Assessing trends and forecasting meteorological drought in South Africa using Savitzky-Golay enhanced hybrid deep learning.

Assessing trends and forecasting meteorological drought in South Africa using Savitzky–Golay enhanced hybrid deep learning - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 10;16:16968. doi: 10.1038/s41598-026-46664-x Assessing trends and forecasting meteorological drought in South Africa using Savitzky–Golay enhanced hybrid deep learning Siphamandla Sibiya 1, 3, ✉ , Shaun Ramroop 1, # , Sileshi Melesse 1, # , Nkanyiso Mbatha 2, # Author information Article notes Copyright and License information 1 School of Mathematics, Statistics, and Computer Science, University of KwaZulu-Natal, Pietermaritzburg Campus, Private Bag X01, Pietermaritzburg, 3209 South Africa 2 Council for Scientific and Industrial Research, Holistic Climate Change, Smart Places, Mering Naude Road, Brummeria, Pretoria, 0001 South Africa 3 Pure and Applied Analytics, Faculty of Natural and Agricultural Sciences, North-West University, Private Bag X1290, Potchefstroom, 2520 South Africa ✉ Corresponding author. # Contributed equally. Received 2025 Nov 24; Accepted 2026 Mar 26; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13230733  PMID: 41963462 Abstract Drought constitutes one of the most significant natural hazards worldwide, exacerbated by climate variability and change, with profound implications for ecosystems, agriculture, and livelihoods. In South Africa, particularly within the drought-prone uMkhanyakude District of KwaZulu-Natal, comprehending rainfall variability and enhancing drought prediction are imperative for sustainable water and food security planning. This study utilized daily rainfall records from six meteorological stations spanning the years 1980 to 2023 to calculate the Standardized Precipitation Index (SPI) at 6-, 9-, and 12-month time scales. Long-term drought trends were evaluated employing Innovative Trend Analysis (ITA) methods, which identified statistically significant decreasing trends at five stations and an increasing trend at Riverview. To augment drought forecasting, a novel hybrid model that integrates the Savitzky–Golay filter with a Temporal Convolutional Network and Long Short-Term Memory (SG–TCN–LSTM) was developed. Comparative assessments against ARIMA, LSTM, TCN, and other hybrid models demonstrated that the SG–TCN–LSTM consistently achieved the lowest Root Mean Square Error (RMSE) values (0.0349–0.1453) and the highest values (0.95–0.99) across all SPI scales, indicating superior predictive accuracy and stability. The integration of signal smoothing with deep learning methodologies enhanced the robustness of forecasts, providing critical insights for proactive drought risk management. This research underscores the potential of hybrid models as reliable early-warning instruments for meteorological drought and establishes a framework that can inform national adaptation strategies. Future research should aim to extend the model to incorporate additional climatic drivers, evaluate its transferability across diverse climatic regions, and investigate its application in operational drought early-warning systems. Keywords: Drought prediction, Standardized precipitation index (SPI), Meteorological drought, Savitzky–Golay (SG) filter, Hybrid models, Innovative trend analysis (ITA) Subject terms: Climate sciences, Environmental sciences, Hydrology, Natural hazards, Water resources Introduction Global warming has increased the frequency and severity of extreme hydroclimatic events, posing significant risks to ecosystems, agriculture, water resources, and human livelihoods 1 – 3 . Drought is particularly complex due to its gradual onset, extended duration, and socio-economic impacts 4 . Meteorological drought, marked by sustained precipitation deficits, often precedes agricultural and hydrological drought, making it a critical early indicator of water scarcity and food insecurity 5 . The assessment of meteorological drought relies on standardized indices, notably the Standardized Precipitation Index (SPI) 6 , the Standardized Precipitation Evapotranspiration Index (SPEI) 7 , and the Palmer Drought Severity Index (PDSI) 8 . Recent studies highlight the SPEI’s relevance for assessing climate change-driven drought 9 , 10 , especially where temperature impacts are significant. However, the SPI is the only index endorsed by the World Meteorological Organization for monitoring meteorological drought due to its robustness, minimal data requirements, and comparability across climatic regimes 9 , 11 – 15 . Its multiscalar flexibility makes it suitable for medium- to long-term drought assessments. This study uses the SPI at 6-, 9-, and 12-month time scales to capture seasonal and interannual rainfall variability. South Africa exhibits significant vulnerability to climate variability, attributable to its semi-arid climate, reliance on rain-fed agriculture, and insufficient water storage infrastructure 14 , 16 , 17 . Recent drought episodes, notably the 2015–2016 El Niño event, resulted in considerable agricultural and socio-economic detriments across Southern Africa 16 , 18 , 19 . In this context, the uMkhanyakude District of KwaZulu-Natal serves as a particularly climate-sensitive area, characterized by its dependence on groundwater, subsistence farming practices, ecological assets such as the iSimangaliso Wetland Park, and elevated levels of socio-economic vulnerability 20 . Notwithstanding this vulnerability, research addressing drought in this district remains scarce, particularly concerning the development of advanced forecasting frameworks that are tailored to local hydro-climatic conditions. Traditional drought forecasting methodologies, including statistical and regression-based approaches, often struggle to accurately capture the nonlinear dynamics and temporal dependencies of hydroclimatic processes. Consequently, there has been increased adoption of soft computing and machine learning techniques in this field. Hybrid predictive models that integrate signal decomposition, optimization algorithms, and deep learning architectures have been widely used to address various environmental challenges, showing improved performance compared to standalone models in drought prediction across different climatic regions 21 – 26 . Malik et al. 27 demonstrated the effectiveness of hybrid optimization-based models for meteorological drought forecasting using the Effective Drought Index (EDI), showing that SVM-HHO outperformed SVM-PSO in predicting median drought conditions in Uttarakhand, India. Similarly, Piri et al. 28 conducted a comparative analysis of SVM-RSM, SVM-PSO, standalone SVM, and ANN across multiple drought indices, concluding that SVM-PSO achieved superior accuracy and trend detection, while SVM-RSM exhibited the highest operational efficiency. Advancements in deep learning highlight the importance of hybridization 15 , 29 , 30 . Waqas et al. 31 found that wavelet-based hybrid models, especially the WBi-LSTM, outperformed standalone Bayesian Regression Forest (BRF) and Bidirectional Long Short-Term Memory (Bi-LSTM) models at various temporal scales. Particle Swarm Optimization (PSO)-optimized Long Short-Term Memory (LSTM) models also showed improved predictive capabilities in forecasting climate-driven Cyanobacterial Harmful Algal Blooms (CyanoHAB) dynamics under multiple Shared Socioeconomic Pathways (SSP) scenarios 32 . Hybrid sequence-to-sequence frameworks combining Temporal Convolutional Networks (TCN) and Transformer architectures surpassed traditional downscaling techniques and standalone LSTM models in climate projection efforts 33 . In agricultural drought forecasting, Rezaei et al. 34 reported that Support Vector Machine with Particle Optimization Algorithm (SVM-POA) consistently outperformed Adaptive Neuro-Fuzzy Inference System with Particle Optimization Algorithm (ANFIS-POA) and Artificial Neural Network with Particle Optimization Algorithm (ANN-POA) across multiple Agricultural Standardized Precipitation Index (ASPI) timescales. Hybrid modeling has also proven effective in related environmental challenges, such as the statistical downscaling of General Circulation Models (GCMs), where integrated deep learning frameworks enhance climate projection reliability 33 . These advancements demonstrate the strengths of hybrid systems in modeling complex hydroclimatic processes. Nevertheless, several important gaps remain in the current literature. First, many studies focus primarily on forecasting accuracy without first examining long-term drought trend behaviour, which limits understanding of the structural evolution of drought conditions. Second, most hybrid forecasting frameworks rely heavily on signal decomposition techniques such as wavelets or CEEMDAN. Although effective in some contexts, these approaches increase computational complexity and fragment the original signal structure, potentially altering the temporal relationships embedded in the data. Third, very few studies integrate trend detection with a smoothing-based deep hybrid architecture, particularly within the hydroclimatic context of the uMkhanyakude District. To address these limitations, this study proposes a novel Savitzky–Golay Temporal Convolutional Network Long Short-Term Memory (SG–TCN–LSTM) framework for meteorological drought prediction using SPI series spanning 1980–2023. The methodological contribution of this study is threefold: Integrated trend–forecasting framework Innovative Trend Analysis (ITA) is incorporated prior to model development, enabling the identification of long-term drought tendencies and linking structural climate variability with predictive modelling. Smoothing-based preprocessing strategy Instead of decomposing the time series into multiple intrinsic components, a Savitzky–Golay filter is employed to reduce noise while preserving the temporal structure and peak characteristics of the SPI series, thereby maintaining signal integrity for downstream learning. Hybrid deep learning architecture The Temporal Convolutional Network (TCN) captures long-range dependencies through dilated causal convolutions, while the Long Short-Term Memory (LSTM) network models sequential memory and short-term variability. Their integration enables the simultaneous learning of multi-scale drought dynamics. By replacing decomposition-based preprocessing with a smoothing-driven learning framework, the proposed SG–TCN–LSTM architecture reduces computational complexity while preserving the structural properties of the original hydroclimatic signal. Methodologically, this study introduces a streamlined hybrid modelling strategy that integrates trend detection and deep learning forecasting within a unified framework. From a regional perspective, it provides the first integrated trend–forecasting drought assessment for the uMkhanyakude District, offering insights that support water resource planning, agricultural risk management, and climate adaptation in one of South Africa’s most drought-prone regions. Methods The research was conducted in the uMkhanyakude District Municipality, situated in the northern region of KwaZulu-Natal Province, South Africa (32.01°S, 27.62°E) (see Fig. 1 ). The district encompasses an area of approximately 13,855 km 2 , bordered by Mozambique to the north, Eswatini to the northwest, and the Indian Ocean to the east. Its topography varies from low-lying coastal plains along the Indian Ocean to the Lebombo Mountains in the west, with elevations ranging from sea level to over 700 meters. The uMkhanyakude District exhibits significant ecological diversity, comprising a range of habitats from wetlands to savannah. The region experiences a subtropical climate characterized by hot, wet summers (November–March) and cool, dry winters (June–August). Mean annual rainfall fluctuates between 600 and 1,200 mm, with higher precipitation levels recorded in the eastern coastal zone and decreased amounts in the interior. Average temperatures range from 18 °C in winter to 30 °C in summer, although extreme heat events exceeding 40 °C have been documented. This district is of national significance due to its ecological and socio-economic value, as it includes the iSimangaliso Wetland Park, a UNESCO World Heritage Site, along with various game reserves such as Hluhluwe–iMfolozi, Ndumo, Tembe Elephant Park, and Mkuze Game Reserve. Despite its rich biodiversity and tourism potential, the district remains predominantly rural, with many communities dependent on rainfed agriculture and natural resources. The interplay of climate variability, water insecurity, and high poverty levels renders it one of the most drought-prone and socio-economically vulnerable regions in South Africa. Fig. 1. Open in a new tab Elevation Map of the uMkhanyakude District Municipality in KwaZulu-Natal Province, South Africa. Red points indicate the locations of the six meteorological stations used in this study. Daily rainfall data from 1980 to 2023 were obtained from the South African Weather Service (SAWS) for six meteorological stations: Pongolapoort Dam, Mkuze Game Reserve, Hlabisa Mbazwana, False Bay Park, Ingwavuma Manguzi, and Riverview. The long-term rainfall variability for the six meteorological stations is presented in Supplementary Figure S1 . These stations were chosen based on data completeness, spatial coverage, and record continuity. Quality control procedures were implemented to identify and rectify any missing or inconsistent observations. The resulting rainfall series was used to compute the SPI for drought characterization and modeling. Standardized precipitation index calculation The Standardized Precipitation Index (SPI), introduced by McKee et al. 6 , quantifies meteorological drought conditions. It is widely used in drought assessments due to its ability to standardize precipitation anomalies across various temporal scales, facilitating comparisons across regions and climates 35 . In this study, SPI values were derived by fitting a gamma probability distribution to accumulated rainfall data at specific timescales (SPI-6, SPI-9, and SPI-12) and transforming cumulative probabilities into a standard normal distribution. This process ensures that the SPI has a mean of zero and a variance of one, improving the classification of wet and dry conditions 36 , 37 . To address zero rainfall occurrences, cumulative probabilities were adjusted before normalization. SPI categorization followed the framework by McKee et al. 6 , where negative values indicate drought conditions and positive values indicate wet conditions, with thresholds distinguishing moderate, severe, and extreme events (see Table S1 in the Supplementary Information). For conciseness, the detailed mathematical derivation of the gamma distribution parameters, cumulative probabilities, and standardization equations