The energy and water vapor exchange between the land surface and atmospheric boundary layer plays a critical role in regional climate simulations. This paper implemented a hybrid data assimilation and machine learning framework (DA-ML method) into the Weather Research and Forecasting (WRF) model to optimize surface soil and vegetation conditions. The hybrid method can integrate remotely sensed leaf area index (LAI), multi-source soil moisture (SM) observations, and land surface models (LSMs) to accurately describe regional climate and land–atmosphere interactions. The performance of the hybrid method on the regional climate was evaluated in the Heihe River basin (HRB), the second-largest endorheic river basin in Northwest China. The results show that the estimated sensible (
Land–atmosphere interactions are an essential component of the hydrological
cycle and factors that influence climate change (Nelli et al., 2020; Zhou et
al., 2022). Terrestrial components, such as soil and vegetation, play a crucial role in atmospheric processes, such as changes in evapotranspiration (ET), which affect the water vapor content in the atmosphere (Gentine et al., 2019; Sawada et al., 2015; Wu et al., 2023). Soil and vegetation processes directly affect surface water vapor transport and energy circulation, particularly in the arid vegetated area (Erlandsen et al., 2017; Gao et al., 2008; R. Liu et al., 2018; X. Zhang et al., 2017; Zhang et al., 2019; Zhao et al., 2021). Such surface variability affects the available energy distribution at the land surface and has additional effects on the sensible and latent heat fluxes (
The development of Earth observation technology has provided important opportunities to study land–atmosphere interactions using the data assimilation (DA) method (Liang et al., 2021). The Land Data Assimilation System (LDAS) has been widely developed and applied in recent years under various hydrological and vegetation conditions (Wu et al., 2022; Xia et al., 2019). It uses remotely sensed observations to constrain model physical processes and empirical parameters to improve water–energy–carbon flux simulations (Tian et al., 2022; Zhao and Yang, 2018). A series of studies have assimilated satellite-retrieved leaf area index (LAI), land surface temperature (LST), soil moisture (SM), and microwave brightness temperature observations into LSMs and improved simulations of ET, runoff, and gross primary productivity (GPP; Ahmad et al., 2022; X. He et al., 2020, 2021; Ling et al., 2019; Seo et al., 2021; Xie et al., 2017; Xu et al., 2019, 2021). In addition, DA can improve the initial conditions of regional climate models (RCMs) and enhance the capability of the models in the simulation of land–atmosphere interactions (Pan et al., 2017; Yi et al., 2021). Several studies have also shown that the assimilation of air pressure, air temperature, humidity, wind speed, and lightning observations into the Weather Research and Forecasting (WRF) model can improve the simulation of atmospheric state variables and the accuracy of weather prediction (Campo et al., 2009; Cazes Boezio and Ortelli, 2019; Comellas Prat et al., 2021; Grzeschik et al., 2008; Pilguj et al., 2019).
