Search

• Surface water isoscapes (δ18O and d-excess) in China was firstly reproduced using Random Forest model. • The δ18O isoscape in river water demonstrates enrichment in the Western arid zone and depletion in the Tibetan Plateau. • Catchment-scale evapotranspiration and instream evaporative processes contribute to the enrichment of downstream river water isotopes. The availability of comprehensive stable isotope data in China is limited, which hinders a thorough understanding of interpreting runoff sources and land–atmosphere water fluxes on a national scale. In this study, we have undertaken the task of establishing a dataset of surface water isotopes (δ18O and δ2H) to create surface water isoscapes, and identify its controlling factors. Our analysis, utilizing a random forest model (RF), indicates that the isotopic patterns observed in precipitation are well-reflected in river water across China. Specifically, the δ18O isoscape in river demonstrates enrichment in the Western arid zone and depletion in the Tibetan Plateau. These patterns are strongly influenced by hydro-climatic factors such as relative humidity, precipitation, and catchment properties, such as elevation. Notably, elevation is a significant variable in the RF model, governing the isotopic composition (δ18O and d-excess) of rivers throughout China, primarily due to the rainout effect resulting in isotopically-depleted precipitation from lowlands to elevated mountain regions. In contrast, surface water d-excess isoscape reveals a more complex spatial variability in China, mainly associated with contrasting moisture sources including maritime vapor from tropical oceans and inland recycling vapor. In addition, secondary evaporation processes resulted from cumulative dams and developed irrigation systems also contribute to this variability. Hence, catchment-scale evapotranspiration and instream evaporative processes contribute to the enrichment of downstream river water isotopes. The predictive surface water isoscapes will help understand the impact of changes in the hydrological cycle on a larger scale and provide practical guidance for future monitoring efforts and isotopic simulations. [ABSTRACT FROM AUTHOR]

• Different types of multi-year droughts exert distinct impacts on the PR relationship. • Multi-year hydrological droughts have a greater impact on the PR relationship. • Multi-year drought exerts distinct impacts on the PR relationship in northern and southern China. • The representation parameters of the PR relationship can be simulated using drought intensity. Multi-year droughts cause severe water scarcity, ecosystem collapse, and significant socio-economic losses. Understanding the impact of these droughts on the precipitation-runoff (PR) relationship is crucial for managing water resources and mitigating drought-related damage. However, previous research focuses mainly on the influence of multi-year meteorological droughts on the PR relationship, neglecting multi-year hydrological droughts and their compound effects. Studies also generally overlook the speed and volume of precipitation-to-runoff conversion, considering only the quality of PR relationship. Furthermore, there is a lack of effective simulation methods using the intensity of multi-year droughts to understand the PR relationship. To address these gaps, this study defines multi-year compound droughts and employs three key parameters, namely, the determination of coefficient (R 2), linear regression slope, and runoff coefficient, to assess the impact of multi-year droughts on changes in the PR relationship's quality, precipitation-to-runoff conversion speed, and runoff volume. A random forest model is then employed to simulate these parameters based on the intensity of multi-year droughts. Case studies conducted in three basins within the arid-semiarid region of northern China as well as in three basins in the humid region of southern China reveal that multi-year droughts result in water shortages and have a significant impact on the PR relationship. Among different types of multi-year droughts, hydrological droughts have the greatest impact on the PR relationship, followed by compound and meteorological droughts, respectively. During meteorological, hydrological, and compound drought periods, the average rainfall (runoff) is lower by 26 – 30% (33 – 58%), 18 – 28% (36 – 60%), and 18 – 26% (30 – 58%) respectively compared to non-drought periods. Furthermore, the PR relationship remains stable in the rivers of the humid region in southern China. In contrast, the Loess Plateau region in northern China exhibits a substantial decrease in both the speed and volume of precipitation-to-runoff conversion. Interestingly, this decline is not accompanied by a significant increase in the R 2 of the PR relationship. The random forest approach effectively simulates key parameters of the PR relationship based on the intensity of multi-year droughts. These findings hold crucial implications for the long-term regional water resource planning. [ABSTRACT FROM AUTHOR]

