable only for small corrections over short intervals [148]. Yang and Wan [149] noted that traditional multi-temporal compositing or averaging methods neglect the time-varying nature of land surfaces and degrade the temporal resolution of the data, which has driven the evolution of time series reconstruction from simple interpolation to deeper temporal feature extraction.

In recent years, deep-learning-based temporal modeling has effectively overcome performance bottlenecks in scenarios with prolonged gaps or high missing rates [150]. The DrcGAN-STF framework proposed by Jiang et al. [142] reformulated gap filling as a temporal resolution enhancement problem and produced a global daily seamless STSG-SM product for 2015 to 2023, with validation accuracy highly consistent with original SMAP data. The PhyFill model developed by Wei et al. [151] embedded the physical mechanisms of soil moisture dynamics into deep learning, employing precipitation-driven monotonicity constraints and dry–down curve boundary constraints to improve physical consistency. Reconstruction models based on temporal autocorrelation assumptions, however, may fail under anomalous meteorological events, further underscoring the need to incorporate physical constraints [152]. Model-driven data assimilation offers an alternative pathway for addressing temporal challenges by integrating multi-source, multi-scale observations as dynamic constraints into land surface process models [153,154]. Algorithms such as the Ensemble Kalman Filter optimize model states [155], as exemplified by NASA’s SMAP L4 root-zone product [156], yielding spatiotemporally continuous, physically consistent products with profile information. Their performance, however, depends on model structure, parameterization schemes, and the accuracy of forcing data, while also being computationally intensive and costly [157].

Synergistic Strategies Across Sensor Combinations

For specific technical synergies, recent representative studies are summarized in

Table 3. Representative case studies of multi-source data fusion methods for field water status retrieval.

Fusion Platform Type	Fusion Method	Fusion Data Type	Scale Processing Method	Depth (cm)	Main Uncertainty Sources	Uncertainty Mitigation	Performance Metrics	Study
Space–Air	Data-driven downscaling	Landsat + UAV multispectral	Point-pixel mapping	0–20	Canopy/ temporal mismatch	Synchronous sampling	R2 = 0.69 RMSE = 3.88%	Yang et al. [168]
Space–Air	Feature-level fusion (ML)	Sentinel-2 + UAV hyperspectral	Feature statistical resampling	0–10	Spectral saturation	Feature selection	R2 ≈ 0.80	Khose et al. [169]
Space–Ground	Active–passive microwave synergy	SMAP Sentinel-1 SAR	Cross-resolution mapping	0–4	Vegetation/ scale mismatch	Vegetation correction	ubRMSE ≈ 0.035–0.045 cm3/cm3	Gruber et al. [170]
Space–Ground	Statistical downscaling (ML)	SMAP + optical proxies	Pixel-scale mapping	0–5	Vegetation effect	Bias correction	Accuracy preservation	Wei et al. [171]
Air–Ground	Hierarchical data-driven retrieval	UAV multispectral + TIR + in situ	Field-scale modeling	0–60	Indirectness of root zone estimation	Hierarchical modeling	R2 = 0.78 (0–20 cm)	Shi et al. [69]
Air–Ground	Physically constrained collaborative retrieval	Ground Penetrating Radar (GPR) + Sentinel-1	Point-pixel registration	0–10	Soil heterogeneity	Co-located calibration	R2 ≈ 0.74	Atun et al. [172]
Space–Air–Ground	Integral Equation Model (IEM) + WCM + neural network	Sentinel-1/2 + UAV Digital Surface Model (DSM) + In situ	UAV roughness retrieval	0–5	Roughness, vegetation	UAV-based roughness + WCM calibration	R2 = 0.71–0.78 RMSE = 0.023 m3/m3	Chakhar et al. [173]
Space–Air–Ground	Multi-source deep learning fusion	Sentinel-1/2 + UAV + in situ	Multi-scale feature representation	0–10	Sensor bias	Regularization	RMSE reduction	Batchu et al. [174]

. Optical and microwave synergy exploits optical data to quantify vegetation conditions, correcting vegetation-induced biases in microwave signals. Strategies range from physics-based correction (e.g., using the water cloud model) to purely data-driven joint retrieval [158]. Current research increasingly favors hybrid, physics-informed approaches that balance accuracy with interpretability [159]. Active and passive microwave synergy capitalizes on their complementary spatiotemporal characteristics. Joint retrieval and state optimization can be achieved through data assimilation [160], while decision-level optimal fusion based on statistical methods such as Bayesian averaging or optimal interpolation offers an alternative pathway [161]. As shown in Table 3, well-designed strategies can reduce regional-scale ubRMSE to 0.03–0.04 cm³/cm³. GNSS-R and traditional remote sensing fusion can be achieved by assimilating GNSS-R products with authoritative datasets like SMAP to optimize model states or applying machine learning for feature-level downscaling to enhance spatial resolution [162,163]. Evaluations indicate that these collaborative approaches preserve SMAP’s stability while enhancing temporal resolution and event capture capability [164]. Optical and thermal infrared synergy, exemplified by the temperature–vegetation index feature space method, indirectly assesses vegetation water stress and root zone moisture availability through the NDVI–LST relationship. This method is well-suited to rapid regional drought assessment but estimates stress levels rather than absolute water content [165,166,167].

