Machine Learning-Based CYGNSS Soil Moisture Estimates over ISMN sites in CONUS

Abstract

Soil moisture (SM) derived from satellite-based remote sensing measurements plays a vital role for understanding Earth’s land and near-surface atmosphere interactions. Bistatic Global Navigation Satellite System (GNSS) Reflectometry (GNSS-R) has emerged in recent years as a new domain of microwave remote sensing with great potential for SM retrievals, particularly at high spatio-temporal resolutions. In this work, a machine learning (ML)-based framework is presented for obtaining SM data products over the International Soil Moisture Network (ISMN) sites in the Continental United States (CONUS) by leveraging spaceborne GNSS-R observations provided by NASA’s Cyclone GNSS (CYGNSS) constellation alongside remotely sensed geophysical data products. Three widely-used ML approaches—artificial neural network (ANN), random forest (RF), and support vector machine (SVM)—are compared and analyzed for the SM retrieval through utilizing multiple validation strategies. Specifically, using a 5-fold cross-validation method, overall RMSE values of 0.052, 0.061, and 0.065 cm3/cm3 are achieved for the RF, ANN, and SVM techniques, respectively. In addition, both a site-independent and a year-based validation techniques demonstrate satisfactory accuracy of the proposed ML model, suggesting that this SM approach can be generalized in space and time domains. Moreover, the achieved accuracy can be further improved when the model is trained and tested over individual SM networks as opposed to combining all available SM networks. Additionally, factors including soil type and land cover are analyzed with respect to their impacts on the accuracy of SM retrievals. Overall, the results demonstrated here indicate that the proposed technique can confidently provide SM estimates over lightly-vegetated areas with vegetation water content (VWC) less than 5 kg/m2 and relatively low spatial heterogeneity.

Cramer–Rao Lower Bound for SoOp-R-Based Root-Zone Soil Moisture Remote Sensing

“…In particular, the compounding effects of surface roughness and topography could be significant over contributing areas within a spaceborne receiver's footprint while will generally be on the order of several hundreds of meters. While previous research has observed successful surface SM estimation from GNSS-R signals under a dominantly coherent signal assumption [33]- [35], the effects of topography and surface roughness deserve an in-depth analysis for these spatial scales [36]- [38]. Under a moderately smooth surface assumption (≈ 1.5 cm), coherent signals at L-band and lower will be minimally effected by surface roughness.…”

Evaluations of Machine Learning-Based CYGNSS Soil Moisture Estimates against SMAP Observations

“…To achieve this, CYGNSS uses eight small satellites that orbit the tropics (within ±38 • latitudes). The considerable amount of CYGNSS land observation data also measured in this region has greatly contributed to the development of new SM retrieval approaches from spaceborne GNSS-R [21][22][23][24][25][26][27][28][29].…”

“…In this paper, we apply a machine learning (ML) framework that is similar to those previously developed by [25,26] to perform SM predictions at a high spatio-temporal resolution. The framework uses in-situ SM data sets provided by the International Soil Moisture Network (ISMN) to train CYGNSS observations that fall within approximately 9 km × 9 km grid centering each ISMN site.…”

“…Several physics-aware features from both CYGNSS observables and other space-borne ancillary data are incorporated into the model to contend with the nonlinear relationship between the CYGNSS signals and SM. While Eroglu et al [25] established artificial neural network (ANN) retrieval over limited ISMN sites as a proof-of-concept using a ten-fold validation method, Senyurek et al [26] generalized the framework by utilizing a larger and more diverse dataset and developing ANN, random forest (RF), and SVM approaches with 5-fold, site-independent, and year-based cross-validation strategies. This paper extends these ML-based studies to global scales by including more reference stations and applying an independent validation strategy.…”

Evaluation of the Land GNSS-Reflected DDM Coherence on Soil Moisture Estimation from CYGNSS Data

“…The problem is that the differences in estimator performance can lead to different results, and the I/F signal dataset used is too small, lacking sufficient persuasive power. Based on the different assumptions, several SM inversion methods have also been developed, such as spatial averaging, combine linear regression method, machine-learning method, and the global inversion accuracy of SM can reach about 0.05 cm 3 /cm 3 [21][22][23][24][25][26][27][28][29][30].…”