Estimation of Surface Moisture Content using Sentinel-1 C-band SAR Data Through Machine Learning Models

Statistical Exploration of SENTINEL-1 Data, Terrain Parameters, and in-situ Data for Estimating the Near-Surface Soil Moisture in a Mediterranean Agroecosystem

“…At this point, it should be briefly mentioned that both the measuring instrument (Stevens HydraProbe) and the timing of the field observations in tune with the overpasses of S1 are directly comparable to several recently published studies (Ayehu et al, 2020;Datta et al, 2020;Han et al, 2020;Ma et al, 2020), and thus are regarded as reliable from a technical point of view.…”

“…High spatiotemporal SAR-based retrieval of θ predominantly focusses on bare land surfaces (e.g., Datta et al, 2020), sparsely vegetated areas such as grassland and meadows (e.g., Xu et al, 2020), or agricultural farmland when the vegetation cover is low such as covered by crop residues (e.g., Ayehu et al, 2020). This is mainly due to the unique challenge resulting from the SAR backscatter's high sensitivity to surface characteristics (e.g., roughness) and vegetation properties (e.g., vegetation canopy architecture).…”

“…Empirical backscatter models present a simplification by investigating the interaction of microwaves with site-specific surface characteristics but are limited to the area under investigation and require the input of high-quality reference data, e.g., surface roughness (Barrett et al, 2009). Data-driven machine learning (ML) approaches were investigated for the operational and systematic mapping and monitoring of θ (e.g., Santi et al, 2016;Adeyemi et al, 2018;Cai et al, 2019;Efremova et al, 2019;Datta et al, 2020). Provided the availability of data sufficiently characterizing the system for which θ is to be estimated, they present an adequate alternative (Ezzahar et al, 2019) as they expand the concept of the above backscatter models and allow integration of additional environmental covariates (parameters) beyond those defined for physical modeling.…”

“…They can account for the non-linearity of relationships, which is mainly present in RS-based θ retrieval in vegetated and heterogeneous areas (e.g., Chakrabarti et al, 2015). Random forest regression and support vector machine applied on pixels (Hajdu et al, 2018;Datta et al, 2020) or image objects (Ezzahar et al, 2019) and K-Nearest Neighbors for nonparametric regression analysis (Datta et al, 2020) are reported to lead to predictions of increasing accuracy, with even reported less required data compared to mechanistic models (Adeyemi et al, 2018), and enhanced readability of results (Attarzadeh et al, 2018;Ezzahar et al, 2019). ML approaches are increasingly used to estimate θ in agricultural watersheds across humid to arid climate regimes and at different scales (e.g., Quesney et al, 2000;Hachani et al, 2019).…”

“…Commonly, studies employ the backscatter intensity Sigmanaught (σ • ) either in single (VV or VH) polarization (e.g., Quesney et al, 2000;Attarzadeh et al, 2018;Xu et al, 2020) or dual (VV + VH) polarization (e.g., Dabrowska-Zielinska et al, 2018;Efremova et al, 2019;Datta et al, 2020) from SAR amplitude information only. The topographic effect on radar backscatter, i.e., foreshortening, layover, and shadowing (Chen and Wang, 2008), is considered in the preprocessing of SAR data such as by terrain flattening and orthorectification to remove radiometric distortions (Small, 2011) and to correct distortions induced by the sensor inclination and topography, respectively (Filipponi, 2019).…”