In summary, selecting an appropriate fusion approach requires balancing multiple factors: data availability in the target region, vegetation complexity, the need for interpretability, and computational resources. Hybrid paradigms integrating physical mechanisms with data-driven intelligence are emerging as a key direction for achieving high-precision and generalizable soil moisture retrieval over complex underlying surfaces.

3.2.3. Uncertainty Quantification

Multi-source data fusion does not merely eliminate errors. Inevitably, it also introduces new scientific challenges. Uncertainty quantification is therefore a core component of reliability assessment [31]. Based on their sources and nature, uncertainties can be divided into two broad categories: aleatoric uncertainty, which arises from the inherent randomness of observational data and measurement errors and is generally irreducible, and epistemic uncertainty, which stems from deficiencies in model structure, parameter uncertainty, or incomplete knowledge and can, in principle, be reduced through model improvements [175].

Uncertainty quantification methods mainly include the following. (1) Error propagation models and triple collocation analysis (TCA) to quantify fusion results’ confidence intervals and identify error sources [176]. As shown in Table 3, several studies have employed TCA, including standard TCA and categorical TCA, to quantify the quality of multi-source remote sensing soil moisture products and trace error sources. (2) Uncertainty assessment embedded within data assimilation methods, such as the Ensemble Kalman Filter (EnKF) [177]. (3) Systematically considering potential error sources, such as those in radiative transfer processes, during model training [19]. (4) Indirectly reflecting uncertainty reduction through improvements in accuracy metrics, such as R² [15].

Uncertainty quantification in deep learning models is particularly important yet highly challenging. Traditional deep models generally do not provide estimates of prediction error [178]. To address this, researchers have explored several technical pathways, including Bayesian neural networks [179], Monte Carlo dropout [180], mixture density networks, and deep ensembles [181]. As Table 3 also reveals, many existing studies lack explicit discussion or systematic quantification of fusion result uncertainty. The black-box nature of these models impedes interpretability and complicates confidence interval estimation [182]. Kuang et al. [183] attempted to improve model transparency through a three-stage interpretability analysis framework, but a systematic quantification framework remains largely absent. Consequently, developing systematic approaches to uncertainty characterization and advancing hybrid paradigms that integrate physical understanding with data-driven methods represent critical directions for achieving reliable, high-precision soil moisture retrieval over complex surfaces [184].

In summary, proactive coordination at the observation design level reduces mismatches at the source, scale transformation and temporal resampling at the algorithmic level address residual mismatches, and uncertainty quantification throughout the entire workflow provides essential safeguards for product reliability. Together, these three components form a complete methodological system spanning data acquisition to product generation.

4. Accurate Acquisition and Application of Field Water Status for Agricultural Irrigation and Drainage

Space–air–ground integrated monitoring provides core support for scientific irrigation and drainage decision making through the accurate acquisition of field water status data. As illustrated in Figure 5, this section first examines large-scale surface soil moisture monitoring for regional water resource allocation, analyzing two data fusion strategies: high-frequency, medium-to-low resolution fusion and high-resolution periodic fusion. We then discuss statistical, physical, and data-driven methods for estimating crop root zone soil moisture, emphasizing approaches that extrapolate root zone conditions from surface observations. Finally, in the context of paddy fields, we systematically describe the technical workflow for planting area identification and water depth retrieval, presenting an integrated framework for accurate field water status acquisition.

4.1. Methods for Large-Scale Surface Soil Moisture Monitoring

The core objective of dynamic large-scale surface soil moisture monitoring is to provide spatiotemporally continuous data for regional drought assessment and macro-scale agricultural water allocation, which requires a balance between three core dimensions: coverage range, temporal resolution, and spatial resolution. Given the limitations of a single data source, two mainstream fusion scheme pathways have been formed in practice, with representative cases summarized in Table 4.

The high-frequency, medium-to-low resolution monitoring strategy takes passive microwave remote sensing products as its core, combined with active microwave data for collaborative optimization. It aims to ensure high temporal continuity and large-scale coverage and is suitable for national-level and basin-level dynamic drought monitoring [177]. The technical combination is based on passive microwave data such as SMAP (L-band passive microwave, 3-day revisit cycle, 9 km resolution) and SMOS (L-band passive microwave, 3-day revisit cycle, 36 km resolution) fused with active microwave data, including ASCAT (C-band active microwave, 1-day revisit cycle, 500 m resolution) and Sentinel-1 (C-band active microwave, 12-day revisit cycle, 10 m resolution), to achieve spatiotemporal complementarity through decision-level fusion [185]. The high-frequency, medium-to-low resolution pathway features temporal continuity and low susceptibility to weather interference but has limited spatial resolution and difficulty characterizing field-scale heterogeneity, with a ubRMSE generally ranging from 0.04 to 0.06 cm³/cm³ [186].