• A new method for improving river medium–high flow estimation using Sentinel-2 was proposed. • Sentinel-2 imagery and Random Forest allow early determination of river flow classes (e.g., low, medium, and high flow). • Pixels closely related to variations in river discharge show significant movement, which is highly similar to the trend of river boundary diffusion. • Identifying multiple characterizing wet pixels by flow classification can significantly improve medium–high flow estimation. River discharge monitoring is one of the most important parts of water resources management. However, due to hydrological stations' high operating and maintenance costs, discharge monitoring data has gradually become scarce. Remote sensing technology has great potential to supplement observations in this field. The close relationship between the reflectance ratio of dry (C) and wet (M) pixels around the river in the near-infrared (NIR) band and river discharge has been effectively proved, and many CM-based methods have also achieved good results, but often underestimate medium–high flow in rivers. However, medium–high flow usually occurs during the rainy season or flood-prone periods, and their underestimation not only affects the risk assessment of possible or already occurred floods but may also cause damage to people's lives, property, and social assets. To improve this phenomenon, we proposed a new CM method called CM Hierarchical Classification (CMHC) based on the Sentinel-2 dataset and Google Earth Engine (GEE) cloud platform to monitor the river discharge at different classes and achieve high-precision discharge estimation at medium–high flow. The study was conducted at four stations along the Po River in Italy, where measured discharge data were used to derive the calibration curve and evaluate the performance of the estimated results. The Nash-Sutcliffe efficiency (NS) values for the estimates reached an average of 0.84 and up to 0.89, indicating the excellent performance of the method. Compared with existing methods, our method substantially improves the monitoring accuracy of medium–high flow of the river. Furthermore, additional experiments in the Han River in China also confirmed the applicability of the CMHC method, implying that this method can be used to improve the ability of Sentinel-2 to estimate river discharge, especially for medium–high flow, and provide support for river discharge monitoring on a global scale. [ABSTRACT FROM AUTHOR]

[Display omitted] • A random forest model refined by the beluga whale optimization was established. • A new random forest model was constructed for flood disaster risk assessment. • The new model has good applicability in flood disaster risk assessment. To scientifically characterize regional flood disaster risks, a comprehensive regional flood disaster risk assessment index system was constructed based on the R cluster-factor analysis method. To improve the regional flood disaster risk identification ability, a random forest model refined by the beluga whale optimization (BWO-RF) was proposed. It was applied to the flood disaster risk assessment of the Jiansanjiang Branch of the Beidahuang Agricultural Reclamation Group in Heilongjiang Province, China. The research results show that the constructed BWO-RF model has strong generalization ability and optimization efficiency. The evaluation results have high rationality and stability, which fully demonstrates the applicability of the BWO-RF model in flood disaster risk assessment. On a time scale, the flood risk of the Jiansanjiang Branch shows a clear downward trend. However, due to weak economic development and insufficient water conservancy and flood control planning and construction, the risk of flood disasters fluctuates greatly in the later period, and the prevention and control situation is still not optimistic. On a spatial scale, the risk of flood disasters on farms in the northeast of Jiansanjiang Branch is low, while the risk on farms in the southwest is higher. Future flood control and drainage should focus on this area. The research results provide a technical method system for reference in flood disaster risk assessment and provide valuable information for reducing regional flood disaster risks. [ABSTRACT FROM AUTHOR]

