Field-scale soil moisture estimated from Sentinel-1 SAR data using a knowledge-guided deep learning approach

Yi Yu a,b,∗ , Patrick Filippi a , Thomas F. A. Bishop a · 2025

Read Field-scale soil moisture estimated from Sentinel-1 SAR data using a knowledge-guided deep learning approach by Yi Yu a,b,∗ , Patrick Filippi a , Thomas F. A. Bishop a (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00265.txt.

Introduction

Yi Yu, Patrick Filippi, and Thomas Bishop estimate field-level soil moisture from satellite radar by guiding a deep learning model with physical knowledge, in a 2025 paper. Radar can sense soil wetness, but vegetation and surface roughness scramble the signal. A classic semi-physical model helps, though its rough edges limit accuracy across varied farmland.

Their model feeds a memory-based network with radar readings stripped of vegetation effects, plus satellite greenery data. A custom loss keeps the network honest to the physical model and to sane bounds. Checked against ground sensors at held-out sites, the method cut error and tracked moisture well across cover types. The authors call the results preliminary but promising.

Catalog record

CREATORYi Yu a,b,∗ , Patrick Filippi a , Thomas F. A. Bishop a
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-00265

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Field-scale soil moisture estimated from Sentinel-1 SAR data using a knowledge-guided deep learning approach Yi Yu a,b,∗ , Patrick Filippi a , Thomas F. A. Bishop a Abstract Soil moisture (SM) estimation from active microwave data remains challenging due to the complex interactions between radar backscatter and surface characteristics. While the water cloud model (WCM) provides a semi-physical approach for understanding these interactions, its empirical component often limits performance across diverse agricultural landscapes. This research presents preliminary efforts for developing a knowledge-guided deep learning approach, which integrates WCM principles into a long short-term memory (LSTM) model, to estimate field SM using Sentinel-1 Synthetic Aperture Radar (SAR) data. Our proposed approach leverages LSTM’s capacity to capture spatiotemporal dependencies while maintaining physical consistency through a modified dual-component loss function, including a WCM-based semi-physical component and a boundary condition regularisation. The proposed approach is built upon the soil backscatter coefficients isolated from the total backscatter, together with Landsat-resolution vegetation information and surface characteristics. A four-fold spatial cross-validation was performed against in-situ SM data to assess the model performance. Results showed the proposed approach reduced SM retrieval uncertainties by 0.02 m 3 /m 3 and achieved correlation coefficients (R) of up to 0.64 in areas with varying vegetation cover and surface conditions, demonstrating the potential to address the over-simplification in WCM. † † address: a The University of Sydney, Sydney, NSW 2006, Australia b CSIRO Agriculture and Food, Canberra, ACT 2601, Australia 1 Introduction Soil moisture (SM) is a critical state variable in land surface hydrology that regulates the exchange of water and heat energy between the land surface and the atmosphere [ 1 ] . An accurate and high-resolution estimation of SM is crucial as an initial step in agricultural-related and broader environmental studies. Despite its importance, obtaining high-resolution and spatially complete estimates of surface SM remains challenging, particularly at the field scale necessary for agricultural applications [ 2 ] . Satellite remote sensing has emerged as a powerful tool for monitoring SM content, offering various approaches to achieve this goal. The most common operational method for retrieving SM utilises the passive microwave electromagnetic spectrum [ 3 ] , which can offer several advantages, including day and night operability, minimal atmospheric interference, and reliability and wide applicability in SM retrievals. However, the trade-off is their coarse spatial resolution, usually in the range of tens of kilometres, which limits their applicability in agricultural field-scale studies. Despite the explorations aimed at deriving high-resolution SM estimates [ 4 , 5 ] , knowledge gaps persist regarding their reliability and uncertainty characterisation across environmental conditions [ 6 , 7 ] . Alternatively, C-band Synthetic Aperture Radar (SAR) sensors present potential for SM retrieval at higher resolution [ 8 ] . The SAR systems operate by actively transmitting electromagnetic waves and measuring the returned backscatter signal, offering distinct advantages over passive microwave sensors [ 9 ] . The strength of SAR lies in its capacity to achieve high spatial resolution, typically