are provided in the Supplementary Information. Innovative trend analysis The Innovative Trend Analysis (ITA) method, first introduced by Şen 38 , offers a key advantage over traditional approaches such as the Mann–Kendall test, as it does not require assumptions regarding serial correlation, data normality, or sample size. The procedure begins by dividing the time series into two equal sub-series, which are then independently arranged in ascending order. The first half of the series is plotted on the x-axis, while the second half is plotted on the y-axis. A graphical illustration of the ITA procedure is provided in Supplementary Figure S2 . If the plotted points lie along the 1:1 reference line (45° line), the series exhibits no discernible trend. In contrast, points concentrated above the line indicate an increasing trend, whereas clustering below the line signifies a decreasing trend 38 , 39 . Savitzky–Golay filter algorithm The Savitzky-Golay filter, introduced by Savitzky and Golay in 1964 40 , is a strong tool for signal denoising and smoothing. This method smooths signals by employing convolution on time-domain signals and polynomial least-squares in each convolution window 41 . This study utilised the Savitzky–Golay (SG) filter, a polynomial function fitting method to eliminate noise from atmospheric conditions. This method is very efficient for analysing datasets characterised by attributes that include widths, maxima, and minima, assuming the data points are evenly distributed along a specified axis 42 . Its superiority over other preprocessing approaches is attributed to its distinctive integration of smoothing and the retention of critical signal properties. Schafer 43 , Guo et al. 44 , and Sibiya et al. 15 described the algorithm of SG in detail. The SG filter is computationally efficient and can be executed in real-time systems. For accurate analysis of the dataset, the generated curve must be continuous and smooth 45 . The implementation of the SG filter significantly reduced the influence of SPI outliers resulting from atmospheric disturbances. Autoregressive integrated moving average The Autoregressive Integrated Moving Average (ARIMA) model, originally introduced by Box and Jenkins 46 , was employed as a benchmark predictive model in this study. ARIMA is a stochastic linear time-series model that combines autoregressive (AR), differencing (I), and moving average (MA) components to capture temporal dependencies in stationary time series. The general ARIMA model is denoted as , where p represents the autoregressive order, d denotes the degree of differencing required to achieve stationarity, and q corresponds to the order of the moving average component. The model development followed the classical Box–Jenkins methodology, and a schematic illustration of this procedure is provided in Figure S3 in the Supplementary Material to improve readability of the main manuscript. The modelling process consisted of the following steps: Stationarity testing The Augmented Dickey–Fuller (ADF) test was applied to assess the stationarity of the SPI time series. When non-stationarity was detected, appropriate differencing was applied. Model identification Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) plots were examined to determine suitable values for the parameters p and q . In addition, the auto_arima() function implemented in Python was used to automatically identify candidate model structures. Model estimation Model parameters were estimated using the maximum likelihood estimation (MLE) approach. Model selection The Akaike Information Criterion (AIC) was used to select the optimal model among competing ARIMA specifications. The ARIMA models were fitted separately for each SPI timescale (SPI-6, SPI-9, and SPI-12) across the six meteorological stations considered in this study. Long and short-term memory network The Long Short-Term Memory (LSTM) network, introduced by Hochreiter and Schmidhuber 47 , is a specialized architecture within recurrent neural networks (RNNs) designed to capture long-term dependencies in sequential data. Unlike conventional RNNs, which are prone to vanishing and exploding gradient problems, LSTM networks employ gated mechanisms that regulate the flow of information through the network, enabling stable learning over extended time horizons. An LSTM cell consists of three principal gates: the forget gate , input gate , and output gate , which collectively control the updating and propagation of the cell state. The forget gate determines which information from the previous state should be retained or discarded, the input gate regulates the incorporation of new information into the cell state, and the output gate governs the generation of the hidden state used for prediction. Through this gating mechanism, LSTM networks effectively preserve relevant historical information while filtering out noise, making them well suited for modelling nonlinear temporal dynamics in hydrological and climatological time series. In this study, the LSTM model was employed to capture nonlinear patterns in the SPI series. A schematic representation of the LSTM architecture is provided in Figure S4 in the Supplementary Material, while the detailed mathematical formulation of the gating operations is also included in the Supplementary Material to maintain the conciseness of the main manuscript. Temporal convolutional network The Temporal Convolutional Network (TCN) is a deep learning architecture designed for sequence modelling that employs convolutional operations to capture temporal dependencies in time-series data 48 . Unlike RNNs, which process sequences sequentially, TCNs utilize convolutional structures that allow parallel computation during training and prediction, thereby improving computational efficiency and scalability. The TCN architecture incorporates two key features: causal convolution and dilated convolution. Causal convolution ensures that predictions at time t depend only on current and past observations, thereby preventing information leakage from future time steps 49 . Dilated convolution extends the receptive field of the network by introducing a dilation factor that skips certain input elements, allowing the model to capture long-range temporal dependencies without requiring excessively deep networks 50 . In this way, the receptive field grows exponentially with the network depth. To improve training stability in deep architectures, TCNs employ residual blocks, which incorporate shortcut connections between layers. These residual connections facilitate gradient propagation and mitigate the vanishing gradient problem commonly encountered in deep neural networks 51 . Each residual block typically consists of stacked dilated causal convolution layers followed by nonlinear activation functions and dropout regularization to prevent overfitting. In this study, the TCN model was employed to capture complex temporal patterns in the SPI series. A schematic illustration of the TCN residual block structure is provided in Figure S5 in the Supplementary Material, while the detailed mathematical formulation of the convolution operations and residual connections is also included in the Supplementary Material to maintain the conciseness of the main manuscript. The development of the hybrid ARIMA-LSTM model The ARIMA–LSTM hybrid model was developed to capture both linear and nonlinear patterns in the SPI time series. In