• A spatial random forest-based downscaling and merging method (SRF-DM) is proposed. • SRF-DM performs better than the original satellite-based precipitation products. • SRF-DM is more accurate than the classical machine learning-based methods. • Spatial autocorrelation is the most important variable. • SRF-DM has a good performance even when the gauge density is low. Publicly available satellite precipitation products (SPPs) are essential inputs for hydrological models. However, the existing SPPs suffer from coarse resolutions and large uncertainties, which greatly hinder their widespread applications. Thus, to fully utilize the unique characteristics of each individual SPP, a spatial random forest (SRF)-based downscaling and merging (SRF-DM) method is proposed to fuse multiple SPPs and gauge observations in this paper. The proposed method incorporates spatial autocorrelation information between precipitation observations through a covariate. This covariate is estimated using ordinary kriging interpolation on spatial neighbors that have a high correlation with the target location. The proposed method was used to generate 1-km daily precipitation data using the datasets from five satellite precipitation products (IMERG, GSMaP, CHIRPS, MORPH, and PERSIANN) and gauge observations in the period from January 1, 2015 to December 31, 2019 over Sichuan province, China. The performance of SRF-DM was compared with the original SPPs, XGBoost-based downscaling and merging (XGB-DM), two double-layer machine learning (ML) methods, namely RF-RF and RF-artificial neural network (RF-ANN), and three single ML methods, namely RF, ANN, and RF-based Merging Procedure (RF-MEP). The experimental results demonstrate the following findings: (i) SRF-DM is more accurate than the original SPPs across different temporal scales (i.e., daily, monthly, and seasonal) in terms of statistical and categorical accuracy measures, (ii) SRF-DM outperforms the six ML-based methods in estimating high precipitation and capturing spatial precipitation patterns, (iii) the spatial autocorrelation between neighboring precipitation data is the most significant factor in both downscaling and merging, and (iv) the performance of SRF-DM is influenced by gauge density, but it still yields good results even with a limited number of gauges. On the whole, SRF-DM is considered an effective tool for generating daily precipitation products with high accuracy and resolution, which is primarily attributed to its satisfactory performance in regions with diverse precipitation patterns and topography. [ABSTRACT FROM AUTHOR]

• Soil moisture products trained by in-situ observations based on machine learning are better than their corresponding raw gridded products, mainly in southeast China. • Merging three coarse-resolution SM datasets together show better results, but it still fails to outperform individual high-resolution fusion product. • Merging three coarse-resolution products and the high-resolution product CSSPv2 further increases. While machine learning (ML) is used to correct gridded soil moisture (SM) products by fusing in-situ observations, the contribution of land surface model (LSM)- and satellite-based SM products are yet to be validated over a large area, leading to imprudent adoption of SM product(s) for data fusion. In this study, single or multiple SM products from CSSPv2 LSM simulation, ERA5 and GLDASv2.1 reanalysis, and ESA-CCI satellite data with different resolutions are used to train ML models and generate daily SM estimates at 0.0625° resolution with in-situ measurements as target and relevant variables as auxiliary. Three widely used ML methods, namely Random Forest (RF), LightGBM, and XGBoost, were compared. Validations over independent in-situ stations during 2012–2017 showed the improvement of fusion products against their corresponding raw products, with KGE and CC increased by 87 % and 6 %, and RMSE decreased by 22 % for SM at surface and rootzone layers. Regionally, ML-based SM estimates improve mainly in southeast China. Merging three coarse-resolution SM datasets (i.e., ERA5, GLDASv2.1 and ESA-CCI) together with in-situ observations further increases KGE and CC by 15 % and 5 % against individual fusion products, but it still fails to outperform the individual high-resolution fusion product of ML/CSSPv2. Merging all four gridded SM products with in-situ data shows advantage against the ML/CSSPv2, with KGE and CC increased by 9 % and 7 %. The results are consistent by using different ML methods. This study suggests the importance of high-resolution LSM for SM data fusion, even with the emergence of ML approaches. [ABSTRACT FROM AUTHOR]

River flow prediction plays a crucial role in flood control and water resource management. As an effective approach, the stepwise cluster analysis (SCA) has been well-accepted in river flow prediction but still encounters difficulties to acquire accurate predictions in sample-sparse regimes. This study proposes a new ensemble model by incorporating the bagging method with the SCA. The proposed Bagged Stepwise Cluster Analysis (BSCA) is applied to predict daily river flow in the Upper Min River, China. The results show that BSCA effectively improves the prediction accuracy of SCA. Particularly, it addresses the limited end clusters problem of SCA and thus enhances the extreme-value prediction. It also has a better overall prediction performance over the classical random forest and decision tree models. In addition, BSCA is capable of quantifying uncertainties through a probabilistic prediction, which can be used to assess risks of extreme flows and provide reliable river flow predictions. The proposed new model can be applied to other water systems and environmental engineering problems. • A new ensemble model (BSCA) is first proposed for river flow prediction. • The model effectively improves the prediction accuracy of SCA. • Probabilistic prediction is integrated into BSCA to quantify uncertainties. • The model provides a way to assess risks of extreme flows. • The BSCA can help flood control and water resource management. [ABSTRACT FROM AUTHOR]