Machine learning (ML) algorithms have been increasingly applied in Earth and
environmental modeling studies to predict land surface variables at various
spatial and temporal scales (Jung et al., 2020; Reichstein et al., 2019; Xu et al., 2018). Compared to physical models, ML technology can fluently and
accurately establish nonlinear and complex relationships between diverse
independent variables (Koppa et al., 2022; Nearing et al., 2018). Thus,
ML-based approaches can create beneficial pathways for knowledge discovery
in process models, based on extensive data (Moosavi et al., 2021; Reichstein et al., 2019). The main improvements are focused on model approximation,
parameterization, bias correction, and hybrid modeling (Brajard et al., 2020; He et al., 2022; Jia et al., 2021; Xu et al., 2014; Zhao et al., 2019). Several studies have shown that the integration of the DA and ML methods can enhance the reliability of predictions and reduce simulation errors by including physical information in observed data (Brajard et al., 2020; Buizza et al., 2022; Forman and Xue, 2017; Gottwald and Reich, 2021; He et al., 2022). Forman and Xue (2017) integrated a ML model (as a measurement operator) into a DA system to improve the estimation of the snow water equivalent. Zhao et al. (2019) used a physics-constrained ML method to improve the
The Heihe River basin (HRB) is a typical endorheic river basin in the arid and semi-arid regions of Northwest China (Li et al., 2013). The upstream mountain region is mainly covered by alpine meadow and Qinghai spruce, has a complex topography, and receives abundant precipitation. The midstream oasis of the HRB is mainly characterized by irrigated croplands, while the downstream oasis is characterized by riparian forests, and at the periphery of the oasis, there is vast desert (Xu et al., 2020). In the HRB, precipitation is the main water resource input in mountainous areas and determines the growth of vegetation in the oasis region, in addition to supporting urban and population development (Li et al., 2018, 2021). Precipitation, snow, and changes to permafrost in the upstream mountains can affect mountain runoff and SM, evaporation, and the groundwater table in the mid- and downstream oases. Strong land–atmosphere interactions in the HRB affect the water and energy exchange between the surface and atmosphere and influence the sustainability of the oasis (Gao et al., 2008; Pan et al., 2021b). The oasis–desert local circulation in the HRB can lead to the microclimate features in oasis–desert areas, which include the cooling and wetting effect and wind shield effect of the oasis and the humidity inversion effect within the surrounding desert (Liu et al., 2020).
In recent decades, several comprehensive experiments have been implemented over the HRB to study land–atmosphere interactions, including the Heihe River basin field experiment (Hu et al., 1994), Watershed Allied Telemetry Experimental Research (WATER; Li et al., 2009), and Heihe Watershed Allied Telemetry Experimental Research (HiWATER; Li et al., 2013). In recent years, many mesoscale climate models and high-resolution computational fluid dynamics (CFD) models have been used to analyze the effects of land–atmosphere interactions on the regional climate (R. Liu et al., 2018, 2020; Xie et al., 2018; X. Zhang et al., 2017; Zhang et al., 2021a). X. Zhang et al. (2017) added an irrigation scheme to the WRF model and identified strong cooling and wetting effects on irrigated cropland in the midstream of the HRB. Liu et al. (2020) investigated the oasis–desert microclimate effects based on an improved CFD model and found that the oasis had a cold and wet island effect and wind shield effect. Zhang et al. (2021b) applied the WRF-Hydro model in the HRB and emphasized the role of lateral flow in the regional precipitation circulation. These studies illustrate that mesoscale climate models can be used as essential tools to better understand regional climate and land–atmosphere interactions in the HRB. However, the advantages of improving the representation of soil and vegetation processes in affecting regional climate via the coupled DA and ML framework have not been fully exploited, especially in basins with complex underlying surfaces. Therefore, this study aims to investigate the improvement in the hybrid DA and ML framework for regional climate and land–atmosphere interactions in the HRB, based on the WRF model, and to further reveal its physical mechanisms.
The goals of this study were to (1) couple the hybrid DA and ML (DA-ML)
framework to the WRF model and improve the estimation of LAI and SM, (2) validate the
The HRB (37.7–42.7
To provide a high-resolution land cover and soil texture dataset that matched the WRF simulation period, the regional land cover and soil texture product generated by Zhong et al. (2014) and Song et al. (2016), with a spatial resolution of 30 m, was employed. These datasets were downloaded from the National Tibetan Plateau Data Center (TPDC; Pan et al., 2021a;
In this study, SM observations from the ecohydrological wireless sensor
networks (WSNs) up- and midstream regions of the HRB are used as an independent validation to evaluate the SM estimates from the WRF (DA-ML). The validation SM dataset in the upstream regions was mainly covered by grassland and obtained by averaging SM observations from 40 nodes. There are nine network nodes installed in the LAS source area at Daman station that measured SM at the depths of 10 cm every 5 min (Che et al., 2019; S. Liu et al., 2018) (
WRF model setup. Note that RRTM is for the rapid radiative transfer model.
