Model-based Tensor Decomposition for Soil Moisture Estimation from Polarimetric SAR Time Series

Kurzfassung

Soil moisture is an essential climate variable that is important for hydrology, climate modeling, and agriculture. Currently, most operational products provide a resolution on the order of kilometers [1], which is insufficient for some applications, such as precision agriculture. To close this gap, this work focuses on data obtained with Synthetic Aperture Radar (SAR) sensors. SAR provides high-resolution and weather-independent information about the dielectric properties of the Earth's surface and, therefore, is an important data source for high-resolution soil moisture products. Future missions like ROSE-L will offer short revisit times combined with a large swath, highlighting the importance of SAR for continuous soil moisture monitoring. In recent years, several approaches to retrieve soil moisture from SAR have been proposed, including short-term backscatter change detection [2], differential interferometry [3], and polarimetry [4]. One of the main challenges is the presence of vegetation or crops covering the ground. While longer wavelengths (e.g., L-band) have a larger penetration and see the ground through the vegetation, the presence of plants cannot be ignored as they strongly influence the signal even at L-band. In this work, we focus on polarimetric decomposition techniques from [4] that allow the separation of the signal into the ground and the vegetation contributions based on the differences in the scattering mechanisms. The data is approximated by a physical model composed of three components: ground, vegetation, and dihedral scattering (interactions of ground and vegetation). The inversion from a single polarimetric acquisition can be ambiguous since there are more model parameters than polarimetric observables. Existing inversion schemes typically require strong assumptions, use simpler models, or set some parameters to fixed values to resolve the ambiguities. This limits the model's validity range as it cannot accurately describe the data in some cases. Instead of using a simpler model, we address the inversion ambiguities by expanding the observation space and adding an additional data dimension, such as a time series of acquisitions. Instead of a single polarimetric matrix, we work with tensors representing stacks of matrices, building on the previous work in [5]. Joint inversion of several matrices reduces the ambiguities and allows the use of models with more parameters to accurately model the surface and vegetation, extending the model's validity range. The inversion is implemented in PyTorch and formulated as an optimization problem that is iteratively solved by gradient descent using automatic differentiation. We evaluate the proposed method on high-resolution airborne F-SAR data obtained during experimental campaigns, including CROPEX 2014 and HTERRA 2022. The model characterizes the dominant scattering mechanisms and provides soil moisture estimates in more regions compared to a simpler X-Bragg model. The model performance depends on the crop type and the phenological stage, indicating that the model can benefit from additional data dimensions like SAR interferometry to improve the ground and vegetation separation. References [1] T. Schmidt, M. Schrön, Z. Li, et al., "Comprehensive quality assessment of satellite-and model-based soil moisture products against the COSMOS network in Germany," Remote Sensing of Environment, 2024. [2] A. Balenzano, F. Mattia, G. Satalino, and M. W. J. Davidson, "Dense temporal series of C- and L-band SAR data for soil moisture retrieval over agricultural crops," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2011. [3] F. De Zan, A. Parizzi, P. Prats-Iraola, and P. L´opez-Dekker, "A SAR interferometric model for soil moisture," IEEE Transactions on Geoscience and Remote Sensing, 2013. [4] I. Hajnsek, T. Jagdhuber, H. Schon, and K. P. Papathanassiou, "Potential of estimating soil moisture under vegetation cover by means of PolSAR," IEEE Transactions on Geoscience and Remote Sensing, 2009. [5] N. Basargin, A. Alonso-Gonz´alez, and I. Hajnsek, "Constrained tensor decompositions for SAR data: Agricultural polarimetric time series analysis," IEEE Transactions on Geoscience and Remote Sensing, 2023.