the first stage, the ARIMA model was applied to model the linear structure of the SPI data. The residuals obtained from the ARIMA forecasts were then used as inputs to the LSTM network, which was trained to learn the remaining nonlinear relationships in the series. The final prediction was obtained by combining the ARIMA forecasts with the residual forecasts generated by the LSTM model. This hybrid approach leverages the strengths of both models: ARIMA effectively captures short-term linear dependencies and autocorrelation structures, while LSTM networks are capable of learning complex nonlinear temporal dynamics. Consequently, the ARIMA–LSTM framework provides improved predictive performance compared with the individual models. A schematic representation of the ARIMA–LSTM hybrid framework is provided in Figure S6 in the Supplementary Material. The development of the hybrid SG-TCN-LSTM model The SG–TCN–LSTM architecture was developed following a structured, ablation-informed design framework aimed at addressing three inherent characteristics of SPI time series: (i) high-frequency stochastic variability, (ii) multi-scale temporal feature structure, and (iii) nonlinear drought persistence with regime-dependent memory. Rather than stacking deep learning components heuristically, each module was incorporated to resolve a specific modeling limitation observed in standalone approaches. Phase I: Deterministic Noise Attenuation via Savitzky–Golay Filtering SPI series derived from precipitation records frequently exhibit short-term fluctuations attributable to transient meteorological disturbances. When directly supplied to deep neural networks, such variability increases gradient variance, amplifies overfitting risk, and obscures physically meaningful drought signals. To mitigate this effect, a Savitzky–Golay (SG) filter is first applied: 1 where n denotes the window length and R the polynomial order. Unlike simple moving averages, the SG filter performs local polynomial regression, preserving curvature and turning points that are critical for identifying drought onset and recovery phases. Ablation experiments excluding the SG stage resulted in increased validation error and reduced stability across stations, confirming that controlled smoothing enhances downstream feature extraction rather than oversimplifying the signal. Phase II: Hierarchical Temporal Feature Extraction via TCN The filtered series is subsequently processed by a Temporal Convolutional Network (TCN) composed of stacked residual blocks with dilated causal convolutions: 2 3 The TCN was selected for its capacity to construct hierarchical temporal representations through controlled receptive field expansion while preserving causal structure. Dilated convolutions enable efficient modeling of structured short- and medium-range dependencies without recurrent feedback, thereby improving training stability and parallelization efficiency. However, although TCNs can theoretically approximate long-range dependencies by increasing depth or dilation factors, empirical evaluation revealed diminishing predictive returns when expanding the network solely to capture drought persistence. Deeper TCN configurations increased parameter complexity and computational burden without proportional gains in generalization performance. This observation suggests that convolutional receptive fields alone are insufficient to fully model the nonlinear accumulation effects characteristic of prolonged drought regimes. Phase III: Nonlinear Regime Persistence Modeling via LSTM To address this limitation, the TCN-derived feature representation is supplied to an LSTM layer, which produces hidden state and subsequent forecast output: 4 where w , bias, and relu (.) represent a weight matrix, bias vector, and ReLU activation function, respectively. Finally, is adopted as the input to the output layer to obtain for the next time point 5 where is a weight parameter for the fully connected layer, q denotes a bias parameter, and linear (.) is a linear function. The LSTM contributes adaptive memory mechanisms via gated cell states, enabling selective retention and forgetting of temporal information. This dynamic state evolution is particularly suited to SPI dynamics, where drought intensification and recovery exhibit nonlinear accumulation and lag effects 52 . Unlike convolutional layers with fixed receptive fields, LSTMs maintain internal memory trajectories that evolve according to the sequence history, thereby capturing regime persistence more effectively. This staged architecture reduces the representational burden imposed on any single component, improves bias–variance balance, and enhances generalization across SPI scales and stations. Importantly, performance gains were not achieved merely through increased parameterization, as larger standalone TCN or LSTM architectures did not consistently replicate the hybrid’s forecasting accuracy without overfitting risk. Overall, the hybrid SG–TCN–LSTM model constitutes a functionally decomposed forecasting framework in which each module addresses a clearly defined statistical property of drought time series, resulting in improved robustness and predictive skill across diverse meteorological contexts (Fig. 2 ). Fig. 2. Open in a new tab Proposed SG-TCN-LSTM model including the SG filter, two TCN residual block layers, an LSTM layer, a fully connected layer, and an output layer. Model evaluation indicators To measure the prediction accuracy of the model, the root mean square error (RMSE), directional symmetry (DS), and coefficient of determination ( ) are used as evaluation indicators. The smaller the RMSE performance indicator value and the larger the DS, the closer they are to the actual value. In addition, the closer the value of to 1, the better the fit of the prediction model. The specific equations are as follows: 6 7 8 where where n is the number of data points, is the observed value, is the predicted value and is the mean of the observation data. Results and discussion Climate descriptive analysis and trends To enhance the understanding of drought dynamics within the study area, key climate parameters were analysed utilizing the SPI. Figure 3 illustrates the annual means of drought-relevant variables. The SPI was analyzed at 6-, 9-, and 12-month timescales to assess drought variability and long-term precipitation patterns in the uMkhanyakude district. The SPI time series indicates multiple drought episodes throughout the study period (1980–2023), with a notably prolonged drought occurring between 2004 and 2009. A second significant drought episode was recorded from 2014 to 2016, coinciding with the intense 2015–2016 El Niño event that substantially diminished rainfall across Southern Africa. These occurrences underscore the heightened vulnerability of the region to climatic variability and extended precipitation deficits. Fig. 3. Open in a new tab Observed SPI values at the 6-, 9-, and 12-month timescales calculated from the rainfall data measured at the uMkhanyakude district rainfall data for the period from 1980 to 2023. The blue and red shades indicate moist and dry conditions. The ITA results, summarized in Fig. 4 and Table S2 in Supplementary Materials reveal a consistent drying trend across the majority of monitoring stations, particularly at longer accumulation timescales. Five stations–False Bay Park, Hlabisa Mbazwana, Pongolapoort Dam, Mkuze Game Reserve, and