• Nine state-of-the-art Chla algorithms for inland waters were evaluated. • Sensitivity of Chla algorithms to AC was quantified for the first time. • An RFR algorithm insensitive to AC uncertainty was developed for turbid lakes. • OLCI Chla products showed most lakes in eastern China are eutrophic. A robust and reliable chlorophyll-a (Chla) concentration algorithm is still lacking for optically complex waters due to the lack of understanding of the bio-optical process. Machine learning approaches, which excel at detecting potential complex nonlinear relationships, provide opportunities to estimate Chla accurately for optically complex waters. However, the uncertainties in atmospheric correction (AC) may be amplified in different Chla algorithms. Here, we aim to select one state-of-the-art algorithm or establish a new algorithm based on machine learning approaches that less sensitive to AC uncertainties. Firstly, nine state-of-the-art empirical, semianalytical, and optical water types (OWT) classification-based Chla algorithms were implemented. These existing algorithms showed good performance by using in situ database, however, failed in actual OLCI applications due to their sensitivity to AC uncertainties. Thus, four popular machine learning approaches (random forest regression (RFR), extreme gradient boosting (XGBoost), deep neural network (DNN), and support vector regression (SVR)) were then employed. Among them, the "RFR-Chla" model performed the best and showed less sensitivity to AC uncertainties. Finally, the Chla spatiotemporal variations in 163 major lakes across eastern China were mapped from OLCI between May 2016 and April 2020 using the proposed RFR-Chla model. Generally, the lakes in eastern China are severely eutrophic, with an average Chla concentration of 33.39 ± 6.95 μg/L. Spatially, Chla in the south of eastern China was significantly higher than those in northern lakes. Seasonally, Chla was high in the summer and autumn and low in the spring and winter. This study provides a reference for water quality monitoring in turbid inland waters suffering certain AC uncertainties and supports aquatic management and SDG 6 reporting. [ABSTRACT FROM AUTHOR]

• Large-scale vegetation restoration is ongoing in the karst region of China. • Using random forests, we assessed vegetation restoration alone impact on soil moisture (SM). • Vegetation restoration alone resulted in a slight increase in average SM, which led to decreased SM variation coefficient. • Vegetation restoration effectively stabilizes soil moisture. • The reduction of soil moisture due to warming is offset by vegetation restoration. Large-scale vegetation restoration projects have been ongoing in the karst region of southwest China for several years. However, the influence of vegetation restoration alone on soil moisture (SM) changes in this region remains unclear. In this study, a random forest model was constructed using meteorological data, the ERA5-Land SM dataset, and the enhanced vegetation index to analyze SM changes in the karst region of southwest China caused by vegetation restoration alone. The results showed that on a monthly scale, vegetation restoration introduced an increase in SM that was mainly concentrated in January–April and a decrease that was mainly concentrated in June–August. On a seasonal scale, vegetation changes caused an increase in mean SM in spring and winter and a decrease in mean SM in summer and autumn. The largest increase in SM owing to vegetation changes was in spring, with an average increase of 0.019 m3/m3, and the largest decrease was in summer, with an average decrease of 0.010 m3/m3. However, the annual average SM increased only slightly. On the seasonal and annual scales, the coefficient of variation values for changes in SM caused by vegetation restoration alone were dominated by decreases, with a large decrease from January to April. In addition, vegetation restoration effectively moderated the drying tendency of soil caused by climate warming. [ABSTRACT FROM AUTHOR]

