Conference Agenda
Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).
|
Daily Overview |
| Session | |
|
Land Applications I Location: Red Hall Session Chair: Felix Isundwa, University of Stirling Session Chair: Krištof Oštir, University of Ljubljana | |
    | |
2:50pm - 3:10pm
Scattering Physics and Deep Learning: an Explainable Physics–Informed AI Framework for Soil Moisture Retrieval 1: Tor Vergata University of Rome, Italy; 2: Jet Propulsion Laboratory, California Institute of Technology; 3: Istituto Nazionale di Geofisica e Vulcanologia, Rome, Italy Soil moisture is a crucial parameter in hydrology and agronomy applications [1,2] playing an important role in water resource and irrigation management. Synthetic Aperture Radar (SAR) data have been extensively used in several studies to retrieve the soil moisture across various land covers from bare to vegetated soil, including cropland, grassland, and forest [3,4]. Several scattering models have been developed to retrieve soil moisture beneath vegetation [5–7]. Using first–order radiative transfer (RT) backscattering models requires a complete description of vegetation structure. Thus, extensively used semi–physical models such as the Water Cloud Model (WCM) [8], rely on vegetation information (e.g., vegetation water content and biomass) derived from optical vegetation indices including LAI and NDVI. However, such indices provide a partial description of vegetation structure, leading to soil moisture retrieval errors. In this context, Deep Learning (DL) techniques can be synergically used with physics–based models to optimize their parameters and compensate the lack of vegetation structure description. In this study, we apply the Physics-informed Residual Network (ResNet) model “RTNet”, proposed and validated using airborne data and published in [9] and [10], and specifically designed for accurate and physically meaningful soil moisture retrieval. Particularly, we apply the RTNet to UAVSAR acquisitions at HV polarization, together with multi-source remote sensing (RS) data, e.g., vegetation water content (VWC), soil texture, and weather information. While the HV polarization is used as input to the AI model, the HH polarization is employed, as described later, for the model optimization task. The RTNet is designed to simultaneously perform two optimization tasks. The first one aims to estimate the parameters of a first–order RT model, namely the four scattering contributions (i.e., surface, double, volume, and triple scattering) and the attenuation term. The second task focuses on soil moisture retrieval starting from the optimized scattering components, inspired by the purely physical modeling approach proposed in [11]. Thus, two loss functions are defined: the first minimizes the difference between the measured total backscatter at HH polarization and the estimated one, computed as the sum of the optimized attenuated scattering components, while the second minimizes the difference between the retrieved soil moisture and in–situ measurements. These two losses are properly combined to guide the RTNet training, ensuring an accurate and consistent soil moisture retrieval by coupling it with the optimization of the scattering mechanisms based on physical modeling. To guarantee the physical validity of the estimation and to avoid ill–posed solutions, additional physical constraints are imposed within the loss function. The methodology is validated using soil moisture measurements from the SMAPVEX12 campaign dataset] and the UAVSAR L–band acquisitions aggregated to a 100 meter resolution over six test sites characterized by different land cover types. On the other hand, the optimized RT parameters, i.e., the scattering mechanisms, are evaluated through a comparison with well–established polarimetric decompositions, e.g., the Freeman–Durden three–component decomposition (FD3) [12], the Yamaguchi four–component decomposition with rotation of the coherency matrix (Y4R) [13], to assess the strengths and limitations of the proposed approach. As a future development, the framework will be applied to airborne data, including L–band NISAR data, offering the opportunity to evaluate the robustness of the proposed framework against global soil-vegetation condition variabilities. To conclude, this methodology goes beyond retrieving soil moisture. Instead, through the ingestion of data from heterogeneous sensors and the decomposition of the radar signal into its scattering components, it provides a comprehensive tool useful for investigating soil–vegetation interactions, seasonal changes, and other applications such as land cover characterization and crop type identification, opening new possibilities for development, training, and validation of explainable physics–informed AI models. References [1] Schaufler, G., Kitzler, B., Schindlbacher, A., Skiba, U., Sutton, M.A., & Zechmeister-Boltenstern, S. (2010). Greenhouse gas emissions from European soils under different land use: effects of soil moisture and temperature. European Journal of Soil Science, 61. [2] Sheffield, J., Goteti, G., Wen, F., & Wood, E.F. (2004). A simulated soil moisture based drought analysis for the United States. Journal of Geophysical Research, 109. [3] Hosseini, M., & Mcnairn, H. (2017). Using multi-polarization C- and L-band synthetic aperture radar to estimate biomass and soil moisture of wheat fields. Int. J. Appl. Earth Obs. Geoinformation, 58, 50-64. [4] Baghdadi, N.N., El-Hajj, M., & Zribi, M. (2016). Coupling SAR C-Band and Optical Data for Soil Moisture and Leaf Area Index Retrieval Over Irrigated Grasslands. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 9, 1229-1243. [5] Baghdadi, N.N., & Zribi, M. (2006). Evaluation of radar backscatter models IEM, OH and Dubois using experimental observations. International Journal of Remote Sensing, 27, 3831 - 3852. [6] Fung, A.K., & Chen, K. (2004). An update on the IEM surface backscattering model. IEEE Geoscience and Remote Sensing Letters, 1, 75-77. [7] Bracaglia, M., Ferrazzoli, P., & Guerriero, L. (1995). A fully polarimetric multiple scattering model for crops. Remote Sensing of Environment, 54, 170-179. [8] Attema, E., & Ulaby, F.T. (1978). Vegetation modeled as a water cloud. Radio Science, 13, 357-364. [9] Huang, X., Papale, L.G., Lavalle, M., Del Frate, F., Fattahi, H., Chan, S.K., Lohman, R.B., Xu, X., & Kim, Y. (2025). Optimizing the Radiative Transfer Model Using Deep Neural Networks for NISAR Soil Moisture Retrieval. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 18, 12697–12712. [10] Papale, L.G., Huang, X., Anconitano, G., Lavalle, M., & Del Frate, F. (2025). Integrating Deep Neural Networks and a Radiative Transfer Model for Soil Moisture Retrieval. IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium, 350-354. [11] Hajnsek, I., Jagdhuber, T., Schon, H., & Papathanassiou, K. P. (2009). Potential of estimating soil moisture under vegetation cover by means of PolSAR.IEEE Transactions on Geoscience and Remote Sensing, 47(2), 442-454 [12] Freeman, A., & Durden, S.L. (1998). A three-component scattering model for polarimetric SAR data. IEEE Trans. Geosci. Remote. Sens., 36, 963-973. [13] Yamaguchi, Y., Sato, A., Sato, R., Yamada, H., & Boerner, W. (2010). Four-component scattering power decomposition with rotation of coherency matrix. 2010 IEEE International Geoscience and Remote Sensing Symposium, 1327-1330. | |
