Search

• The selective EEMD-TFT model is introduced for predicting multi-step discharge. • Decomposing nonlinear and nonstationary data improves prediction performance. • The selective EEMD-TFT outperforms other seq-to-seq models in comparison. • It provides a more robust prediction and is less sensitive to forecast horizons. Karst groundwater is a critical freshwater resource for numerous regions worldwide. Monitoring and predicting karst spring discharge is essential for effective groundwater management and the preservation of karst ecosystems. However, the high heterogeneity and karstification pose significant challenges to physics-based models in providing robust predictions of karst spring discharge. In this study, an interpretable multi-step hybrid deep learning model called selective EEMD-TFT is proposed, which adaptively integrates temporal fusion transformers (TFT) with ensemble empirical mode decomposition (EEMD) for predicting karst spring discharge. The selective EEMD-TFT hybrid model leverages the strengths of both EEMD and TFT techniques to learn inherent patterns and temporal dynamics from nonlinear and nonstationary signals, eliminate redundant components, and emphasize useful characteristics of input variables, leading to the improvement of prediction performance and efficiency. It consists of two stages: in the first stage, the daily precipitation data is decomposed into multiple intrinsic mode functions using EEMD to extract valuable information from nonlinear and nonstationary signals. All decomposed components, temperature and categorical date features are then fed into the TFT model, which is an attention-based deep learning model that combines high-performance multi-horizon prediction and interpretable insights into temporal dynamics. The importance of input variables will be quantified and ranked. In the second stage, the decomposed precipitation components with high importance are selected to serve as the TFT model's input features along with temperature and categorical date variables for the final prediction. Results indicate that the selective EEMD-TFT model outperforms other sequence-to-sequence deep learning models, such as LSTM and single TFT models, delivering reliable and robust prediction performance. Notably, it maintains more consistent prediction performance at longer forecast horizons compared to other sequence-to-sequence models, highlighting its capacity to learn complex patterns from the input data and efficiently extract valuable information for karst spring prediction. An interpretable analysis of the selective EEMD-TFT model is conducted to gain insights into relationships among various hydrological processes and analyze temporal patterns. [ABSTRACT FROM AUTHOR]

• A hybrid framework is proposed for predicting karst spring discharge using historical data. • The hybrid framework achieves good prediction accuracy in non-stationary spring discharge sequences and has strong robustness against disturbances. • To the best of our knowledge, this study is the first to apply PMI and VMD to the prediction of spring discharge. • The time series decomposition technology has the greatest impact on improving prediction accuracy. Accurately predicting karst spring discharge is crucial for effective regional water resource management. However, the uneven distribution (or absence) of meteorological stations within karst spring catchments and the temporal discontinuity of meteorological data make the prediction of spring discharge difficult. To overcome a model's dependence on meteorological data, this paper proposes a novel hybrid framework that uses historical spring discharge as the input for the LSTM model. The framework consists of four components—hyperparameter optimization, variable screening, decomposition-combination, and error suppression—to improve prediction accuracy. Specifically, particle swarm optimization, genetic algorithms, and Bayesian optimization (BSO) were used to optimize the hyperparameters of LSTM. Partial correlation, grey correlation analysis, and partial mutual information (PMI) were adopted to analyze the correlations between historical and current discharge to determine the optimal input variables of the model. Seasonal and trend decomposition using loess, ensemble empirical mode decomposition, and variational mode decomposition (VMD) were used to extract multi-scale change features in the spring discharge sequence, which reduced the learning complexity of LSTM. Finally, the Autoregressive Moving Average (ARMA) model was used to correct the output error, ensuring prediction accuracy. In total, 57 models were established to evaluate the performance of the hybrid framework in three karst spring catchments with different hydrogeological conditions and strongly nonlinear discharge. The results show that the hybrid framework achieves excellent prediction performance for all three karst springs. In comparison to the five benchmark models that take meteorological data as the input, the hybrid framework outperforms the best benchmark models in two of the spring catchments. This study provides a high-quality alternative for predicting spring discharge in regions without meteorological data. [ABSTRACT FROM AUTHOR]