Ingwavuma Manguzi exhibit declining SPI trends, whereas Riverview demonstrates a gradual increase in precipitation. The drying signal becomes increasingly pronounced from SPI-6 to SPI-12, indicating that long-term precipitation deficits are intensifying in the region. This pattern suggests that the likelihood of prolonged drought events may escalate in the future, potentially impacting water availability, agricultural productivity, and ecosystem sustainability in northern KwaZulu-Natal. From a climatological standpoint, these trends are likely influenced by large-scale climate variability, such as the El Niño–Southern Oscillation (ENSO), which is recognized for its role in suppressing summer rainfall across southern Africa during El Niño phases. The observed spatial heterogeneity, particularly the wetter conditions at Riverview, may also reflect local topographic or microclimatic influences on rainfall distribution. Overall, the trend analysis emphasizes the necessity for reliable drought monitoring and forecasting systems to facilitate proactive water resource management. Fig. 4. Open in a new tab Innovative trend analysis of monthly total rainfall data measured at six meteorological stations within the uMkhanyakude district, South Africa. SPI time series filtering The study introduces the SG data preprocessing methodologies for the development of a hybrid model aimed at mitigating noise and enhancing the quality of the SPI time series. Figure 5 presents the time series data for SPI-6, SPI-9, and SPI-12 plotted with their respective SG-filtered smoothed data, while the remaining results are provided in the Supplementary Materials (see Figure S7 – S9 ). The SG filter was employed to smooth the time series, with its effectiveness dependent on the selected parameters: a window size of 21 and a polynomial order of 5, determined through a trial-and-error approach during the experimental phase. As depicted in Fig. 5 , the SG filter demonstrates considerable efficacy in smoothing the SPI time series, as evidenced by the filtered curve’s close adherence to the original data. However, it is crucial to emphasize that the SG filter functions solely as a preprocessing technique aimed at noise reduction and series stabilization, without engaging in forecasting. The SG filter contributes by furnishing cleaner input data to the deep learning models, thereby enhancing their convergence and stability. Consequently, the hybrid model, which integrates SG, TCN, and LSTM, capitalizes on the strengths of each component: SG eliminates high-frequency noise, TCN extracts multi-scale features, and LSTM captures long-term memory effects. This integration yields more reliable and accurate forecasts compared to the application of SG in isolation, as evidenced by the performance metrics across all monitoring stations. Fig. 5. Open in a new tab Temporal Variability of the SPI at Different Locations and Time Scales: The graphical representations illustrate the original SPI data (represented by blue lines) alongside the Savitzky-Golay filtered trends (depicted in red) for SPI-6, SPI-9, and SPI-12 across diverse geographical regions within the uMkhanyakude district. Hyperparameter tuning settings All models were implemented in Python 3.10 utilizing Google Colab, a cloud-based platform that offers GPU acceleration to enhance computational efficiency. Various deep learning architectures, including LSTM, TCN, and the proposed SG–TCN–LSTM hybrid framework, were developed using the TensorFlow/Keras library. The ARIMA model was executed with the help of the pmdarima and statsmodels packages. Data preprocessing, normalization, and performance evaluation were conducted utilizing NumPy, Pandas, and Scikit-learn, while the SG filter was employed through the SciPy library. The SPI time series from each meteorological station was segmented using a chronological 80/20 split, wherein the initial 80% of observations were allocated for model training, and the remaining 20% were designated for out-of-sample testing. This method preserves the temporal structure of the data and mitigates the risk of information leakage that may occur from random shuffling in time-series forecasting tasks. Forecasting models were trained within a supervised sliding-window framework for one-step-ahead prediction, utilizing lagged SPI observations to forecast the subsequent time step. Hyperparameter optimization for the deep learning models was conducted using Bayesian optimization, which systematically explores the hyperparameter space to minimize validation loss and enhance predictive accuracy. The complete ranges for hyperparameter searches and optimal configurations for the ARIMA, TCN, and LSTM models are provided in Tables S3 – S5 in the Supplementary Material. For the ARIMA models, optimal (p,d,q) parameters were ascertained using the procedure based on the Hyndman–Khandakar stepwise algorithm, with model selection guided by the minimization of the Akaike Information Criterion (AIC). Prior to model estimation, the stationarity of the SPI series was assessed using the ADF test, and differencing was applied as necessary to ensure stationarity. All models were trained independently for each SPI timescale and meteorological station, thereby ensuring methodological rigor and facilitating a consistent evaluation of forecasting performance across varying climatic conditions. Early stopping was implemented during training to prevent overfitting and enhance model generalization. While Bayesian optimization enhances model performance through the efficient identification of optimal hyperparameter combinations, it is computationally intensive. Consequently, the optimization process was meticulously designed to balance computational cost and predictive accuracy. The resulting parameter configurations represent the optimal settings for the SPI datasets analyzed in this study. Comparative performance of forecasting models To assess the capability of drought prediction, six forecasting models were developed: ARIMA, LSTM, TCN, ARIMA-LSTM, TCN-LSTM, and the proposed SG-TCN-LSTM hybrid model. The performance of these models was evaluated using root mean square error (RMSE), the coefficient of determination ( ), and directional symmetry (DS) metrics. The results indicate that hybrid models consistently outperform standalone models across all SPI timescales and monitoring stations. For example, Figures 6 and 7 present the results for the False Bay Park station, while the results for the remaining stations are provided in the Supplementary Materials (see Figures S10 – S19 ). The ARIMA model, which captures linear temporal structures, provides a reasonable baseline prediction but shows limitations in representing the nonlinear relationships inherent in the SPI series. In contrast, deep learning models such as LSTM and TCN improve forecasting performance by effectively capturing nonlinear dynamics and complex temporal dependencies present in the rainfall data. Among the two deep learning architectures, the TCN model generally achieved superior prediction accuracy compared to the LSTM model. This improvement is likely attributable to the dilated causal convolution mechanism employed by TCN, which enables the model to capture multi-scale temporal patterns while maintaining computational efficiency through parallel processing. Nevertheless, both models exhibited limitations when applied independently, particularly in