The high-resolution, periodic monitoring strategy takes the synergy between active microwave and optical remote sensing as its core, combined with a space–air–ground verification framework. It aims to improve the spatial resolution to the field scale (10–100 m) and support county-level and township-level refined irrigation planning [187]. The technical combination takes Sentinel-1 SAR (C-band, 10 m resolution, 12-day revisit cycle) as the microwave data source fused with Sentinel-2 optical data (10 m resolution, 5-day revisit cycle) and UAV multispectral data (1–5 m resolution, on-demand observation) to achieve high-resolution retrieval through feature-level fusion [21,188]. The high-resolution, periodic pathway has high spatial resolution and is able to characterize field-scale heterogeneity, but it is greatly affected by cloud cover, with its temporal continuity dependent on interpolation algorithms. Its ubRMSE can be reduced to 0.03–0.04 cm³/cm³ [119].

However, these strategies cannot fundamentally solve the passive microwave and discontinuous SAR observations’ excessively coarse native resolution. Future breakthroughs may rely on the observational revolution brought about by new-generation remote sensing constellations (e.g., SAR constellations with higher revisit frequency, TerraSAR-X and TanDEM-L) [189,190] as well as advances in edge computing [191] and real-time assimilation technology [192] to better balance macro-scale monitoring and micro-scale management.

Table 4. Representative case studies of large-scale surface moisture monitoring.

Fusion Platform Type	Data Combination	Final Temporal Resolution	Final Spatial Resolution	Crop Type	Validation Dataset	Performance Metric	Study
Space–Ground	SMAP +MODIS + topographic and soil factors	Daily	1 km	Multi-crop farmland	ISMN	ubRMSE ≈ 0.041 m3·m−3; R ≈ 0.72	Xu et al. [193]
Space–Ground	SMAP L4 + Sentinel-1 SAR + soil factors	Daily	1 km	Rice, wheat, maize	87 stations + oven drying	ubRMSE ≈ 0.040 m3·m−3; R > 0.6 for 60% of stations	Xu et al. [194]
Space–Ground	SMAP + ASCAT + Joint UK Land Environment Simulator (JULES) LSM	Daily	50 km	Grassland, Cropland, Mixed cover	Ground stations	ΔR ≈ +0.05	Seo et al. [195]
Air–Ground	UAV (multispectral + thermal) + GPR	Flight campaign	cm–m	Vineyard	GPR + oven drying	R2 = 0.879; RMSE = 0.066 m3·m−3	Guan et al. [196]
Air–Ground	UAV (RGB + multispectral + thermal) + in situ	Flight campaign	5–30 m	Maize	Layered oven drying + TDR	R2 = 0.61; RMSE ≈ 2%	Zhang et al. [197]
Air–Ground	GPR + UAV (RGB + thermal)	Flight campaign	Meter level	Maize/ wheat	GPR + oven drying/TDR	R2 = 0.83 (10 cm), 0.79 (30 cm); RMSE = 1.9% (10 cm), 3.2% (30 cm)	Vahidi et al. [198]
Space–Air–Ground	SAR, optical, and microwave + UAV + meteorological and soil factors	Daily–weekly	0.1–1 km	Mixed farmland crops	Regional station network	R2 = 0.822; RMSE = 0.038 m3·m−3; RRMSE = 16.46%	Li et al. [199]
Space–Air–Ground	UAV multispectral + Sentinel-1/2 + in situ	Multi- temporal	10 m	Winter wheat	180 in situ samples	0–20 cm: R2 = 0.901; 20–40 cm: R2 = 0.884	Yu et al. [119]
Space–Air–Ground	UAV multispectral + Sentinel-1 + in situ	Single flight	0.5 m	Soybean	Gravimetric samples	R2 = 0.82–0.87	Zhao et al. [200]

Table 4. Representative case studies of large-scale surface moisture monitoring.

Fusion Platform Type	Data Combination	Final Temporal Resolution	Final Spatial Resolution	Crop Type	Validation Dataset	Performance Metric	Study
Space–Ground	SMAP +MODIS + topographic and soil factors	Daily	1 km	Multi-crop farmland	ISMN	ubRMSE ≈ 0.041 m3·m−3; R ≈ 0.72	Xu et al. [193]
Space–Ground	SMAP L4 + Sentinel-1 SAR + soil factors	Daily	1 km	Rice, wheat, maize	87 stations + oven drying	ubRMSE ≈ 0.040 m3·m−3; R > 0.6 for 60% of stations	Xu et al. [194]
Space–Ground	SMAP + ASCAT + Joint UK Land Environment Simulator (JULES) LSM	Daily	50 km	Grassland, Cropland, Mixed cover	Ground stations	ΔR ≈ +0.05	Seo et al. [195]
Air–Ground	UAV (multispectral + thermal) + GPR	Flight campaign	cm–m	Vineyard	GPR + oven drying	R2 = 0.879; RMSE = 0.066 m3·m−3	Guan et al. [196]
Air–Ground	UAV (RGB + multispectral + thermal) + in situ	Flight campaign	5–30 m	Maize	Layered oven drying + TDR	R2 = 0.61; RMSE ≈ 2%	Zhang et al. [197]
Air–Ground	GPR + UAV (RGB + thermal)	Flight campaign	Meter level	Maize/ wheat	GPR + oven drying/TDR	R2 = 0.83 (10 cm), 0.79 (30 cm); RMSE = 1.9% (10 cm), 3.2% (30 cm)	Vahidi et al. [198]
Space–Air–Ground	SAR, optical, and microwave + UAV + meteorological and soil factors	Daily–weekly	0.1–1 km	Mixed farmland crops	Regional station network	R2 = 0.822; RMSE = 0.038 m3·m−3; RRMSE = 16.46%	Li et al. [199]
Space–Air–Ground	UAV multispectral + Sentinel-1/2 + in situ	Multi- temporal	10 m	Winter wheat	180 in situ samples	0–20 cm: R2 = 0.901; 20–40 cm: R2 = 0.884	Yu et al. [119]
Space–Air–Ground	UAV multispectral + Sentinel-1 + in situ	Single flight	0.5 m	Soybean	Gravimetric samples	R2 = 0.82–0.87	Zhao et al. [200]