ranging from 10 to 30 metres, while maintaining consistent data acquisition capability regardless of atmospheric conditions or solar illumination [ 10 ] . However, SAR-based SM retrieval also presents inherent signal complexities compared to passive microwave sensors [ 11 ] , due to the multiplicative nature of radar backscatter interactions with surface characteristics, including vegetation cover, surface roughness and soil texture [ 12 ] . The SAR-based SM retrieval predominantly relied on either empirical or data-driven approaches. The empirical and semi-empirical approaches, such as the change detection method [ 11 ] and the water cloud model (WCM) [ 13 ] , attempt to characterise the complex interactions between radar backscatter and vegetation and surface properties [ 14 ] . These models typically decompose the total backscatter signal into contributions from soil and vegetation, establishing analytical relationships based on theoretical understanding and experimental observations. However, their performance may be constrained by the simplified assumptions and site-specific parametrisation that may not generalise well across diverse agricultural landscapes [ 15 ] . In contrast, recent developments in data-driven approaches, particularly deep neural networks, demonstrate potential for field-scale SM estimation by capturing complex, non-linear relationships between SAR observations and predictors. Nonetheless, the purely data-driven models were usually criticised due to the lack of interpretability [ 16 ] . The integration of semi-physical and data-driven approaches, hence, presents an opportunity to embed the physical insights into the learning capacity of deep neural networks [ 16 ] . By reformulating the semi-empirical relationships within a deep learning framework while preserving physical constraints, such hybrid approaches can potentially overcome the limitations of both paradigms, including both the oversimplification inherent in empirical models and the physical inconsistency risks associated with purely data-driven methods [ 17 ] . Therefore, this research presents preliminary efforts to estimate field-scale SM from Sentinel-1 SAR data using a knowledge-guided deep learning approach, to explore the possibility of embedding the knowledge from a WCM into a neural network architecture. 2 Study area and data 2.1 Study area and in-situ data The Yanco agricultural region was chosen as study area for this research, which is situated within the Murrumbidgee Catchment and exemplifies a semi-arid agricultural environment of southeastern Australia. Fig. 1 shows an 80 km × 80 km area of (a) the Sentinel-1 backscatter coefficient data acquired on 01 Feb 2019; (b) the synthesised NDVI data derived from the fusion of MODIS and Landsat surface reflectance coincident with the SAR acquisition time; and (c) the land cover information [ 18 ] . The in-situ near-surface SM measurements (0-5 cm) from the OzNet Hydrological Monitoring Network [ 19 ] were collected for cross-validation, with sites denoted by black squares in Fig. 1 (c). The distribution of OzNet sites was relatively clustered (Fig. 1 c). They were then partitioned into four distinct groups to facilitate a four-fold cross-validation, where the detailed information for the folds and clusters can be found in [ 7 ] (see their Table II). Figure 1: (a) The spatial distribution of Sentinel-1 backscatter coefficient on 01 Feb 2019; (b) the synthesised NDVI derived from the fusion of MODIS and Landsat data on 01 Feb 2019; and (c) the land cover classification based on the ESA WorldCover 10 m v200 dataset [ 18 ] . 2.2 Sentinel-1 SAR data The Sentinel-1 mission, operated under the ESA’s Copernicus program, comprises twin C-band (5.405 GHz) SAR satellites: Sentinel-1A and -1B [ 10 ] . The mission offers high-resolution SAR imagery with a spatial resolution of 5 × 20 meters in the Interferometric Wide (IW) swath mode, which represents the primary acquisition mode over land. The constellation achieves a temporal resolution of 6 days at the equator when both satellites are operational, though actual revisit frequency varies by geographical location (e.g., a typical 12-day revisit cycle for eastern Australia). The Ground Range Detected (GRD) product was collected for this research, which undergoes systematic preprocessing using the Sentinel-1 Toolbox to generate calibrated, ortho-corrected imagery. The preprocessing workflow encompasses thermal noise removal, radiometric calibration, and terrain correction. The final backscatter coefficients are converted to decibel scale through a logarithmic transformation (10\times log10(x)) . We specifically focused on the vertical transmit / horizontal receive (VH) polarisation, which has demonstrated sensitivity to SM variations. Each scene includes an

Read the complete text