their ability to capture abrupt fluctuations in the SPI series. Fig. 6. Open in a new tab Comparison of observed and predicted SPI values at the False Bay Park station for three accumulation timescales: SPI-6, SPI-9, and SPI-12 over the period 2015–2023. The black line represents the observed SPI series, while the colored lines correspond to predictions generated by ARIMA, TCN, LSTM, ARIMA-LSTM, TCN-LSTM, and SG-TCN-LSTM models. Fig. 7. Open in a new tab Taylor diagrams comparing the performance of ARIMA, TCN, LSTM, ARIMA-LSTM, TCN-LSTM, and SG-TCN-LSTM models in predicting SPI-6, SPI-9, and SPI-12 at the False Bay Park station. Hybrid models that integrate different modeling strategies further enhanced forecasting accuracy. The ARIMA-LSTM model improved prediction performance by decomposing the linear and nonlinear components of the SPI series. However, the TCN-LSTM hybrid model demonstrated greater predictive capability compared to ARIMA-LSTM, suggesting that convolution-based feature extraction can effectively complement recurrent neural networks in capturing complex climatic patterns. The proposed SG-TCN-LSTM model achieved the best overall performance across all stations and timescales. The RMSE values ranged from 0.2018 to 0.2910 for SPI-6, 0.1275 to 0.2119 for SPI-9, and 0.0487 to 0.0897 for SPI-12, while the corresponding values ranged between 0.9456 and 0.9723 for SPI-6, 0.9785 and 0.9862 for SPI-9, and 0.9953 and 0.9991 for SPI-12. These findings indicate that the SG-TCN-LSTM framework significantly reduces prediction error while enhancing the reliability of drought forecasts (Table 1 ). Table 1. Comparison of prediction performance of the models. Model SPI-6 SPI-9 SPI-12 RMSE DS RMSE DS RMSE DS False Bay Park ARIMA 0.3504 0.8945 0.8422 0.3481 0.8795 0.8325 0.3169 0.9421 0.8446 LSTM 0.3184 0.9121 0.8462 0.2431 0.9567 0.8595 0.1667 0.9687 0.8969 TCN 0.2864 0.9196 0.8106 0.2336 0.9582 0.8562 0.1597 0.9692 0.9011 ARIMA+LSTM 0.2706 0.9260 0.8135 0.1765 0.9732 0.8586 0.1405 0.9694 0.9054 TCN+LSTM 0.2416 0.9328 0.8135 0.1565 0.9785 0.8864 0.0949 0.9995 0.8951 SG-TCN+LSTM 0.2018 0.9556 0.8012 0.1275 0.9862 0.8757 0.0812 0.9974 0.8959 Mkuze Game Reserve ARIMA 0.3572 0.8642 0.8419 0.3475 0.8957 0.8792 0.2202 0.9697 0.8547 LSTM 0.3265 0.9019 0.8756 0.2514 0.9500 0.8539 0.1329 0.9843 0.9023 TCN 0.3265 0.9195 0.8826 0.3131 0.8732 0.8356 0.1499 0.9817 0.8934 ARIMA+LSTM 0.3179 0.9194 0.8541 0.2560 0.9546 0.8541 0.1085 0.9957 0.8956 TCN+LSTM 0.3129 0.9182 0.8512 0.1512 0.9854 0.8851 0.1095 0.9971 0.8937 SG-TCN+LSTM 0.2289 0.9682 0.8673 0.1531 0.9853 0.8451 0.0818 0.9971 0.9114 Pongolapoort Dam ARIMA 0.4470 0.8797 0.8624 0.2933 0.9668 0.8119 0.1918 0.9882 0.8992 LSTM 0.4470 0.9023 0.8756 0.2295 0.9568 0.8422 0.1738 0.9915 0.9065 TCN 0.3221 0.8959 0.8228 0.2094 0.9851 0.8921 0.1438 0.9851 0.9031 ARIMA+LSTM 0.3221 0.8969 0.8228 0.2045 0.9852 0.8912 0.1438 0.9852 0.9031 TCN+LSTM 0.2221 0.9365 0.8254 0.1814 0.9852 0.8621 0.1125 0.9993 0.9015 SG-TCN+LSTM 0.2134 0.9723 0.8914 0.1529 0.9862 0.8562 0.0568 0.9978 0.9114 Hlabisa Mbazwana ARIMA 0.4704 0.9347 0.8624 0.2436 0.9597 0.8891 0.1918 0.9845 0.9013 LSTM 0.3471 0.9041 0.8532 0.2143 0.9673 0.8657 0.1667 0.9893 0.8934 TCN 0.3471 0.9397 0.8512 0.1765 0.9751 0.8864 0.1405 0.9951 0.8945 ARIMA+LSTM 0.3269 0.9549 0.8515 0.2195 0.9589 0.8815 0.1145 0.9989 0.8952 TCN+LSTM 0.3150 0.9499 0.9115 0.2179 0.9599 0.8859 0.0975 0.9985 0.9021 SG-TCN+LSTM 0.2910 0.9616 0.8941 0.1537 0.9851 0.8853 0.0897 0.9987 0.9114 Ingwavuma Manguzi ARIMA 0.4712 0.8715 0.8571 0.2704 0.8792 0.8750 0.2252 0.9794 0.8964 LSTM 0.3348 0.8981 0.8738 0.2544 0.9024 0.8955 0.1614 0.9893 0.8975 TCN 0.3466 0.9095 0.8735 0.2242 0.9491 0.8935 0.1365 0.9935 0.8935 ARIMA+LSTM 0.3158 0.9495 0.8923 0.2279 0.9567 0.8949 0.1225 0.9954 0.8931 TCN+LSTM 0.3045 0.9579 0.9112 0.2293 0.9572 0.8951 0.1083 0.9983 0.8947 SG-TCN+LSTM 0.2735 0.9687 0.8845 0.2119 0.9785 0.9113 0.0914 0.9991 0.8979 Riverview ARIMA 0.4375 0.8132 0.8106 0.1708 0.9074 0.8013 0.1577 0.9507 0.7973 LSTM 0.3245 0.8954 0.8105 0.1597 0.9321 0.8037 0.1385 0.9694 0.8056 TCN 0.3245 0.8951 0.8125 0.1348 0.9437 0.8112 0.1297 0.9781 0.8142 ARIMA+LSTM 0.3214 0.9454 0.8112 0.1275 0.9751 0.8256 0.1108 0.9895 0.8241 TCN+LSTM 0.3047 0.9521 0.8323 0.1083 0.9851 0.8412 0.0945 0.9981 0.8409 SG-TCN+LSTM 0.2735 0.9456 0.8445 0.1089 0.9856 0.8456 0.0487 0.9953 0.8503 Open in a new tab Influence of drought trends on forecasting performance An important observation from the results is the relationship between drought trends and forecasting performance. Stations exhibiting stronger decreasing precipitation trends generally show slightly higher prediction errors compared to stations with more stable or increasing precipitation patterns. This behaviour reflects the influence of non-stationarity, which introduces structural changes in the statistical properties of the time series and increases forecasting uncertainty. For example, stations such as Ingwavuma Manguzi and Hlabisa Mbazwana, which show strong drying trends in the ITA analysis, exhibit slightly higher RMSE values compared to the Riverview station, where an increasing precipitation trend was observed. Despite this challenge, the SG-TCN-LSTM model maintains relatively stable prediction accuracy across all stations, suggesting that the hybrid framework is robust to moderate levels of non-stationarity. This robustness can be attributed to the complementary roles of the three model components. The SG filter reduces high-frequency noise and stabilizes the SPI series, allowing the deep learning models to focus on meaningful temporal structures. The TCN component extracts multi-scale temporal features using dilated causal convolutions, while the LSTM component captures long-term dependencies within the data. The integration of these mechanisms enables the hybrid model to better adapt to evolving climatic patterns compared to standalone models. Environmental and policy implications The increasing drought tendencies observed across the majority of monitoring stations highlight significant environmental and socio-economic risks for the uMkhanyakude district. The region is heavily dependent on rainfall-driven water resources for agriculture, domestic consumption, and ecosystem sustainability. Prolonged precipitation deficits may therefore exacerbate water scarcity, reduce crop yields, and increase pressure on already vulnerable rural communities. Accurate drought forecasting plays a crucial role in mitigating these risks by providing early warning signals that support proactive decision-making. The improved predictive performance of the SG-TCN-LSTM model suggests that advanced hybrid forecasting frameworks can contribute to more reliable drought early-warning systems. Such systems could assist water resource managers in optimizing reservoir operations, implementing timely water restrictions, and planning climate adaptation strategies. From a policy perspective, integrating machine learning-based forecasting models into national drought monitoring systems could enhance the effectiveness of climate risk management strategies in South Africa. Improved forecasting accuracy may enable government agencies to implement targeted interventions, such as drought relief programs, agricultural advisories, and water allocation policies, before severe impacts occur. Additionally, the development of robust drought prediction tools aligns with broader climate resilience