• Inconsistency between vegetation greenness and productivity occurred mainly after 2000. • Significant greening in humid areas has not led to a synchronous increase in productivity. • Prolonged drought significantly reduced vegetation productivity but promoted greening in humid zones. • Vegetation greenness and productivity responses to prolonged drought were opposite in humid and arid ecosystems. Recent studies have shown that climate stress may cause asynchronous changes between greenness and productivity, but the mechanism how greenness and productivity respond to drought remains unclear. The significant greening of vegetation in China has attracted worldwide attention; nevertheless, it is unclear whether large-scale greening will translate into productivity in the context of recent droughts. In this study, we used the vegetation index dataset to extract long-term trends in vegetation greenness and productivity from 1982 to 2018, and then chose the multi-scale Standardized Precipitation Evapotranspiration Index (SPEI) index to measure how vegetation greenness and productivity change in response to drought. The results show that significant vegetation greening did not cause a simultaneous increase in productivity during 1982–2018. Vegetation greenness and productivity trends were relatively consistent before 2000 and diverged after 2000, especially during the prolonged drought of 2000–2009. The opposite trend between greenness and productivity in humid ecosystems during drought periods mainly caused this trend difference. In humid ecosystems, short-term drought promotes vegetation greening, while long-term drought has no significant impact on greenness. However, the inhibitory effect of drought on productivity deepens with an increase in drought duration. Therefore, prolonged droughts offset vegetation greenness-induced gross primary productivity. We also found that prolonged drought causes a significant decrease in greenness but has little effect on productivity in arid ecosystems, possibly because vegetation productivity is more sensitive to soil water than atmospheric drought. Our results highlight the complexity of vegetation growth in the context of climate change. Furthermore, it is crucial to pay attention to the impact of climate stress on regional carbon sequestration functions, considering the significant carbon sink potential of forest areas in southern China. [ABSTRACT FROM AUTHOR]

Highlights • Improved the decomposition-ensemble framework of EEMD-based forecast models. • Proposed an adaptive forecast model using EEMD-ANN hybrid approach. • Compared the performance of proposed model with ANN, ANFIS, SVM and SAR (1). • Recommended to use the proposed model in wet season and SAR (1) model in dry season. Abstract It remains a challenge to obtain an accurate middle and long-term runoff forecast, especially in flood seasons. The forecast performance can be improved using ensemble empirical mode decomposition (EEMD) to produce cleaner signals as model inputs. In many EEMD based forecast models, the entire time series are decomposed into several sub-series, and each sub-series is divided into calibration and validation datasets and forecasted by some common models, such as artificial neural network (ANN), and finally an ensemble forecast is obtained by summing the forecasted results of each sub-series. In such a decomposition-ensemble framework, some future information is used, and thus it is a hindcast experiment. Attempts have also been made to propose a real forecast experiment, which, however, often cannot be able to adapt to the non-stationary changes in runoff due to the lack of adaptive ability. Therefore, this study tries to improve the decomposition-ensemble framework and propose an adaptive middle and long-term runoff forecast model especially for flood seasons. Unlike a hindcast experiment, it does not use any future information; and unlike conventional forecast experiments, its decomposition and forecast models are adjusted adaptively as long as new runoff information is added. EEMD is used to decompose the original time series and ANN is used to forecast each sub-series, hence the name an adaptive EEMD-ANN (AEEMD-ANN) model. This model is applied to forecast the 1-month ahead streamflow of three stations in China, and the results show that the AEEMD-ANN model can improve forecast accuracy in flood seasons, but it is not as good as ANN, adaptive neuro-fuzzy inference system (ANFIS), support vector machine (SVM) and seasonal first-order autoregressive (SAR (1)) models in dry seasons. While the conventional forecast models, especially the SAR (1) models, are more suitable in dry season. Therefore, it is recommended to use SAR (1) model in dry season and AEEMD-ANN model in flood season to forecast the monthly runoff in Yangtze River Basin. [ABSTRACT FROM AUTHOR]

Climate indices characterize climate systems and may identify important indicators for long-term precipitation, which are driven by climate interactions in atmosphere-ocean circulation. In this study, we investigated the climate indices that are effective indicators of long-term precipitation in South Korea, and examined their relationships based on statistical methods. Monthly total precipitation was collected from a total of 60 meteorological stations, and they were decomposed by ensemble empirical mode decomposition (EEMD) to identify the inherent oscillating patterns or cycles. Cross-correlation analysis and stepwise variable selection were employed to select the significant climate indices at each station. The climate indices that affect the monthly precipitation in South Korea were identified based on the selection frequencies of the selected indices at all stations. The NINO12 indices with four- and ten-month lags and AMO index with no lag were identified as indicators of monthly precipitation in South Korea. Moreover, they indicate meaningful physical information (e.g. periodic oscillations and long-term trend) inherent in the monthly precipitation. The NINO12 indices with four- and ten- month lags was a strong indicator representing periodic oscillations in monthly precipitation. In addition, the long-term trend of the monthly precipitation could be explained by the AMO index. A multiple linear regression model was constructed to investigate the influences of the identified climate indices on the prediction of monthly precipitation. Three identified climate indices successfully explained the monthly precipitation in the winter dry season. Compared to the monthly precipitation in coastal areas, the monthly precipitation in inland areas showed stronger correlation to the identified climate indices. [ABSTRACT FROM AUTHOR]