4.2. Methods for Crop Root Zone Soil Moisture Acquisition

Recent studies on root zone soil moisture have advanced soil moisture acquisition methods, but notable limitations persist, with representative studies summarized in Table 5. At the in situ monitoring level, FDR sensors and multi-depth networks now enable layered monitoring from 10 cm to 60 cm and beyond, accommodating different crops’ root zone requirements [201]. In process-based modeling, Chen et al. [202] embedded the Richards equation within the C-WOFOST model, providing important insights into root zone water uptake processes.

Although remote sensing has enabled large-scale soil moisture estimation [203], microwave sensors’ penetration depth remains confined to the top few centimeters of the soil surface, which is approximately 0 to 5 cm for the L-band [204]. Enhancing vertical detection capability at the sensor level is therefore an essential pathway for acquiring root zone moisture information. Extending the operating wavelength is the most straightforward strategy. C-band sensors probe to approximately 2 cm, the L-band reaches about 5 cm, and the P-band is theoretically capable of achieving substantially greater detection depths [205]. Sadeghi et al. [206] integrated AirMOSS airborne P-band SAR with the Richards equation and obtained an RMSE of 0.06 to 0.10 m³/m³ over the 0–50 cm depth interval, confirming the unique value of P-band observations. A complementary strategy for single-wavelength detection is multi-frequency synergistic observation, which exploits the differential penetration depths and sensitivities of distinct microwave bands to extract vertical moisture profile information. Brakhasi et al. [204] performed synergistic retrieval using L-band and P-band brightness temperature data over the 0–28 cm range, achieving an RMSE below 0.04 m³/m³ and providing proof-of-concept validation for multi-frequency vertical moisture distribution interpretation. Building on this foundation, the integrated use of multi-angle remote sensing data holds promise for further mitigating the confounding effects of vegetation canopies and soil surface roughness. GNSS-R opens an alternative avenue from the perspective of signal source diversification. The FY-3 GNSS-R constellation has already improved sampling density by a factor of 3.3 relative to a single satellite, with accuracy comparable to that of SMAP products [164]. In addition, near-surface geophysical techniques have notable potential. Wu et al. [207] employed UAV-mounted ground-penetrating radar and achieved an inversion accuracy with R exceeding 0.7 over the 10–40 cm depth interval.

Even the most advanced sensor observations, however, cannot directly cover the complete root zone profile over the entire crop growing season. Extrapolating remotely sensed surface moisture to the root zone therefore remains an enduring challenge [58]. This extrapolation relies on three categories of variables: remotely sensed observables, meteorological and environmental characteristics, and soil physicochemical properties. The manner in which they are utilized is the common foundation of various retrieval methods [208].

Recent methodological developments fall into three broad categories. (1) Statistical approaches map surface time series to root zone estimates using filtering or accumulation techniques, including linear regression [209], exponential filtering [210], and cumulative functions [211]. (2) Physics-based methods integrate remote sensing observations with land surface or hydrological models, iteratively updating model states to infer deep-layer moisture conditions, among which data assimilation into land surface models is widely recognized as the most accurate pathway for root zone moisture retrieval [212]. (3) Data-driven machine learning methods directly regress root zone moisture using multi-source data [213]. Existing reviews have systematically mapped this landscape. Li et al. [209], for instance, categorized the methods into four types, including empirical, semi-empirical, physics-based, and machine learning approaches, while Kasim et al. [214] classified them according to whether prior knowledge of surface soil moisture is required.