objectives under national and regional climate adaptation frameworks. Conclusion and future remarks This study examined meteorological drought dynamics and forecasting performance in the uMkhanyakude district, South Africa, using the Standardized Precipitation Index (SPI) at 6-, 9-, and 12-month timescales. The trend analysis revealed a consistent drying tendency across most monitoring stations, with five stations exhibiting decreasing precipitation trends, particularly at longer accumulation periods. These results indicate an increasing risk of prolonged drought conditions in the region, which may have significant implications for agricultural productivity, water resource availability, and ecosystem stability. To improve drought prediction, several forecasting models were evaluated, including ARIMA, LSTM, TCN, and hybrid frameworks. The findings demonstrate that hybrid approaches outperform individual models, with the proposed SG–TCN–LSTM model providing the most reliable predictions across all stations and SPI timescales. The integration of Savitzky–Golay filtering with deep learning architectures enables effective noise reduction, multi-scale feature extraction, and improved modelling of nonlinear temporal dependencies in hydroclimatic time series. From a practical perspective, the enhanced forecasting capability of the SG–TCN–LSTM framework provides a valuable tool for strengthening drought early-warning systems and supporting evidence-based water resource management in drought-prone regions. However, this study presents two primary limitations. First, the analysis is predicated on rainfall data from a limited number of stations within the uMkhanyakude District, which may constrain the generalizability of the findings to other regions. Second, the investigation exclusively utilizes precipitation-based indices, specifically the SPI, and does not account for other significant hydroclimatic variables, such as temperature or soil moisture. Furthermore, uncertainties in long-term forecasting persist due to the inherent variability and non-stationarity of climate systems. Future research should explore the integration of additional climate drivers (e.g., ENSO indices, temperature, and evapotranspiration) and multi-source datasets to further improve drought prediction accuracy and strengthen climate resilience strategies under increasing hydroclimatic variability. Supplementary Information Supplementary Material 1. (9.1MB, zip) Author contributions S.S. and N.B. Conceptualization, methodology, software, validation; S.S. formal analysis, writing–original draft preparation, N.B., S.R. and S.M. writing–review and editing, N.B., S.R., and S.M. supervision. All authors have read and agreed to the published version of the manuscript. All authors reviewed the manuscript. Funding The authors declare no funding. Data availability The rainfall datasets used in this study were obtained from the South African Weather Service (SAWS) and are not publicly available due to licensing restrictions. Derived datasets, including computed SPI time series and model prediction outputs, are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Shaun Ramroop, Sileshi Melesse, and Nkanyiso Mbatha have contributed equally to this work. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-46664-x. References 1. Ge, W. et al. Quantitative evaluation of drought risk related to vegetation productivity in china. J. Hydrol. 623 , 129877 (2023). [ Google Scholar ] 2. Wang, L., She, D., Xia, J., Meng, L. & Li, L. Revegetation affects the response of land surface phenology to climate in loess plateau, china. Sci. Total Environ. 860 , 160383 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 3. Liu, J., Xia, J. & Wang, M. Comparative study on bivariate statistical characteristics of drought in shandong using spi and spei. Sci. Rep. 15 , 11268 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Zhang, Y., Keenan, T. F. & Zhou, S. Exacerbated drought impacts on global ecosystems due to structural overshoot. Nat. Ecol. Evolut. 5 , 1490–1498 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Li, Y. et al. Spatiotemporal heterogeneity in meteorological and hydrological drought patterns and propagations influenced by climatic variability, lulc change, and human regulations. Sci. Rep. 14 , 5965 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. McKee, T. B., Doesken, N. J., Kleist, J. et al. The relationship of drought frequency and duration to time scales. In Proceedings of the 8th Conference on Applied Climatology , vol. 17, 179–183 (California, 1993). 7. Vicente-Serrano, S. M., Beguería, S. & López-Moreno, J. I. A multiscalar drought index sensitive to global warming: the standardized precipitation evapotranspiration index. J. Clim. 23 , 1696–1718 (2010). [ Google Scholar ] 8. Palmer, W. C. Meteorological drought. us. Weather Bureau Res. Paper 45 , 1–58 (1965). [ Google Scholar ] 9. Ceppi, A., Achite, M., Toubal, A. & Caloiero, T. Mapping drought characteristics in northern algerian basins using the era5-land dataset. Sci. Rep. 15 , 10720 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Nikoo, M. R., Zarei, E. & Al-Wardy, M. Non-stationary drought patterns in hyper-arid regions: Spatiotemporal and multi-timescale drought analysis. Sci. Total Environ. 1000 , 180401 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Danandeh Mehr, A., Sorman, A. U., Kahya, E. & Hesami Afshar, M. Climate change impacts on meteorological drought using spi and spei: case study of ankara, turkey. Hydrol. Sci.J. 65 , 254–268 (2020). [ Google Scholar ] 12. Liu, C., Yang, C., Yang, Q. & Wang, J. Spatiotemporal drought analysis by the standardized precipitation index (spi) and standardized precipitation evapotranspiration index (spei) in sichuan province, china. Sci. Rep. 11 , 1280 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Yaseen, Z. M., Ali, M., Sharafati, A., Al-Ansari, N. & Shahid, S. Forecasting standardized precipitation index using data intelligence models: regional investigation of bangladesh. Sci. Rep. 11 , 3435 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Onyeuwaoma, N., Sivakumar, V. & Bade, M. Modelling drought in south africa: meteorological insights and predictive parameters. Environ. Monitor. Assessment 196 , 965 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Sibiya, S., Mbatha, N., Ramroop, S., Melesse, S. & Silwimba, F. Forecasting of standardized precipitation index using hybrid models: A case study of cape town, south africa. Water 16 , 2469 (2024). [ Google Scholar ] 16. Shayanmehr, S., Rastegari Henneberry, S., Sabouhi Sabouni, M. & Shahnoushi Foroushani, N. Drought, climate change, and dryland wheat yield response: An econometric approach. Int. J. Environ. Res. Public Health 17 , 5264 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Nyamane, S., Abd Elbasit, M. A. & Obagbuwa, I. C. Harnessing deep learning for meteorological drought forecasts in the northern cape, south africa. Int. J. Intell. Syst. 2024 , 7562587 (2024). [ Google Scholar ] 18. Baudoin, M.-A., Nortje, K., Naik, M., Rouault, M. & Vogel, C. South africa: El niño impacts and management in south africa: lessons learned for an ‘el niño ready’ nation. In El Niño Ready Nations and Disaster Risk Reduction: 19 Countries in Perspective , 271–305 (Springer, 2022). 