Estimation of field-scale soil moisture content and its uncertainties using Sentinel-1 satellite imagery
The PhD defence of Harm-Jan Benninga will take place (partly) online.
The PhD defence can be followed by a live stream.
Harm-Jan Benninga is a PhD student in the department of Water Resources. Supervisor is prof.dr. Z. Su from the Faculty of Geo-Information Science and Earth Observation (ITC).
The soil moisture content (SMC) expresses the amount of water in the unsaturated zone. The variable is essential for vegetation growth and hydrological processes. SMC can be estimated from satellite microwave observations across spatial domains. This thesis focuses on the field-scale, at which more direct relations between the ground conditions (SMC, surface roughness and vegetation) and satellite microwave observations are expected. The uncertainties in the field-scale SMC retrievals were studied and decomposed in uncertainties originating from in situ references (Usp and Us,S1), satellite observations (US1) and model parameters (Up).
In situ measurements
Two regional networks were employed for the in situ monitoring of SMC, namely the Twente network in the east and the Raam network in the south of the Netherlands. Both networks have stations with sensors at depths of 5 cm, 10 cm, 20 cm, 40 cm and 80 cm. From soil-specific calibrations follow station probe measurement uncertainties (Usp) of 0.018 m3 m-3 to 0.023 m3 m-3 for the Raam network and 0.028 m3 m-3 for the Twente network. The sensor's influence zone, determined in Raam soil, is 3 cm – 4 cm. A vertical mismatch and a horizontal mismatch between SMC retrievals from satellites and the station measurements cause a spatial mismatch uncertainty (Us,S1). Using measurements inside four agricultural fields, the Us,S1 estimate is 0.051 m3 m-3.
Sentinel-1 satellite observations
The Sentinel-1 satellites provide microwave backscatter (σ0) observations, which can be used for field-scale SMC retrieval. The σ0 observations are found to be disturbed by frozen conditions below an air temperature of 1 °C, snow during Sentinel-1's morning overpasses on meadows and cultivated fields, and interception after more than 1.8 mm of rain in the 12 hours preceding a Sentinel-1 overpass. Dew was not found to be of influence. After masking based on these rules, the Sentinel-1 σ0 observations still contain radiometric uncertainty (sS1) originating from calibration uncertainties, sensor instabilities and speckle. σ0 over forests is assumed time-invariant; the observed deviations were used to estimate the sS1. The sS1 improves from 0.85 dB (for a surface area of 0.25 ha) to 0.30 dB (10 ha) for the VV polarization and from 0.89 dB (0.25 ha) to 0.36 dB (10 ha) for the VH polarization, following approximately an inverse square root dependency on the surface area over which the σ0 observations are averaged. The retrieval uncertainty due to sS1 (US1) is low (-0.01 m3 m-3 to +0.01 m3 m-3) dry soils and large surface areas and high (-0.10 m3 m-3 to +0.17 m3 m-3) for wet soils and small surface areas.
Uncertainty under sparsely vegetated conditions
The uncertainties involved in surface scattering simulations and SMC retrievals were investigated. The surface roughness parameters that are input to the integral equation method (IEM) surface scattering model were calibrated for two sparsely vegetated meadows and two fallow maize fields. A Bayesian framework was used for the calibration as well as for deriving the model parameter uncertainty (Up) and total uncertainty (Utotal-B). The resulting Utotal-B successfully reproduces the uncertainty estimated empirically against the in situ references. The combination of the derived Up with Usp, Us,S1 and US1 also constitute the total SMC retrieval uncertainty. The main uncertainty originates from the in situ references (Usp and Us,S1) and the Sentinel-1 observations (US1), whereas the contribution from the surface roughness parameters (Up) is small.
Accounting for vegetation effects over meadows
For the two meadows the surface roughness parameter distributions are similar, time-invariant and independent of Sentinel-1's ascending/descending orbits. These are promising results for the operational retrieval of SMC over meadows across a larger region because they suggest that using a single set of surface roughness parameters is permitted. The IEM surface scattering model and the Tor Vergata (TV) vegetation scattering and absorption model were parameterised for grass-covered soil surfaces. A Sentinel-2 leaf area index (LAI) product provides field-scale vegetation information, as was demonstrated by validation against in situ measurements on two meadows and four maize fields. However, uncertainty propagation shows that the Sentinel-2 LAI uncertainty of 0.71 m2 m-2 has a large impact on SMC retrievals. The SMC retrievals for 21 meadows in the Twente region, validated against adjacent in situ station references, have mean Pearson correlation coefficients of 0.55 for IEM and 0.64 for TV-IEM, root mean square deviations (RMSD) of 0.14 m3 m-3 for IEM and 0.13 m3 m-3 for TV-IEM, and RMSDs relative to the range of the SMC references (RRMSD) of 24% for both IEM and TV-IEM. The performance metrics for IEM and TV-IEM, i.e. without and with a vegetation correction, are similar if the same retrieval-reference pairs are considered.
Perspectives
In conclusion, the quantification of uncertainty contributions helps to comprehend SMC retrieval accuracy. A large part of the uncertainty originates from the in situ references and the Sentinel-1 σ0 observations. The uncertainty in the Sentinel-2 LAI estimates also has a large impact. The thesis' methods and findings lead to several directions for future research. Future research could focus on the uncertainty sources with the largest contributions to effectively improve the SMC retrievals, assessing the general applicability and improvement of the SMC retrieval scheme, functionally evaluating the SMC retrievals in potential applications, and representing field-scale SMC as an ensemble of SMC products.