These three categories have distinct boundaries and trade-offs. Statistical approaches impose the lowest input requirements and are suitable for regions where only remote sensing products are available and ground truth data are lacking, offering computational simplicity [215]. However, they suffer from poor parameter transferability and lack a physical description of vertical water transport. Physics-based assimilation methods, by virtue of their explicit constraints on vertical water movement, demonstrate notable advantages in physical interpretability and vertical resolution [216]. For instance, Seo et al. [195] improved the correlation coefficient by approximately 0.03 at a global resolution of 9 km, and Zhou et al. [217] achieved an ubRMSE of 0.045 m³/m³ at a regional scale of 25 km. Their accuracy, however, depends on the quality of meteorological forcing data and soil hydraulic parameters, and their computational cost is substantial [218]. Machine learning methods are distinguished by their flexibility and computational efficiency [219]. Sahaar et al. [220] achieved R values ranging from 0.76 to 0.86 at resolutions from 70 m to 1 km, while Yu et al. [119] improved R² by 9.53% to 10.52% and reduced RMSE by 11.11% to 31.25% using an eXtreme Gradient Boosting (XGBoost) model. However, these methods rely on the spatial representativeness of training samples and may produce physically implausible predictions under extreme conditions [31].

Two cross-cutting challenges persist. First, root zone moisture extrapolation depends on surface moisture accuracy, yet existing models oversimplify root system architecture and water uptake dynamics. Second, the prohibitive cost of acquiring depth-resolved ground truth data for validation introduces evaluation uncertainty. Future progress requires integrating multi-source observations with physically constrained models, advancing deeper-profiling sensor technologies, and establishing standardized validation networks across diverse agroecosystems.

Table 5. Representative case studies of crop root zone soil moisture acquisition.

Fusion Platform Type	Depth (cm)	Data Combination	Method	Final Temporal Resolution	Final Spatial Resolution	Validation Dataset	Performance Metric	Study
Space–Ground	0–100	SMAP + ASCAT + LSM	Surface-driven + EnKF	Daily	9 km	ISMN	ΔR ≈ 0.03	Seo et al. [195]
	0–50	SMAP + ASCAT + ground meteorological	Variable Infiltration Capacity model (VIC) model + EnKF	Daily	25 km	Soil moisture stations	ubRMSE = 0.045 m3/m3	Zhou et al. [217]
	0–100	SMAP + MODIS + ECOSTRESS + meteorological	Weak surface dependence + ML	Daily–3 h	70 m– 1 km	ISMN	R = 0.76–0.86	Sahaar et al. [220]
Air–Ground	0–28	Tower-based L + P band brightness temperature	L/P band + Njoku and Kong layered coherent radiative transfer forward model + Particle Swarm Optimization (PSO)	Hour–Daily	Tower footprint	Profile probes	RMSE < 0.04 m3/m3	Brakhasi et al. [204]
	0–50	AirMOSS P-band	P-band + Richards equation + global optimization	Flight campaign	30–50 m	AirMOSS	RMSE ≈ 0.06–0.10 m3/m3	Sadeghi et al. [206]
	10–40	UAV-GPR + TDR	GPR full-wave + Lambot equation + Look-Up Table (LUT) + Topp	Flight campaign	Meter level	TDR	R > 0.7	Wu et al. [207]
Space–Air–Ground	0–100	Weather Research and Forecasting model (WRF) + GNSS Zenith Total Delay (ZTD) + meteodrones + in situ	3DVAR + SPHY model	Daily	2.5 km	In situ soil moisture sensors	Bias = 35%	Lagasio et al. [221]
	0–40	UAV multispectral + Sentinel-1/2 + in situ	Partial Least Squares Regression (PLSR) + XGBoost + SHapley Additive exPlanations (SHAP)	5 days	10 m	Ground sampling	R2 increased by 9.53–10.52%, RMSE reduced by 11.11–31.25%	Yu et al. [119]

Table 5. Representative case studies of crop root zone soil moisture acquisition.

Fusion Platform Type	Depth (cm)	Data Combination	Method	Final Temporal Resolution	Final Spatial Resolution	Validation Dataset	Performance Metric	Study
Space–Ground	0–100	SMAP + ASCAT + LSM	Surface-driven + EnKF	Daily	9 km	ISMN	ΔR ≈ 0.03	Seo et al. [195]
	0–50	SMAP + ASCAT + ground meteorological	Variable Infiltration Capacity model (VIC) model + EnKF	Daily	25 km	Soil moisture stations	ubRMSE = 0.045 m3/m3	Zhou et al. [217]
	0–100	SMAP + MODIS + ECOSTRESS + meteorological	Weak surface dependence + ML	Daily–3 h	70 m– 1 km	ISMN	R = 0.76–0.86	Sahaar et al. [220]
Air–Ground	0–28	Tower-based L + P band brightness temperature	L/P band + Njoku and Kong layered coherent radiative transfer forward model + Particle Swarm Optimization (PSO)	Hour–Daily	Tower footprint	Profile probes	RMSE < 0.04 m3/m3	Brakhasi et al. [204]
	0–50	AirMOSS P-band	P-band + Richards equation + global optimization	Flight campaign	30–50 m	AirMOSS	RMSE ≈ 0.06–0.10 m3/m3	Sadeghi et al. [206]
	10–40	UAV-GPR + TDR	GPR full-wave + Lambot equation + Look-Up Table (LUT) + Topp	Flight campaign	Meter level	TDR	R > 0.7	Wu et al. [207]
Space–Air–Ground	0–100	Weather Research and Forecasting model (WRF) + GNSS Zenith Total Delay (ZTD) + meteodrones + in situ	3DVAR + SPHY model	Daily	2.5 km	In situ soil moisture sensors	Bias = 35%	Lagasio et al. [221]
	0–40	UAV multispectral + Sentinel-1/2 + in situ	Partial Least Squares Regression (PLSR) + XGBoost + SHapley Additive exPlanations (SHAP)	5 days	10 m	Ground sampling	R2 increased by 9.53–10.52%, RMSE reduced by 11.11–31.25%	Yu et al. [119]