[Display omitted] • We developed a hybrid approach integrating physical-based flood models with random forest for flood susceptibility assessment. • We identified the relatively significant flood-inducing factors. • The geographical and hydromorphological factors had greater effects on riverine flooding. • This flood susceptibility assessment is suitable in ungauged areas. Flooding is one of the most destructive natural hazards that has caused catastrophic effects worldwide. Recently, machine learning methods have become widespread in flood susceptibility assessments due to their efficiency and robustness. However, flood inventory data acquired from historical flooding records and field surveys in existing studies are sparse and not enough to develop accurate machine learning models. Thus, we developed a hybrid approach that integrates hydrodynamic model (HEC-HMS/RAS), rapid flood model (height above the nearest drainage, HAND) and machine learning model (random forest, RF) for flood susceptibility assessment. The proposed approach was then implemented in the Xinluo watershed, a flood-prone mountainous catchment in China. The results showed that the spatially continuous flood inventory map generated from HEC-HMS/RAS model satisfies the big data requirements of the RF model, and the proposed hybrid approach demonstrates the effectiveness and efficiency in flood susceptibility assessment with overall accuracy and area under the ROC curve values of 0.915 and 0.912, respectively. The geographical and hydromorphological factors had greater effects on riverine flooding with average contribution of 27.8% and 27.6%. In comparison, the average contribution of land cover and meteorological factors were 24.3% and 20.3%, respectively. The method proposed in this study is effective in assessing and forecasting flood susceptibility and can facilitate resilience planning as well as integrated flood management. [ABSTRACT FROM AUTHOR]

• Machine learning provides a feasible framework for estimating actual evapotranspiration at weather stations without flux observations. • Seasonal ET simulations performed better than that of daily, 8-day and monthly using pooled data from all flux towers. • The predictors with high contribution to the model are Tmax, Tmin, Tmean, Prcp, SWI, Rn, NDVI, LAI, FPAR, SWC and Tsoil. • Changes in the importance ranking and value of predictor variables partly explained the variations of model performance. The scarcity of site-scale actual evapotranspiration (ET) measurements poses a challenge for modelling and verifying regional ET. Taking the grassland ecosystem in arid and semi-arid regions of Northern China (ASNC) as an example, we proposed a framework for estimating ET at weather stations without flux observations through a machine learning approach by combining data from MODIS and fifteen flux towers distributed in alpine grassland (AG) and temperate grassland (TG) across ASNC. First, we analyzed the temporal characteristics of grassland ET at site scale. Second, we applied the machine learning approach Random Forest (RF) to develop a robust model for simulating site-scale grassland ET, and used random and spatial cross-validations (CVs) and three metrics to evaluate the RF models performance. Two strategies (using pooled data from all flux towers for a general model, and dividing data by grassland type to develop models specific to AG and TG) and four temporal resolutions (daily, 8-day, monthly and seasonal) were used to develop the RF models. Third, we investigated how the importance of predictor variables for estimating grassland ET changed in different CVs, strategies and temporal resolutions. The ET of AG and TG showed similar dynamic patterns but differed in magnitude. The RF models showed good performance in both strategies and all four temporal resolutions (R2 > 0.64, MAE < 0.53 mm d−1, RMSE < 0.72 mm d−1). However, seasonal ET simulations performed better than that of daily, 8-day and monthly using pooled data from all flux towers, especially in spatial CV. Meteorological variables (temperature, precipitation and radiation), vegetation (NDVI, LAI and FPAR) and soil (soil water content and soil temperature) were strong predictors of variation in grassland ET. Changes in the importance ranking and value of predictor variables partly explained the variations of model performance. The robust model performance in spatial CV proved that the framework developed in this study was reliable when applied to weather stations without flux observations, thereby overcoming the scarcity of site-scale actual ET measurements. [ABSTRACT FROM AUTHOR]