19. Mzimela, J. H., Moyo, I. & Tshabalala, T. Spatiotemporal analysis of agricultural drought variability in the umkhanyakude district municipality, kwazulu-natal. Environ. Monitor. Assessment 197 , 178 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Maziya, M., Nkonki-Mandleni, B. & Van Niekerk, J. Smallholder farmers’ choice of climate change adaptation strategies in the umkhanyakude district in kwazulu-natal, south africa. South African J. Agric. Extension 52 , 97–127 (2024). [ Google Scholar ] 21. Aghelpour, P., Mohammadi, B., Mehdizadeh, S., Bahrami-Pichaghchi, H. & Duan, Z. A novel hybrid dragonfly optimization algorithm for agricultural drought prediction. Stochastic Environ. Res. Risk Assessment 35 , 2459–2477 (2021). [ Google Scholar ] 22. Roushangar, K., Ghasempour, R. & Alizadeh, F. Uncertainty assessment of the integrated hybrid data processing techniques for short to long term drought forecasting in different climate regions. Water Resour. Manag. 36 , 273–296 (2022). [ Google Scholar ] 23. Hinge, G., Piplodiya, J., Sharma, A., Hamouda, M. A. & Mohamed, M. M. Evaluation of hybrid wavelet models for regional drought forecasting. Remote Sens. 14 , 6381 (2022). [ Google Scholar ] 24. Ghasempour, R., Roushangar, K. & Alizadeh, F. Hybrid models for drought forecasting: Integration of multi pre-processing-data driven approaches and non-linear garch time series model. Arabian J. Geosci. 16 , 361 (2023). [ Google Scholar ] 25. Sewal, N. & Virmani, C. A hybrid prediction of drought using artificial intelligence and machine learning. In 2023 5th International Conference on Advances in Computing, Communication Control and Networking (ICAC3N) , 480–485 (IEEE, 2023). 26. Mukhtar, A. et al. A new hybrid weighted regional drought index to improve regional drought assessment. Water Resour. Manag. 38 , 5541–5558 (2024). [ Google Scholar ] 27. Malik, A., Tikhamarine, Y., Sammen, S. S., Abba, S. I. & Shahid, S. Prediction of meteorological drought by using hybrid support vector regression optimized with hho versus pso algorithms. Environ. Sci. Pollut. Res. 28 , 39139–39158 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 28. Piri, J., Abdolahipour, M. & Keshtegar, B. Advanced machine learning model for prediction of drought indices using hybrid svr-rsm. Water Resour. Manag. 37 , 683–712 (2023). [ Google Scholar ] 29. Wan, W. & Zhou, Y. Deep hybridnet for drought prediction based on large-scale climate indices and local meteorological conditions. Stochastic Environ. Res. Risk Assessment 39 , 4725–4747 (2024). 30. Ariyaei, A., Sharafati, A. & Hameed, A. S. A hybrid deep learning approach to predict meteorological drought based on several remote sensing products. Earth Sci. Inf. 18 , 471 (2025). [ Google Scholar ] 31. Waqas, M., Hliang, P. T., Dechpichai, P., Wangwongchai, A. et al. A deep learning perspective on meteorological droughts prediction in the mun river basin, thailand. AIP Adv. 14 (2024). 32. Al-Wardy, M., Zarei, E., Nikoo, M. R. & Nazari, R. Climate-driven projections of cyanobacterial harmful algal bloom expansion in coastal waters. Sci. Total Environ. 992 , 179940 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 33. Zarei, E. et al. Hybrid deep learning downscaling of gcms for climate impact assessment and future projections in oman. J. Environ. Manag. 376 , 124522 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 34. Rezaei, M., Moghaddam, M. A., Piri, J., Azizyan, G. & Shamsipour, A. A. Drought prediction using advanced hybrid machine learning for arid and semi-arid environments. KSCE J. Civ.Eng. 29 , 100025 (2025). [ Google Scholar ] 35. Tsakiris, G. & Vangelis, H. Towards a drought watch system based on spatial spi. Water Resour. Manag. 18 , 1–12 (2004). [ Google Scholar ] 36. Khan, M. M. H., Muhammad, N. S. & El-Shafie, A. Wavelet based hybrid ann-arima models for meteorological drought forecasting. J. Hydrol. 590 , 125380 (2020). [ Google Scholar ] 37. Pande, C. B. et al. Forecasting of spi and meteorological drought based on the artificial neural network and m5p model tree. Land 11 , 2040 (2022). [ Google Scholar ] 38. Şen, Z. Innovative trend analysis methodology. J. Hydrol. Eng. 17 , 1042–1046 (2012). [ Google Scholar ] 39. Şen, Z. Trend identification simulation and application. J. Hydrol. Eng. 19 , 635–642 (2014). [ Google Scholar ] 40. Savitzky, A. & Golay, M. J. Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36 , 1627–1639 (1964). [ Google Scholar ] 41. Wu, J.M.-T. et al. Applying an ensemble convolutional neural network with savitzky-golay filter to construct a phonocardiogram prediction model. Appl. Soft Comput. 78 , 29–40 (2019). [ Google Scholar ] 42. Kim, S.-R. et al. Application of the savitzky-golay filter to land cover classification using temporal modis vegetation indices. Photogrammetric Eng. Remote Sens. 80 , 675–685 (2014). [ Google Scholar ] 43. Schafer, R. W. What is a savitzky-golay filter?[lecture notes]. IEEE Signal Process. Magazine 28 , 111–117 (2011). [ Google Scholar ] 44. Guo, Y. et al. An enhanced pca method with savitzky-golay method for vrf system sensor fault detection and diagnosis. Energy Build. 142 , 167–178 (2017). [ Google Scholar ] 45. Chen, X. et al. Response of global agricultural productivity anomalies to drought stress in irrigated and rainfed agriculture. Sci. China Earth Sci. 67 , 3579–3593 (2024). [ Google Scholar ] 46. Box, G. & Jenkins, G. Forecasting and control. San francisco, Analysis (1976). [ Google Scholar ] 47. Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput. 9 , 1735–1780 (1997). [ DOI ] [ PubMed ] [ Google Scholar ] 48. Wang, Y., Liu, Z., Hu, D. & Zhang, M. Multivariate time series prediction based on optimized temporal convolutional networks with stacked auto-encoders. In Asian Conference on Machine Learning , 157–172 (PMLR, 2019). 49. Lu, Z., Zheng, C. & Yang, T. Application of offshore visibility forecast based on temporal convolutional network and transfer learning. Comput. Intelli. Neurosci. 2020 , 8882279 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Bai, S., Kolter, J. Z. & Koltun, V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271 (2018). 51. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 770–778 (2016). 52. Sibiya, S., Ramroop, S., Melesse, S. & Mbatha, N. Meteorological drought trend analysis and forecasting using a hybrid sg-ceemdan-arima-lstm model based on spi from rain gauge data. Natural Hazards Earth Syst. Sci. 26 , 315–342 (2026). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1. (9.1MB, zip) Data Availability Statement The rainfall datasets used in this study were obtained from the South African Weather Service (SAWS) and are not publicly available due to licensing restrictions. Derived datasets, including computed SPI time series and model prediction outputs, are available from the corresponding author upon reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.9 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Evaluating deep learning models for streamflow prediction under RCP 4.5 and RCP 8.5 climate scenarios#1136569
• An Input-Frugal Deep Learning Framework for Weather-Driven National Crop-Yield Forecasting: A Case Study of Brazilian Soybean#1143048
• Enhanced Rainfall Forecasting Through Deep Learning Optimization Using Long Short-Term Memory Networks#1017859
• A hybrid deep learning approach for winter wheat yield prediction: evidence from leveraging multi-source data.#923696
• Assessing the environmental costs of multi-scale recurrent neural networks for sustainable extreme rainfall nowcasting.#123091