4.3. Methods for Paddy Field Identification, Flooded Area Mapping, and Water Depth Estimation

Paddy field water status remote sensing involves two interconnected tasks: mapping the spatial extent of fields and surface water and estimating the water depth within them. The presence of surface water renders remote sensing signals a complex mixture originating from three components: the water body itself, the submerged soil (or crop residue) background, and the rice canopy, which undergoes dynamic changes throughout phenological stages [222]. During periods of high vegetation cover, microwave and optical signals are strongly modulated by canopy scattering and absorption, leading to substantial attenuation or masking of depth-sensitive signatures. This fundamental complexity distinguishes paddy field monitoring from soil moisture retrieval in upland systems [223].

4.3.1. Paddy Field Delineation and Flooded Area Mapping

Paddy field and flooded area mapping provides essential spatial constraints for monitoring water depth dynamics. Approaches have evolved from traditional thresholding and empirical models through temporal feature analysis to machine learning and deep learning.

Traditional threshold and empirical methods exploit spectral or index-based characteristics to distinguish target features from other land cover types using fixed or adaptive thresholds or empirical regression equations for classification [224]. Optical-based approaches often combine phenological trajectories of indices such as the Normalized Difference Vegetation Index (NDVI) and Land Surface Water Index (LSWI) to establish multi-index decision rules [225]. SAR-based methods leverage the specular reflection properties of open water and apply thresholds to VV or VH polarization data [226]. However, paddy fields exhibit dynamic backscatter behavior throughout the growing season, necessitating multi-temporal data and adaptive thresholds to capture their unique temporal signatures [227].

Temporal feature recognition methods identify rice by analyzing time series trajectories from high-temporal-resolution satellite data combined with threshold-based rules or similarity measures [228]. Randriamihaja et al. [229] applied the Dynamic Time Warping classification algorithm to categorize paddy fields into distinct flood dynamic classes, capturing the seasonal heterogeneity of inundation extent and duration. Paddy fields exhibit characteristic patterns in SAR time series that enable their discrimination from other land cover types [230]. Zhang et al. [231] proposed a Reliable Scatterer classification framework using Sentinel-1 dual-polarization time series, achieving accurate separation of flooded and non-flooded areas. Optical and SAR data fusion allows for more comprehensive rice phenology characterization [232], as demonstrated by Fatchurrachman et al. [233], who combined monthly Sentinel-1 VH polarization with Sentinel-2 NDVI time series for rice identification and growth stage discrimination. For areas with varying inundation frequencies, differentiated temporal anomaly detection strategies can be designed to identify flood events or seasonal inundation [234]. More recently, unsupervised Gaussian Mixture Models have been applied to multi-temporal satellite data for monitoring flood impacts on crops, automatically extracting probability density functions for flooded versus non-flooded areas via the Flood Separability Index [235].

The strength of temporal phenology-based approaches lies in reformulating spatial identification as a temporal pattern recognition problem, circumventing the complexities of vegetation–water mixing in single-date imagery [236]. However, their accuracy remains contingent on the quality and density of time series data [237], and performance may degrade in regions with anomalous phenology or complex cropping systems [238,239].

4.3.2. Paddy Field Water Depth Estimation: From Clear Water to Vegetated Conditions

A fundamental limitation exists in current remote sensing research on paddy field water depth monitoring: existing methods primarily focus on water presence detection (flooding) while accurate water depth quantification remains an unresolved challenge, particularly in areas with large-scale, dynamic conditions, vegetation cover, or turbid water [2,61]. This limitation constrains both disaster loss assessment and the advancement of precision agriculture and refined water resource management [240]. Current approaches fall into two main categories: methods applicable to clear shallow water and techniques for vegetated areas.

For clear shallow water areas, early retrieval models relied on empirical or semi-analytical relationships linking spectral information to measured water depth [241,242]. The Stumpf model is a typical empirical method requiring calibration with in situ depth data, together with radiometric calibration and atmospheric correction [243]. In specific shallow sea areas, it can achieve errors within ±0.5 m [244]. Other conventional methods include log-ratio [245] and multi-band models [246], which are straightforward but sensitive to variations in water quality and bottom composition, with limited environmental transferability [247]. These constraints have spurred the development of adaptive and hybrid frameworks. The adaptive blended algorithm approach (ABAA) model integrates physical principles with empirical analysis and reveals a systematic pattern: as water depth increases, the bands most sensitive to depth transition from longer to shorter wavelengths [248]. Liu et al. [241] coupled ABAA with the optimization-based UMOPE model, establishing a robust retrieval framework that requires no field measurements. The downscaled bathymetric mapping approach (DBMA) model couples optimization-based and empirical components, achieving RMSE values below 2 m for clear water depths of 0–12 m and below 5 m for turbid water depths of 0–5 m [249]. However, beyond certain depth thresholds, reduced sensitivity of water-leaving reflectance leads to systematic underestimation [250].