• We generated the annual IMERG at a 0.01° resolution using GBDT from 2015 to 2018. • GBDT is robust to simulate and downscale the IMERG over the areas with complex topography. • GBDT improves the spatial resolution without reducing the accuracy over Mainland China. Remote sensing brings unprecedented opportunities to estimate precipitation at regional, continental, and even global scales. Nevertheless, the spatial resolutions of current mainstream satellite precipitation estimates are still too coarse to be directly used in the hydrological and meteorological applications over the medium and small scales. The spatio-temporal distribution of precipitation is often affected by various land surface factors, such as geographical location, vegetational cover, topographic characteristics, and ground temperature. Based on the relationship between precipitation and multi-geospatial factors, this study proposed a new spatial downscaling approach named as gradient boosting decision tree (GBDT) to downscale the annual satellite precipitation estimates of Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) from 0.1° to 0.01° gridded resolution over Mainland China from 2015 to 2018. As for the preprocessing of downscaling algorithm, the entire study area of Mainland China was separated into four different climate regions due to the spatial discrepancy of the relationship between precipitation and geospatial factors. To validate the effectiveness of GBDT approach, we compared it with other two types of downscaling methods based on random forests (RF) and support vector machine (SVM), respectively. Our evaluation results show that the geographical location (including both latitude and longitude) seems to be relatively more important and stable for GBDT modelling than other land surface factors across the four climate regions in Mainland China. In terms of large spatial scales, the GBDT and RF algorithms generally are superior to SVM for downscaling the IMERG precipitation, as SVM reduced the considerable accuracy of the downscaled results. Last but not least, GBDT exhibits more robust features in downscaling the satellite precipitation retrievals than RF over complex terrains, where the amount of precipitation has strong spatial heterogeneity. [ABSTRACT FROM AUTHOR]

• A new method for meteorological interpolation was propose by a decomposing way. • RF and IDW was coupled in predicting meteorological surfaces with covariates. • RF-IDW well depicts the relationship between meteorological value and covariates. • RF-IDW have good performance in the complex region. Spatially interpolated temperature and precipitation are hydrological variables that are widely applied in models of ecology, hydrology, agronomy, and other environmental sciences. The accuracy of meteorological data directly determine model performance. However, the interpolation of temperature and precipitation over complex areas is affected by the value observed with surrounding meteorological stations and also by aspects of the physical environment, such as terrain and land cover effects. These effects significantly limit the precise interpolation of temperature and precipitation. To normalize these effects, a new method coupling random forest (RF) and inverse distance weighting (IDW) was proposed, RF-IDW. The proposed method was applied to generate the climate surfaces of both precipitation and temperature in China in 2015. During the interpolation with RF-IDW, the temperature and precipitation were decomposed into a basic part (indicating the atmospheric properties) and local parts (representing the effects of the physical environment). The complex correlations were determined between precipitation/temperature and multiple-covariates of landcover, normalized difference vegetation index (NDVI), elevation, slope, longitude, soil moisture, latitude and Land Surface Temperature (LST) derived from Moderate Resolution Imaging Spectroradiometer (MODIS) data. Next, 10-fold and leave-one-region-out cross-validation were applied to verify the accuracy of RF-IDW for the interpolation of meteorological values. We also compared the RF-IDW performance to that of RF, IDW, Anusplin, and MLR to further investigate its feasibility. Based on 10-fold cross-validation, the coefficient of determinations (R2) between interpolated value and observed values were 0.962 for precipitation and 0.990 for temperature, indicating that the RF-IDW can accurately interpolate climate surfaces. The performance comparison showed that the RF-IDW reduced the mean absolute error by 55 mm to 120 mm for precipitation interpolation and 0.06 ℃ to 0.8 ℃ for temperature interpolation, especially in the complex region. The result of monthly interpolations implied the good robustness of the RF-IDW. Overall, the proposed RF-IDW approach allows improved spatial interpolation of complex region and suggest alternate methods for the spatial interpolation of hydrological variables. [ABSTRACT FROM AUTHOR]