The advanced research WRF model version 4.0.3 (Skamarock et al., 2019) was
used in this study. The WRF is a state-of-the-art numerical weather and climate model designed by the National Center for Atmospheric Research (NCAR) for meteorological research and numerical weather predictions (Wang et al., 2021). The model source code is available at the official repository for WRF (
In this study, the hybrid model proposed by He et al. (2022) based on the DA and ML methods was incorporated into the WRF model to improve the LAI, SM, and ET simulations. The hybrid approach relies on the DA method to update the vegetation dynamics of the Noah-MP model and the ML method to construct a three-layer SM surrogate model. Compared with the direct assimilation of coarse-resolution remotely sensed SM, the hybrid model can improve the estimation of SM and ET on the heterogeneous land surface. This is because in situ SM profile observations are used to construct an ML-based surrogate model to improve SM and ET estimation on complex underlying surfaces.
In the DA part, the remotely sensed LAI was assimilated using the ensemble
Kalman filter (EnKF) method to update the leaf biomass (LFMASS) and optimize
the specific leaf area (SLA) in the Noah-MP model. LAI is estimated as the
product of leaf biomass predictions and SLA (LAI
In the ML part, the normalized soil texture (ST), land cover (LC), air
temperature and humidity (Ta and RH), wind speed (
The coupled land–atmosphere DA-ML system consists of two steps. In the first
step, the meteorological forcing data were generated from the WRF model at
time
Two experiments were conducted using the WRF model to investigate the effects of the DA-ML method. These two experiments consisted of an experiment under natural conditions without DA and ML (WRF (OL), where OL is open loop) and another experiment implementing DA and ML (WRF (DA-ML)). The simulation covered the period of the vegetation growing season (from 1 April to 30 September 2015), and the first month was used for the spin-up. The computation details about the WRF (OL) and WRF (DA-ML) are shown in Table A1. The differences between the WRF (DA-ML) and WRF (OL) simulations were used to investigate the effects of LAI and SM integration. The root mean square deviation (RMSD) and coefficient of determination (
Figure 3 shows the monthly averaged LAI estimates from the WRF (OL), WRF (DA-ML), and GLASS products. As indicated, the WRF model failed to capture the magnitude and seasonality of the LAI. This is because the simulation of LAI dynamics in Noah-MP is controlled by the planting date, harvest date, and growing degree days in the cropland (X. Liu et al., 2016). In addition, an inaccurate specification of the SM saturation, Vcmax25, and Clapp–Hornberger
Seasonal variations in the LAI estimates for cropland, grassland, forest, and shrubland in the HRB.
The SM estimates from the WRF (OL) and WRF (DA-ML) are validated over the up- and midstream WSNs in Fig. 4. The SM estimates from the WRF model were markedly lower than WSN observations for cropland because the impacts of irrigation events on SM estimates are not fully considered in the Noah-MP model (He et al., 2022; Zhang et al., 2020). The Noah-MP model also slightly underestimates SM in the upstream regions because it ignores the effects of dense root systems and soil organic matter on SM estimation (Chen et al., 2012; Sun et al., 2021). As anticipated, SM predictions from the WRF (DA-ML) are closer to the measurements than those of WRF. WRF (DA-ML) SM retrievals indicate a reasonable response to the precipitation and irrigation events in the midstream cropland. Similarly, the WRF (DA-ML) SM dynamics show a characteristic response to precipitation in the upstream regions. The results also indicate that the SM simulations from the WRF (DA-ML) model find it hard to capture the observed peak values. This is because the prediction accuracy of the ML methods is limited by the training dataset. This also means that if the model is applied under extremely wet conditions with sparse training data, then the performance of the hybrid model will decrease as the number of training samples decreases. In general, the WRF (DA-ML) can use the information contained in remotely sensed LAI and multi-source SM observations to improve land surface conditions.
The time series of SM estimates from the WRF (OL) and WRF (DA-ML) models against the up- and midstream WSN observations in 2015.