To address the core challenge of vegetation cover, the synergy between microwave and optical data has become a cutting-edge research direction [251]. In representative studies, Zhang et al. [223] combined Sentinel-3 radar altimetry with optical imagery to monitor water levels in vegetated wetlands, achieving an RMSE of 0.29 m against in situ data. Simeón et al. [2] found that Sentinel-2 near-infrared reflectance during the tillering stage exhibited a significant negative correlation with water depth over the 0–14 cm range, yielding a retrieval RMSE of 1.2–1.8 cm. More recently, L-band fully polarimetric SAR has demonstrated new potential in this domain. Segami et al. [252] reported that the classification accuracy of inundated and non-inundated paddy fields in Japan reached 88% when plant height remained below 70 cm, but it declined markedly beyond this threshold. Islam et al. [253] observed a similar pattern in Bangladesh, where Alternate Wetting and Drying (AWD) identification accuracy dropped from 94% at early growth stages to 80% under dense canopy conditions. These findings indicate that even with the comparatively strong penetration capability of L-band signals, surface water status signals remain partially masked by vegetation scattering once the canopy has fully developed.

Although these studies have yielded encouraging results under their respective experimental conditions, they reveal the limitations of water depth retrieval in vegetated areas. First, existing methods operate within a narrow applicability window, relying heavily on specific phenological stages or sensor configurations, and no generalized scheme applicable to the entire growing season has been established [254]. Second, most current methods are empirical or semi-empirical models and parameter calibration depends on in situ water depth measurements, hindering their application in regions lacking ground observation data [247]. Third, existing studies have yet to address, from a physical mechanism perspective, whether the weak signal originating from shallow water layers, typically less than 20 cm in depth, can be stably and reliably separated from the strong scattering interference of the vegetation canopy [255]. This defines the essential distinction between paddy field water depth retrieval and water depth retrieval in open water bodies.

Collectively, these findings indicate that indirect water depth retrieval under vegetation cover is achievable only within narrow windows of applicability. The core physical problem of separating shallow water signals from canopy scattering remains unresolved.

4.3.3. Challenges and Future Directions for Paddy Field Water Depth Quantification

Current research primarily focuses on water presence identification or phenological stage analysis, with few studies directly addressing the dynamic quantification of paddy field water depth. Retrieval accuracy is particularly limited when water bodies are affected by complex factors such as sediment and vegetation (e.g., rice canopies) [254]. A second limitation concerns model generalization and adaptability to heterogeneous environments. Fragmented field parcels, complex topography, diverse cultivation practices, and persistent cloud cover and vegetation occlusion impede broad model applicability [256]. Third, limited spatiotemporal data coverage poses significant challenges. Satellite revisit cycles often misalign with the rapid progression of hydrological processes, hindering continuous, high-frequency monitoring of dynamic water level changes [257]. Finally, the interpretability of deep learning models requires strengthening to ensure transparent and defensible decision making in operational contexts [258].

These difficulties in paddy field water depth retrieval can be traced, at the physical mechanism level, to the inherent complexity of remote sensing signal composition. During the rice growing season, the backscatter signals received by SAR sensors represent the superposition of multiple scattering mechanisms, including volume scattering, surface scattering, and double-bounce scattering, and the relative contributions of these components vary substantially across phenological stages [253]. At the seedling stage, specular reflection from the water surface dominates; from the tillering to the jointing stage, double-bounce scattering between rice stems and the water surface intensifies markedly; after the heading and flowering stage, volume scattering from the dense canopy becomes overwhelmingly dominant, with scattering components from the water body severely attenuated [259]. Quantitative analyses based on Monte Carlo rice scattering models have further demonstrated that backscatter coefficients are highly sensitive to canopy structural parameters, such as leaf length, width, and thickness, and the magnitude of these variations far exceeds signal changes induced by shallow water depth fluctuations [260]. This implies that water depth information and vegetation structural information are deeply coupled in remote sensing observations. Decoupling water depth information from such mixed signals requires accurate isolation of the scattering contribution of the vegetation layer [255].

Whether one attempts to extract the double-bounce component through polarimetric decomposition or detect water level dynamics using differential InSAR, both approaches are constrained by the same core problem: the signal component associated with water depth accounts for too small a fraction of the total backscatter and is readily overwhelmed by minor variations in canopy structural parameters [261]. Specifically, although the Freeman–Durden decomposition can separate scattering components, the sensitivity of the double-bounce component to stem density and hill diameter far exceeds its sensitivity to shallow water depth variations, rendering decoupling highly unstable in practice [262]. InSAR can detect vertical displacements with millimeter-level precision, yet routine irrigation and rainfall during the rice growing season rapidly alter surface dielectric properties and the micro-geometric structure of the canopy, causing complete decorrelation between image pairs and consequent loss of interferometric phase information [263]. Moreover, even when longer-wavelength radar signals can penetrate the canopy and reach the water’s surface, this process is accompanied by two-way attenuation caused by canopy absorption and scattering [264]. The attenuation coefficient is itself a function of the leaf area index, canopy water content, and plant density, which vary dramatically over the course of the growing season. Any errors in the estimation of canopy parameters may thus be amplified through the attenuation correction process and ultimately propagated into the water depth retrieval results [265].