• The distribution of the nonlinearity of ecosystem functioning was revealed. • The stability of ecosystem improvement is probably higher. • Ecosystem stability presented bimodal pattern along elevation gradient. • Climatic mean state rather than climatic variability affected the ecosystem stability. Ecosystem water-use efficiency (WUE) is an effective indicator of ecosystem functioning, which is widely used to characterize the coupling relationship between terrestrial carbon and the water cycle. Although the linear trends in WUE have been well reported, the nonlinearity of the secular trends in WUE and their relative contribution to the overall changes were still poorly investigated. In this study, we used ensemble empirical mode decomposition (EEMD) to extract the secular trends of the WUE and their relative contribution (RC) to the WUE changes in China, which represents the stability of the ecosystem functioning. We found that although increasing trends dominated the secular trends of the WUE, nearly half of the study areas had experienced trend shifts. Except that deciduous forests and mixed forests were dominated by trend shifts, most of the land cover types were dominated by increasing trends. The RC exhibited a bimodal pattern along the elevation gradient, with peaks at 20 m and 5000 m. Similarly, the RC peaked at the east and south aspects, and was the lowest for the west and north aspects. In comparison, the RC exhibited a U-shaped pattern with the slope, and was the lowest at 80°. For the monotonically increasing trends and decreasing-to-increasing trends, the high RC mainly occurred on the North China Plain and Loess Plateau, indicating ecosystem improvement and strong stability. Meanwhile, it was low in northeastern China and southwestern China, indicating unsustainable ecosystem improvement. The RC was always very low for increasing-to-decreasing trends and monotonically decreasing trends, indicating low ecosystem stability with a high possibility for improvement. Temperature and the atmospheric carbon dioxide (CO 2) concentration were the two main climatic factors driving the WUE changes. The climatic mean state rather than the climatic variability affected the ecosystem stability. The results of this study highlight the relative contribution of the secular trends to WUE changes and provide a better understanding of the structure and the stability of ecosystem functioning. [ABSTRACT FROM AUTHOR]

• Propose a combination method of time–frequency algorithm and LSTM. • Improve the ability of LSTM handling strong nonlinear multi-frequency sequence. • The prediction performance of EEMD-LSTM is better than SSA-LSTM. Spring discharges from karst aquifers are results of spatially and temporally complex hydrologic processes, such as precipitation, surface runoff, infiltration, groundwater flow as well as anthropogenic factors. These processes are spatially and temporally varying at a multiplicity of scales with nonlinear and nonstationary characteristics. For improving the prediction accuracy of karst springs discharge, this study first applied the time–frequency analysis methods, including singular spectrum analysis (SSA) and ensemble empirical mode decomposition (EEMD) to extract frequency and trend feature of Niangziguan Springs discharge. Then the long short-term memory (LSTM) was used to simulate each frequency and trend subsequence. Subsequently, the prediction of spring discharge was completed by a combination of the simulated results from LSTM. Finally, the performances of LSTM, SSA-LSTM, and EEMD-LSTM under different inputs were compared. The results show that the performance of SSA-LSTM and EEMD-LSTM are better than LSTM, and the EEMD-LSTM model achieved the best prediction performance. [ABSTRACT FROM AUTHOR]

• Physical mechanisms and regression of streamflow series were assessed. • Study combined ensemble empirical mode decomposition (EEMD) with artificial neural network. • Teleconnection analysis using EEMD yielded more physical information about hydrology. • Climate indices more closely correlated with decomposed series than original inflow. Streamflow is the result of complex climatic and hydrological interactions that are driven by atmosphere–ocean circulation. Teleconnection analysis of streamflow is significant for identifying the atmosphere- and climate-related indices of hydrology series. To mine the physical information of streamflow series more effectively, ensemble empirical mode decomposition (EEMD) was employed to decompose streamflow series into inherent stationary components with different periodic oscillations and a trend. In this study, important climate indices were identified based on the correlation coefficients, which were calculated using cross-correlation. According to the selected indicators, the decomposed and streamflow series were regressed using an artificial neural network (ANN) and support vector regression (SVR). In total, 130 climate phenomenon indices, pre-runoff, and 1–12-month time lags were considered as teleconnection variables. The results show that EEMD can be used to extract the period and trend of streamflow, and that the decomposed hydrology components have much stronger correlations than the original runoff with the climate phenomenon indices. The original streamflow was the most closely correlated with the series from the previous month, which is an autocorrelation. However, more physical information was obtained through the teleconnection of the sub-streamflow series. The periodic oscillations was explained by relatively diverse atmospheric circulation and sea surface temperature indices with different time lags. The low-frequency periodic series were better represented than the high-frequency series by the climate phenomenon indices. In addition, Nino 3.4, Nino 4, the Warm-pool ENSO Index, and the ENSO Modoki Index were representative of the high-frequency component; the mid-frequency signals were sensitive to the Solar Flux Index, Total Sunspot Number Index, Pacific Decadal Oscillation Index, and Southern Oscillation Index; the low-frequency series were influenced by the Atlantic Multi-decadal Oscillation Index, North African Subtropical High Area Index, North Atlantic Triple Index, etc.; and the Indian Ocean Basin-Wide Index, Atlantic Multi-decadal Oscillation Index, Western Pacific Warm Pool Strength Index, and East Pacific 850 mb Trade Wind Index expressed well the long-term trend of monthly streamflow. Meanwhile, the regression results of the decomposed series obtained by the ANN and SVR exhibited better statistical performance than those of the original series, especially for the EEMD-SVR. [ABSTRACT FROM AUTHOR]