Speaker
Description
The agricultural asset is facing a transition towards an increasingly sustainable resources management. In this context, soil moisture represents a key geophysical variable for many precision farming applications such as smart irrigation and crop yield estimation. Synthetic Aperture Radar (SAR) has been demonstrated to be one of the most valuable sources of information for accurate and continuative estimation of soil moisture with high spatial and temporal resolutions in almost any weather conditions. However, when observing the soil in agricultural areas, the developing vegetation cover causes additional signal attenuation and scattering mechanisms, leading to less accurate soil moisture retrievals. Artificial Neural Networks (ANNs) have been demonstrated to be an important instrument for Earth Observation (EO) applications designed to retrieve information from Remote Sensing (RS) data. In general, when using AI for geophysical parameters estimation, there is a risk that the physical meaning behind the ANN mapping criteria becomes challenging to understand. Nevertheless, if properly trained, AI algorithms can reproduce the physics associated with the phenomenon we are observing, as shown in our previous work [2]. This study aims to synergically adopt electromagnetic data modelling and a Machine Learning (ML) algorithm to separate the different scattering contributions (i.e., volume, surface and double bounce scattering). More in detail, the component associated with the vegetation cover (volume scattering) is neglected so that only the surface and double bounce scattering can be used for more effective soil moisture retrieval algorithms. For this purpose, the “Tor Vergata” electromagnetic model developed by Bracaglia et al. [3] was adopted to generate a set of simulated Mueller matrices at L-band associated to the backscatter components which address the different canonical scattering mechanisms for VV, HH and HV polarizations. The simulated data were generated considering different values of vegetation- and soil-related variables (plant height and soil moisture/roughness) while setting the sensor configuration variables such as frequency and incidence angle [2]. Moreover, to motivate the choice of isolating and using the soil backscatter for soil moisture estimation, a preliminary analysis on the sensitivity of simulated total backscatter and the soil-related components to soil moisture was performed. As a first result, it is possible to observe that the total backscatter is less sensitive to soil moisture variations if compared to the soil related components. Besides, the Pearson correlation coefficient is higher when the soil backscatter component is considered. Then, the simulated dataset was used to train a properly designed ANN model taking as input the elements of the Mueller matrix and using the backscatter soil-related components (soil and double bounce scattering) in the selected polarizations as targets. In order to assess the performances of the proposed methodology, the ESA BelSAR2018 dataset, providing full-polarimetric SAR data at L-band, acquired with aerial surveys and in-situ soil moisture measurements, is considered to derive the correlation between the soil moisture itself and the soil-related backscatter estimated by the trained ANN model. Moreover, the sensitivity of those components to the soil moisture will be compared with the sensitivity shown by applying the canonical polarimetric decompositions (e.g., Freeman Durden [4], Generalized Freeman Durden [5], Pauli [6], Yamaguchi [7]). In this regard, it will be shown how the proposed approach can help in estimating soil moisture by mitigating the vegetation effect, demonstrating that the isolated soil-related backscatter components retrieved by the ANN are more sensitive to soil moisture if compared to the total received signal.
Keywords: electromagnetic modeling, Machine Learning, Synthetic Aperture Radar, SAR polarimetry, L-band, soil moisture, vegetation, agriculture.
Acknowledgments
We would like to thank the ESA BelSAR2018-Campaign (https://doi.org/10.5270/ESA-bccf2d9) team for the collection of the SAR and on-field datasets used in this work.
References
[1] Xing, M.; Chen, L.; Wang, J.; Shang, J.; Huang, X. Soil Moisture Retrieval Using SAR Backscattering Ratio Method during the Crop Growing Season. Remote Sens. 2022, 14, 3210.
[2] L. G. Papale, F. Del Frate, L. Guerriero and G. Schiavon, "A Physics-Based ML Approach for Corn Plant Height Estimation with Simulated Sar Data," IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, Kuala Lumpur, Malaysia, 2022, pp. 4159-4162.
[3] M. Bracaglia, P. Ferrazzoli, and L. Guerriero, “A fully polarimetric multiple scattering model for crops,” Remote Sens. Env., pp. 170-179, 1995.
[4] A. Freeman and S. L. Durden, “A three-component scattering model for polarimetric SAR data,” IEEE Trans. Geosci. Remote Sens., pp. 963-973, May 1998.
[5] Jong-Sen Lee and Eric Pottier, Polarimetric Radar Imaging: From Basics to Applications, CRC Press, 2009
[6] S.R. Cloude, “Polarisation: Applications in Remote Sensing”, Oxford University Press, ISBN 978-0-19-956973-1, 2009.
[7] Y. Yamaguchi, A. Sato, W. Boerner, R. Sato and H. Yamada, “Four-Component Scattering Power Decomposition With Rotation of Coherency Matrix,” IEEE Trans. Geosci. Remote Sens., pp. 2251-2258, June 2011.