Figure 5 shows the spatial patterns of the averaged LAI and SM estimates from May to September 2015. The WRF simulation significantly underestimated the LAI, particularly in the up- and midstream vegetation areas of the HRB. In addition, it underestimated the SM in the mid- and downstream vegetation
regions. The integration of LAI and SM into the WRF model improved the
estimation of leaf biomass and SM and increased LAI and SM in the HRB. The
maps of estimated LAI and SM from the DA-ML method consistently resembled
the rainfall, vegetation cover, irrigation event, and shallow groundwater
table features (Xu et al., 2018, 2020). The precipitation in the upstream
mountains, irrigation in the midstream oasis, and shallow groundwater in the
downstream oasis enhance SM and provide the necessary water supply for
vegetation growth (Li et al., 2022). Figure 5 also shows the LAI and SM differences between the WRF (DA-ML) and WRF (OL) simulations. The maximum
LAI (SM) difference from the WRF (DA-ML) and WRF (OL) simulations reaches approximately 2.24 m
The LAI and SM estimates from the WRF (OL) and WRF (DA-ML) models during the growing season in 2015 and the average difference in the LAI and SM between the WRF (DA-ML) and WRF (OL) (i.e., WRF (DA-ML) minus WRF (OL)).
Scatterplot of daily sensible and latent heat flux estimates from the WRF (OL) and WRF (DA-ML) models versus measurements at the Arou, Daman, and Sidaoqiao sites.
Statistical indices of daily
Figure 6 compares the daily
Spatial distribution of evapotranspiration estimates obtained from the WRF (OL), WRF (DA-ML), and ETMap during the growing season in 2015.
Figure 7 shows the spatial distribution of ET estimates from the WRF (OL), WRF (DA-ML), and ETMap over the HRB. The results indicate that the ET values from the WRF model were underestimated, especially in the midstream oasis region, which was mainly because the WRF model underestimated the SM and LAI (see Figs. 3 and 4) during the growing season. Compared with the WRF (OL) model, the WRF (DA-ML) method improves the estimation of ET, and the spatial distribution is consistent with that of ETMap because of the effective information contained in the remote sensing LAI and multi-source SM observations. The estimation of ET in the WRF (OL) is sensitive to SM and vegetation dynamics, especially in semi-arid regions. Therefore, the WRF (DA-ML) model will produce more improvements in the mid- and downstream oasis regions compared to the WRF (OL) model. The spatial patterns of ET from the DA-ML method showed a significant gradient from wet to dry, owing to variations in the precipitation and vegetation cover. In the upstream regions of the HRB, the spatial pattern of retrieved ET was mainly controlled by precipitation and vegetation cover. The ET values were higher in areas with heavier precipitation and denser vegetation. In the midstream region, the spatial pattern of ET was well aligned with the oasis caused by crop growth and irrigation. Meanwhile, the ET values were higher in the downstream oasis because of shallow water tables and transpiration from riparian forests (Xu et al., 2018, 2020). The sparsely vegetated areas covered by desert and Gobi in the mid- and downstream regions had the lowest ET values. The results show that the integration of remotely sensed LAI and multi-source SM observations is essential for studying land–atmosphere water vapor fluxes (ET) because of the realistic land surface conditions.
The monthly averaged 2 m air temperature and specific humidity from the WRF (OL), WRF (DA-ML), and corresponding observations at nine sites are shown in Figs. 8 and 9. As indicated, the WRF model overestimated (underestimated)
the air temperature (specific humidity) in the HRB, especially in the midstream oasis (Daman and wetland stations), which was mainly because the WRF model underestimated the SM and LAI (see Figs. 3 and 4) in the HRB. Compared to the WRF model, the WRF-(DA-ML)-simulated seasonal cycles of air
temperature and specific humidity at the nine sites were closer to the
measurements, which was because the integration of remotely sensed LAI and
multi-source SM observations improves the estimation of vegetation dynamics
and SM, decreases the air temperature, and increases the specific humidity. The increased specific humidity was due to the enhanced evaporation from the
soil and stronger transpiration from the expanded vegetation cover. Simultaneously, evaporation absorbs a large amount of energy, thereby
reducing the air temperature (Wen et al., 2012). The discrepancy between the
WRF (OL) and WRF (DA-ML) was amplified in the middle of the growing season
(June, July, and August) due to dense growing vegetation and higher SM caused by several irrigation events. After integrating LAI and SM, the simulated air temperature and specific humidity values from the Daman station decreased and increased by approximately 1.75 K and 1.86 g kg
Monthly averaged air temperature simulations from the WRF (OL) and WRF (DA-ML) versus the observations at nine sites in 2015 (error range denotes the standard deviation).
