Controls of surface soil moisture spatial patterns and their temporal stability in a semi-arid steppe

Soil moisture is an integral quantity parameter in hydrology and agriculture practices. Satellite remote sensing has been widely applied to estimate surface soil moisture. However, it is still a challenge to retrieve surface soil moisture content (SMC) data in the heterogeneous catchment at high spatial resolution. Therefore, it is necessary to improve the retrieval of SMC from remote sensing data, which is important in the planning and efficient use of land resources. Many methods based on satellite-derived vegetation indices have already been developed to estimate SMC in various climatic and geographic conditions. Soil moisture retrievals were performed using statistical and machine learning methods as well as physical modeling techniques. In this study, an important experiment of soil moisture retrieval for investigating the capability of the machine learning methods was conducted in the early spring season in a semi-arid region of Iran. We applied random forest (RF), support vector machine (SVM), artificial neural network (ANN), and elastic net regression (EN) algorithms to soil moisture retrieval by optical and thermal sensors of Landsat 8 and knowledge of land-use types on previously untested conditions in a semi-arid region of Iran. The statistical comparisons show that RF method provided the highest Nash–Sutcliffe efficiency value (0.73) for soil moisture retrieval covered by the different land-use types. Combinations of surface reflectance and auxiliary geospatial data can provide more valuable information for SMC estimation, which shows promise for precision agriculture applications.

Drylands cover about 40% of the world’s land area and support two billion people, most of them living in developing countries that are at risk due to land degradation. Over the last few decades, there has been warming, with an escalation of drought and rapid population growth. This will further intensify the risk of desertification, which will seriously affect the local ecological environment, food security and people’s lives. The goal of this research is to analyze the hydrological and land cover characteristics and variability over global arid and semi-arid regions over the last decade (2010–2019) using an integrative approach of remotely sensed and physical process-based numerical modeling (e.g., Global Land Data Assimilation System (GLDAS) and Famine Early Warning Systems Network (FEWS NET) Land Data Assimilation System (FLDAS) models) data. Interaction between hydrological and ecological indicators including precipitation, evapotranspiration, surface soil moisture and vegetation indices are presented in the global four types of arid and semi-arid areas. The trends followed by precipitation, evapotranspiration and surface soil moisture over the decade are also mapped using harmonic analysis. This study also shows that some hotspots in these global drylands, which exhibit different processes of land cover change, demonstrate strong coherency with noted groundwater variations. Various types of statistical measures are computed using the satellite and model derived values over global arid and semi-arid regions. Comparisons between satellite- (NASA-USDA Surface Soil Moisture and MODIS Evapotranspiration data) and model (FLDAS and GLDAS)-derived values over arid regions (BSh, BSk, BWh and BWk) have shown the over and underestimation with low accuracy. Moreover, general consistency is apparent in most of the regions between GLDAS and FLDAS model, while a strong discrepancy is also observed in some regions, especially appearing in the Nile Basin downstream hyper-arid region. Data-driven modelling approaches are thus used to enhance the models’ performance in this region, which shows improved results in multiple statistical measures ((RMSE), bias (ψ), the mean absolute percentage difference (|ψ|)) and the linear regression coefficients (i.e., slope, intercept, and coefficient of determination (R2)).