• A comprehensive soil moisture downscaling methods comparison scheme was used. • Machine learning methods show high stability but provide smoother spatial patterns. • Residual interpolation can improve the accuracy of the downscaled soil moisture. • High ability on preserving the feature of microwave product could be a disadvantage. • All five downscaled soil moisture results show a declining accuracy. Microwave remote sensing is able to retrieve soil moisture (SM) at an adequate level of accuracy. However, these microwave remotely sensed SM products usually have a spatial resolution of tens of kilometers which cannot satisfy the requirements of fine to medium scale applications such as agricultural irrigation and local water resource management. Several SM downscaling methods have been proposed to solve this mismatch by downscaling the coarse-scale SM to fine-scale (several kilometers or hundreds of meters). Although studies have been conducted over different climatic zones and from different data sets with good results, there is still a lack of a comprehensive comparison and evaluation between them to guide the production of high-resolution and high-accuracy SM data. Therefore, in this study we compared several SM downscaling methods (from 0.25° to 0.01°) based on polynormal fitting, physical model, machine learning and geostatistics over the Qinghai-Tibet plateau where there is a wide range of climate conditions from four aspects, that is, comparison with the original microwave product, comparison with in situ measurements, inter-comparison based on three-cornered hat (TCH) method, and a spatial feasibility analysis. The comparison results show that the method based on a physical model, in this case the Disaggregation based on Physical And Theoretical scale Change (DisPATCh) method, has the highest ability on preserving the coarse-scale feature of original microwave SM product, while to some extent, this ability could be a disadvantage for improving the accuracy of the downscaling results. In addition, soil evaporation efficiency (SEE) alone is not sufficient to represent SM spatial patterns over complex land surface. Geostatistics based area-to-area regression Kriging (ATARK) introduces the highest uncertainty caused by the overcorrection during the residual interpolation process while this process can also improve correlation (R) and correct the bias as well as provide more feasible spatial patterns and details. Two machine learning methods, the random forest (RF) and Gaussian process regression (GPR) show high stability on all comparison results but provide smoother spatial patterns. The multivariate statistical regression (MSR) method performs worst due to the fact that its simple linear regression model could not meet the requirement of SM fitting on complicated land surface. Moreover, all five downscaling methods show a declining accuracy after downscaling. This phenomenon may be caused by the spatial mismatch on fine-scale. In addition, this could also be caused by the tendency that downscaled results will usually provide more spatial details from downscaling predictors, while they cannot capture the temporal changes of the microwave SM product well. In general, this phenomenon tends to be more significant over heterogeneous land surface. All in all, five widely used soil moisture downscaling methods were compared based on a comprehensive comparison scheme to add to the body of knowledge in applicability of downcaling methods under different weather conditions. [ABSTRACT FROM AUTHOR]

• Establishes the random forest classification model for identifying effective reservoirs. • Different floods are employed as training samples to explore the model stability. • The expected total cost with cross-validation is adopted to evaluate the classification accuracy. • The average increase of expected total cost with cross-validation is proposed to assess criteria importance. In the real-time operation of a flood control system, identifying effective reservoirs accurately and adaptively is the premise of establishing a multi-reservoir real-time flood control hybrid operation (MRFCHO) model. The existing effective reservoir intelligent reasoning (ERIR) method is based on the inference rules, which causes the identification results to be greatly influenced by the training samples. This paper establishes a random forest classification (RFC) model for identifying effective reservoirs to solve this problem. The performance of the RFC model is evaluated in terms of model stability and model accuracy, and compared with the ERIR model and other machine learning (ML) models. First, different flood samples are used to establish the models to verify the model stability. Then, the expected total cost under cross-validation is taken as the index to evaluate the classification accuracy. Finally, the average increase of expected total cost with cross-validation is proposed to evaluate the criteria importance and analyze the sensitive factors that affect the classification accuracy of the RFC model. The proposed method is applied to a multi-reservoir system in the Huaihe River basin in China. The results indicate that the RFC model has the characteristic of high classification accuracy, low sensitivity to flood samples and high stability. It displays more dominance in the dynamic identification of effective reservoirs compared to other models. [ABSTRACT FROM AUTHOR]