Monthly averaged specific humidity simulations from the WRF (OL) and WRF (DA-ML) versus the observations at nine sites in 2015 (error range denotes the standard deviation).
Averaged air temperature and
Tables 3 and 4 further compare the simulated air temperature and specific
humidity with the same variables from the station observations. The WRF (OL)
results show a dry bias in the HRB region, which is reduced by the simulation of the WRF (DA-ML). The statistical metrics (i.e.,
Averaged specific humidity and
Figure 10 compares the spatial patterns of the air temperature and specific humidity maps from the WRF (OL) and WRF (DA-ML). Compared with the WRF (OL), significant differences were observed in the WRF (DA-ML). The integration of LAI and SM decreases air temperature and increases specific humidity in the vegetated area of the HRB, particularly in the midstream oasis region. The spatial distribution of specific humidity from the WRF (DA-ML) is consistent with the LAI and SM maps in Fig. 5 and the ET map in Fig. 7. The results show that the improved LAI and SM simulations lead to different land surface dynamic and thermal characteristics between the oasis and desert. This difference leads to oasis–desert interactions and produces microclimatic effects, including the cooling and wetting effects of the oasis. The average simulated air temperature from WRF (OL) and WRF (DA-ML) methods in the midstream oasis were 293.64 and 291.32 K, respectively. In contrast, the near-surface air temperatures over the desert are approximately 294.13 and 293.54 K, respectively. The difference in air temperature between the oasis and desert areas indicates that the oasis areas represent a cold and wet island compared to the surrounding desert. This difference is amplified after the implementation of the DA-ML method. The significant wetting and cooling effects propagate in desert areas to a maximum distance of approximately 5–10 km from the edge of the oasis. In the midstream oasis, the dominant vegetation is irrigated cropland, and the vegetation cover was only approximately 42 % in the original WRF model; however, the vegetation cover was updated to approximately 70 % in the WRF (DA-ML). The different land surface dynamic and thermal characteristics between the oasis and desert can produce oasis–desert interactions and enhance local circulation. The oasis–desert interactions create a water vapor flux from the oasis to the surrounding desert. This transport process is beneficial for increasing desert water vapor and maintaining the sustainability of desert vegetation (Li et al., 2016; Liu et al., 2020). A similar pattern was observed in the downstream oasis. However, because of the decreased SM and vegetation cover (Fig. 5), the downstream oasis exhibited a weaker wet island effect. The results also indicated that enhanced vegetation transpiration increases specific humidity and reduces air temperature owing to increased LAI in the upstream region of the HRB.
Spatial distribution of the air temperature and specific humidity estimates from the WRF (OL) and WRF (DA-ML) during the growing season in 2015 and the average difference in air temperature and specific humidity between the WRF (DA-ML) and WRF (OL) (i.e., WRF (DA-ML) minus WRF (OL)). The blue line indicates the mid- and downstream oasis vertical profile used in Figs. 11 and 12.
Mean vertical profile of differences in air temperature and specific humidity between the WRF (DA-ML) and WRF (OL) (i.e., WRF (DA-ML) minus WRF (OL)) and mean LAI and SM during the growing season in 2015 in the midstream oasis. The dashed and solid lines represent the WRF (OL) and WRF (DA-ML), respectively. The shaded white area represents the change in elevation. The orange bar represents the oasis area.
