**1. Introduction**

Soil moisture (SM) plays a pivotal role in many natural phenomena and processes. For instance, it directly affects crop growth and can be a significant factor in natural disasters such as land degradation, floods, and landslides [1]. These issues have profound impacts, including on food security and the stability of ecological environments, making accurate and real-time monitoring of SM particularly important. However, traditional SM detection methods have notable limitations. These methods primarily rely on direct measurements from ground detectors or meteorological stations, which means they require substantial human and material resources and are time-consuming [2]. Moreover, due to the limitations of these methods, they cannot achieve large-scale, efficient, and low-cost SM retrieval. For vast areas and complex terrains, their detection performance is severely limited. Fortunately, the advent of remote sensing technology provides a new avenue to address this issue. Remote sensing technology can use satellites or drones to monitor the ground from the air, avoiding the difficulties of ground detection and thus achieving large-scale SM retrieval [3,4]. In fact, the European Space Agency (ESA) and the National Aeronautics and Space Administration (NASA) have launched the Soil Moisture and Ocean Salinity (SMOS) satellite [5] and the Soil Moisture Active Passive (SMAP) mission [6] for

**Citation:** Luo, Q.; Liang, Y.; Guo, Y.; Liang, X.; Ren, C.; Yue, W.; Zhu, B.; Jiang, X. Enhancing Spatial Resolution of GNSS-R Soil Moisture Retrieval through XGBoost Algorithm-Based Downscaling Approach: A Case Study in the Southern United States. *Remote Sens.* **2023**, *15*, 4576. https://doi.org/ 10.3390/rs15184576

Academic Editors: Hugo Carreno-Luengo and Chun-Liang Lin

Received: 14 July 2023 Revised: 29 August 2023 Accepted: 14 September 2023 Published: 17 September 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

SM retrieval. Both missions can achieve global SM retrieval with a spatial resolution of about 40 km, and they can revisit the globe every 2–3 days. However, while remote sensing technology and related satellite missions such as SMOS and SMAP provide global SM retrieval capabilities, the resolution of these products is relatively low, making them more suitable for large-scale applications. For medium- and small-scale applications that require more detailed observations, such as irrigation management in farmland or flood warnings in specific areas, these methods may not meet the needs.

The technique of Global Navigation Satellite System-Reflectometry (GNSS-R) represents a novel type in the field of remote sensing. Its internal L-band signal source is adequate and exhibits high penetration capabilities for vegetation, soil, snow, etc. It is capable of all-weather, all-day observation and has excellent potential for SM retrieval [7–9]. GNSS-R receivers are generally installed on the ground or on aircraft. Although they have an excellent detection accuracy, the monitoring range limits its ability to achieve a wide range of SM retrieval [10]. CYGNSS was successfully launched in 2016, with a revisit cycle of 2.8 (median) and 7.2 (average) hours [11], providing ample data for SM retrieval by GNSS-R technique. Thus, using the GNSS-R technique to retrieve SM has become a hot research topic in recent years. Chew et al. [12] showed that there is a strong linear relationship between the surface reflectance of CYGNSS and SMAP SM, and a global SM product with a resolution of 36 km was produced through linear method. Ruf [13] proposed that SMAP SM can be supplemented by using the relative signal-to-noise ratio (rSNR) of CYGNSS to SM retrieval. Al-Khaldi et al. [14] considered that vegetation and surface roughness would affect SM. They proposed a method for CYGNSS SM retrieval through time series. A global SM product of 0.2◦ × 0.2◦ was finally generated. Considering that the terrain, vegetation, and surface roughness have an impact on the GNSS signal, the relationship between the signal and SM is relatively complex and nonlinear. Machine learning has been frequently used in the study of CYGNSS SM retrieval because of its great advantages in handling nonlinear situations. Eroglu et al. [15] combined CYGNSS observables with in situ sites observations, Vegetation Water Content (VWC), Normalized Vegetation Index (NDVI), and topography features. Finally, a daily SM product with a resolution of 9 km was generated using the Artificial Neural Network (ANN). Senyurek et al. [16] obtained the daily SM of the United States with a resolution of 36 km using CYGNSS and in situ site observations based on machine learning algorithms. The results showed that the prediction effect of Random Forest (RF) was the best, with an RMSE of 0.052. Jia et al. [17] pre-classified land cover types and used the eXtreme Gradient Boosting (XGBoost) method for SM retrieval. Compared with the accuracy of SM retrieval without pre-classification, there was an improvement, with an RMSE of 0.052.

However, the SM products obtained from the aforementioned microwave remote sensing data have a coarse resolution, which limits their utility in medium- and small-scale hydrological and agricultural applications. Zhan et al. [18] first introduced an empirical polynomial for downscaling, marking an initial exploration of effective strategies to address this issue. Subsequently, Chauhan et al. [19] improved upon Zhan's method, enhancing its performance. In this empirical polynomial downscaling method, high-resolution SM is expressed as a polynomial function of surface temperature, plant index, and surface reflectance derived from brightness temperature data. This innovative method provides a fresh perspective for tackling the downscaling of SM. Piles et al. [20] further optimized this downscaling polynomial fitting method. Their improvement replaced surface reflectance in the polynomial equation with coarse-resolution brightness temperature data, making the method more flexible and efficient in handling practical problems. Moreover, this polynomial fitting downscaling method has been widely applied in the downscaling of various SM products, such as SMOS and AMSR-E, and also in various high-resolution remote sensing image products, such as MODIS and MSG-SEVIRI. This has been confirmed by many scholars [21–26]. Their research further validates the practicality and broad application value of this method. In order to retrieve daily SM at a 9 km resolution, Das et al. [27] downscaled the coarse-resolution (approximately 40 km) SMAP L-band

brightness temperature data using the high-resolution (1–3 km) L-band Synthetic Aperture Radar (SAR) backscatter observations. Based on artificial intelligence techniques including Support Vector Machines, Artificial Neural Networks, and Associated Vector Machines, Srivastava et al. [28] fused MODIS surface temperature with SMOS SM and enhanced the spatial resolution of SMOS SM by using downscaling methods. The factor used to represent the high-resolution state of SM plays a crucial role in determining the accuracy of the downscaled SM. The downscaled SM has higher accuracy compared to the original coarseresolution SMOS and AMSR-E SM, with the *R* rising from 0.27 to 0.96 [29]. Compared to the observed data, the accuracy of the downscaled SM has improved relative to the products of SMOS and AMSR-E [30]. This means that downscaling methods could be attempted to provide high-resolution SM for products such as SMAP, SMOS, AMSR-E, and NASA-USDA.

The aforementioned research demonstrates both the significant advantages of using GNSS-R technique for SM retrieval and the notable effects of using downscaling methods to enhance the spatial resolution of SM products. However, no studies have yet used the downscaling method to improve the spatial resolution of GNSS-R technique. At present, the spatial resolution achieved by SM retrieval based on spaceborne GNSS-R is limited (up to 9 km). Spatial downscaling of microwave SM is a crucial strategy. It addresses the pressing need for higher spatial resolution SM data, which is essential for local hydrological or agricultural applications. Therefore, this paper proposes a method for constructing a SM downscaling model. This method aims to fuse the CYGNSS observables and auxiliary variables with SMAP SM (36 km) products, forming a nonlinear relationship at the same scale. Finally, a downscaling model will be built based on the XGBoost algorithm to retrieve SM with a spatial resolution of 3 km. In the end, the SM retrieval using GNSS-R technique is successfully spatially downscaled, improving the spatial resolution of SM retrieval.