• Deep learning models have excellent performance for estimating ETo beyond study areas. • Temporal convolution neural network outperformed markedly empirical equations. • T-test method was used to test the performance of proposed models. • Temporal convolution neural network outperformed classical machine learning models. To evaluate the performance of deep learning methods (DL) for reference evapotranspiration estimation and to assess the applicability of the developed DL models beyond the study areas where they were trained, three popular DL models named deep neural network (DNN), temporal convolution neural network (TCN), and long short-term memory neural network (LSTM) were developed to estimate daily reference evapotranspiration (ET o) using incomplete meteorological data in the Northeast plain, China. The performances of the three DL models were compared to two classical machine learning models (CML)—support vector machine (SVM) and random forest (RF)—and empirical equations, including two temperature-based (Hargreaves (H) and modified Hargreaves (MH)), three radiation-based (Ritchie (R), Priestley-Talor (P), and Makkink (M)), and two humidity-based (Romanenko (ROM) and Schendel (S)) empirical models, in two strategies: (1) all proposed models were trained, tested, and compared in each single weather station, and (2) all-weather stations were split into several groups using the K-means method with their mean climatic characteristics. Then, in each group, stations took turns testing the proposed models which were trained by rest of the stations. The results showed that (1) the coefficient of determination (R2) values of the TCN and RF were 0.048 and 0.035 significantly higher than that of MH, respectively, and the relative root mean error (RMSE) values of TCN and RF were substantially 0.096, and 0.074 mm/d lower than that of MH, indicating that TCN and RF performed better than empirical models in the first strategy, and TCN and LSTM exhibited an RMSE that was significantly decreased by 0.069 and 0.079 mm/d, showing that TCN and LSTM outperformed empirical models in the second strategy, compared with the MH method; (2) in both strategies, compared with the Ritchie (R) model, TCN, LSTM, DNN, RF, and SVM increased R2 and decreased RMSE significantly, especially the TCN model; (3) similarly, TCN, LSTM, DNN, RF, and SVM models all augmented R2 and reduced RMSE substantially in comparison to humidity-based empirical models in both strategies, especially the TCN model. Overall, when temperature-based features were available, the TCN and LSTM models performed markedly better than temperature-based empirical models beyond the study areas, and when radiation-based or humidity-based features were available, all of the proposed DL and CML models outperformed radiation-based or humidity-based empirical equations beyond the study areas in which they were trained. [ABSTRACT FROM AUTHOR]

• A random forest (RF)-based model to extend GRACE TWSA back to1948 over China. • A spatially moving-window structure is combined with the RF model. • The RF TWSA match well with GRACE results and another reconstructed GRACE data. • Spatial patterns of the variable importance in RF model are quantified. The changes of terrestrial water storage (TWS) is critical for drought monitoring, water and food security, global water cycle and climate change studies. Currently, the Gravity Recovery and Climate Experiment (GRACE) twin satellites are unique means of observing large-scale water storage variations, but the short time series (2002-present) limits their applications in long term climatic and hydrologic studies. Although the TWS can be calculated from global land surface models, large uncertainties arise due to uncertainties of inputs and the limitations of the models. This study developed a reconstruction model for GRACE TWS anomalies (TWSA) based on the Global Land Data Assimilation System (GLDAS) model outputs by using a Random Forest (RF) regression approach. A Spatially Moving Window (SMW) structure was introduced when training the RF model to address the spatial variations of TWSA, and a linear regression approach (LR) was also used for comparison purpose. Long-term TWSA over China land area were generated based on the proposed approaches and results were validated through cross-validation and comparisons with reference datasets. As a result, the RF-based model outperforms the LR-based model, and the reconstructed TWSA by using the two models both well reproduce GRACE dataset and outperform the TWSA that are derived directly from GLDAS models. Moreover, the TWSA produced by using the presented models have good agreements with another global GRACE-based reconstructed TWSA dataset and in-situ soil moisture measurements. Importance value for each variable in the RF model was quantified as well as the spatial coefficients for each variable in the LR model. The importance values and regression coefficients present varying spatial patterns. Rather than modifying the land surface model structure and inputs, this study provides alternative ways of improving the TWS estimations of GLDAS and extending time range of GRACE datasets. The experiments are expected to promote and enrich the methodologies and theories of combining physical and statistical models for optimal simulations in geoscientific research. [ABSTRACT FROM AUTHOR]