The abovementioned findings show that the proposed WRF (DA-ML) method exhibits strong wetting and cooling effects in the mid- and downstream oasis. These wetting and cooling effects reduce the air warming bias and dry bias in the simulation. Therefore, the WRF (DA-ML) simulation is much closer to the observations than the WRF (OL) simulation. Two vertical profiles were selected in Fig. 10 to further analyze the effect of the DA-ML on the local climate in the mid- and downstream areas. The difference between the
WRF (DA-ML) and WRF (OL) methods is used to represent the enhanced cooling
and wetting effects after improving the LAI and SM simulations. As illustrated in Fig. 11, the enhanced wetting and cooling effects of the
midstream oasis were the strongest in the southern irrigated cropland and gradually decreased in the northern desert areas. The magnitudes of the surface wetting and cooling effects were consistent with the differences in
LAI and SM estimates from the WRF (DA-ML) and WRF (OL). For example, the
difference in LAI and SM peaks at 38.08
Figure 12 shows the same wetting and cooling effects in the downstream oasis. Compared to the midstream irrigated cropland, the downstream oasis wetting and cooling effects were mainly influenced by the growth of riparian forests and shallow groundwater tables. The wetting and cooling effects showed maximum values at 42.01
Mean vertical profile of differences in air temperature and specific humidity between the WRF (DA-ML) and WRF (OL) (i.e., WRF (DA-ML) minus WRF (OL)) and mean LAI and SM during the growing season in 2015 in the downstream oasis. The dashed and solid lines represent the WRF (OL) and WRF (DA-ML), respectively. The orange bar represents the oasis area.
The mean wind vectors at 10 m during the growing season from the WRF (OL) and WRF (DA-ML) in the mid- and downstream oases are shown in Fig. 13. By
comparing the simulated wind speeds in the oasis and the surrounding desert, we found that crops, shelterbelts, and residential areas in the midstream oasis produced a wind shield effect. The wind speed within the oasis is less
than that of the surrounding desert because the drag force of crops, shelterbelts, and residential areas reduces the wind speed and also changes
the wind direction (Liu et al., 2020). In Fig. 13, the heat transfer coefficient (
The integration of the LAI and SM affected the wind speed at the land surface and the local circulation through oasis–desert interactions. Figure 14 shows the zonal mean vertical velocity and local meridional circulation in the midstream oases from the WRF (OL) and WRF (DA-ML). Compared with the flat topography of the downstream oasis, the topography of the midstream oasis generally varies from plains to mountains (from low to high altitude) from north to south. The surface dynamic and thermal characteristics of the oasis and surrounding desert differed significantly; therefore, strong horizontal temperature and humidity field gradients were observed at the intersection of the boundary layer of the mountains, oasis, and surrounding desert (Meng et al., 2015; Wen et al., 2012). The air humidity and vegetation cover in the midstream oasis were enhanced by integrating the LAI and SM, which resulted in stronger evaporation from irrigated cropland than from the surrounding desert. As shown in Fig. 14, the divergence of the lower atmosphere over the midstream oasis is enhanced, and the wet and cold air masses are transferred to the surrounding desert through advection, whereas the dry and hot air is transferred into the oasis from the upper atmosphere. In the upper atmosphere, the desert-to-oasis air masses enhance the background northerly winds, which promote atmospheric water vapor transport in the HRB. However, oasis–desert interactions are weaker in the downstream region (Fig. A1) than in the midstream region under actual weather or climate conditions, which is attributed to the local circulation being weakened by stronger background northerly winds. Overall, the simulation of soil and vegetation characteristics can be improved by integrating LAI and SM and enhancing land–atmosphere interactions in mid- and downstream oases.
The zonal mean vertical velocity and meridional circulation from the WRF (OL) and WRF (DA-ML) models during the growing season in 2015 in the midstream oasis. The shaded white area represents the change in elevation. The orange bar represents the oasis area.