Emerging methodological advances offer pathways to address these challenges. First, the availability of data sources with finer resolution, higher revisit frequency, and expanded spectral coverage will substantially enhance the detail and timeliness of paddy water depth monitoring [266,267,268]. Second, deeper integration of deep learning with big data analytics is enabling the transformation of raw remote sensing measurements into actionable decision support information. Coupled with Explainable Artificial Intelligence (XAI) and physically based modeling, this trend promises to progressively mitigate the “black-box” limitations of purely data-driven approaches [269]. Third, as demonstrated by Wang et al. [270], the integration of remote sensing observations with domain knowledge from hydrological modeling and agronomy can significantly improve the model’s predictive reliability and physical consistency. It is therefore worthwhile to explore the use of multi-source, multi-temporal remote sensing observations as constraints that can be dynamically assimilated into rice hydrological process models, enabling systematic estimation of water depth, which is difficult to observe directly. Future research should draw inspiration from other mature fields and focus on addressing current challenges to propel remote-sensing-based agricultural monitoring and management towards new technological innovations.

5. Challenges and Future Directions in Field Water Status Monitoring

Despite significant advances in space–air–ground integrated monitoring systems, several challenges continue to constrain the operational application of field water status data for agricultural decision making.

First, a substantial body of research remains focused on qualitative presence detection while accurate quantification has yet to achieve the robustness required for broad application. This limitation is particularly evident in two critical areas: paddy field water depth under vegetation cover and depth-resolved root zone soil moisture. A fundamental constraint is that most remote sensing signals originate from the surface, providing only limited vertical information.

Second, many high-precision retrieval algorithms, especially data-driven approaches, are tailored to specific regions, crops, or phenological stages. Their performance often degrades in different agro-climatic conditions due to their reliance on localized statistical relationships rather than transferable physical mechanisms. This lack of generalizability limits their utility across diverse agricultural landscapes.

Third, multi-source data fusion remains hampered by scale mismatch and error propagation. Critically, many advanced retrieval products lack rigorous, spatially explicit uncertainty characterization. This deficiency is particularly problematic for risk-sensitive agricultural decisions where confidence is as important as point estimates.

Fourth, the ultimate goal of informing irrigation management remains inadequately addressed. Most studies focus on enhancing retrieval accuracy but fall short of establishing clear pathways from spatiotemporally resolved water status to field-scale actionable decisions, such as determining when to irrigate and how much water to apply. Moreover, rigorous evaluation of associated water-saving or yield benefits is rarely undertaken, and the decision–support loop remains incomplete.

To overcome these challenges, future research should prioritize the following directions.

First, differentiable programming frameworks should embed physical mechanisms (e.g., the Richards equation for soil water dynamics) into neural networks, achieving physics-constrained deep learning with enhanced interpretability and generalization capabilities.

Second, fusion methods should be advanced from static algorithms to dynamic, adaptive systems capable of evaluating and weighting contributions from different data sources in real time based on environmental conditions such as vegetation density and cloud cover, generating more robust estimates.

Third, space–air–ground-derived water status data should be integrated with crop growth models and optimization algorithms to establish field-scale irrigation and drainage decision support systems, in addition to quantitative predictions of water savings, yield impacts, and environmental co-benefits.

Fourth, open-access global benchmark validation datasets for field water status should be established, with uncertainty quantification a requirement for product generation.

Fifth, it is imperative to capitalize on opportunities presented by new satellite missions, including synergistic L-band and P-band SAR and operational GNSS-R constellations, to overcome data source limitations in vertical profile soil moisture retrieval and advance technical frameworks for remote-sensing-based field water status monitoring.

6. Conclusions

This review systematically examined the space–air–ground integrated framework for field water status monitoring. Satellite remote sensing and ground-based monitoring constitute the dual-core foundation of current systems, with UAV technology emerging as a critical bridge across spatial scales. No single platform can simultaneously satisfy the comprehensive demands of agricultural water management, making synergistic fusion essential. Multi-source data fusion has evolved from simple statistical approaches to sophisticated frameworks integrating physical models with machine learning. However, scale mismatch and error propagation remain unresolved, and systematic uncertainty quantification is rarely implemented. Among agricultural applications, paddy field water depth retrieval represents the most significant research gap. Despite its critical importance for global food security, a quantitative estimation has not yet been achieved. The foundational stage of technological demonstration is largely complete, and the path forward requires a concerted push toward intelligent, reliable, and prescriptive systems rooted in physical understanding and robust across diverse agro-ecosystems.