Figure 15 exhibits the influence of the DA-ML on precipitation in the HRB. The results show that the integrated LAI and SM led to increased precipitation in the upstream regions of the HRB and that the spatial variation in precipitation was very heterogeneous. The increase in precipitation was mainly concentrated in the southeastern part of the HRB, where it reached approximately 1.5 mm d
Comparisons of the daily precipitation estimates from the WRF (OL) and WRF (DA-ML) with the values from the AFD and CMFD references during the growing season in 2015.
The simulated daily precipitation from the WRF (OL) and WRF (DA-ML) was compared with that of the AFD and CMFD references in Fig. 16. Because the water vapor from the East Asian monsoon was blocked by the Tibetan Plateau,
most of the precipitation was concentrated in the southeastern part of the
Qilian Mountains. Figure 16 shows that the high precipitation zone of the
HRB was mainly located in the mountainous areas below 39.5
In this paper, a hybrid data assimilation and machine learning framework
(DA-ML approach) was proposed and implemented into the Weather Research and
Forecasting (WRF) model to optimize the initialization of surface soil and
vegetation variables. Remotely sensed leaf area index (LAI) and multi-source
soil moisture (SM) observations (in situ SM profile observations and remotely sensed SM products) were integrated into the WRF model to improve the soil and vegetation characteristics. The performance of the WRF (DA-ML) framework was tested in the Heihe River basin (HRB) in northwestern China. The results indicated that the integration of remotely sensed LAI and multi-source SM into the WRF model improved the LAI, SM, and evapotranspiration (ET) estimates and regional climate and land–atmosphere interactions. The estimated sensible and latent heat fluxes from the WRF (OL) and WRF (DA-ML) models are validated with the large aperture scintillometer (LAS) observations at the Arou, Daman, and Sidaoqiao sites. For the WRF (DA-ML) approach, the three-site-averaged RMSDs of daily sensible and latent heat fluxes are 30.16 and 48.45 W m
Compared to the WRF model, the seasonal mean air temperature and specific
humidity simulated by the WRF (DA-ML) at the nine sites were closer to the
station measurements. For the WRF model, the nine-site-averaged root mean
square deviation (RMSD) of the air temperature and specific humidity estimates was 1.79 K and 1.08 g kg
The crops, shelterbelts, and residential areas in the midstream oasis produce a wind shield effect because of the stronger surface roughness. The different land surface dynamic and thermal characteristics between the oasis and desert can produce oasis–desert interactions and generate local circulation. In the lower atmosphere, wet and cold air masses are transferred to the surrounding desert by advection, while the dry and hot air over the desert is transferred to the oasis from the upper atmosphere. The results show that the integration of LAI and SM will induce water vapor intensification and promote precipitation in the upstream regions of the HRB. The WRF (DA-ML) simulation captured the temporal and spatial variability in precipitation well and was consistent with the reference data. The results indicate that the 3 km high-resolution grid can consider topographic information and produce accurate precipitation distribution estimates.
The computation details about the WRF (OL) and WRF (DA-ML).
The zonal mean vertical velocity and meridional circulation from the WRF (OL) and WRF (DA-ML) models during the growing season in 2015 in the downstream oasis. The orange bar represents the oasis area.
ERA5 data for the WRF model are freely available via the European Centre for Medium-Range Weather Forecasts
(
XH developed the model code and completed the draft paper, with support from all co-authors. YL, SL, TX, FC, ZL, RL, and LS revised the paper. ZZ provided the original WRF code. ZX, ZP, and CZ provided the methods for processing the observational data. All authors contributed to the synthesis of the results and key conclusions.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We would like to thank all the scientists, engineers, and students who participated in WATER and HiWATER field campaigns. The computation of the WRF model has been supported by sources from the Center for Geodata and Analysis, Faculty of Geographical Science, Beijing Normal University (
This research has been supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDA20100101) and the National Natural Science Foundation of China (grant no. 42171315).
This paper was edited by Bob Su and reviewed by four anonymous referees.