Intercomparison of machine learning methods for statistical downscaling of daily temperature under CMIP6 scenarios: a case study from Iran

Climate change denotes a sustained alteration in long-term climate patterns. This transformation is a direct outcome of the release of greenhouse gases into the atmosphere, a phenomenon rooted in the pre-industrial revolution era (Zahabiyoun et al. 2013). Climate change has evolved into a worldwide challenge with the ability to influence the local scale. Changing climate patterns increase the likelihood of more frequent occurrences of extreme hydrological events, like flash floods or droughts (IPCC 2014). Diverse sectors, such as agriculture, ecosystems, human and food security, as well as natural resource management, face potential threats from these effects, leading to substantial losses. Climate change jeopardizes the sustainable development of nations. Hence, it is crucial to initiate prompt and targeted climate actions aligned with the objectives of sustainable development.

Taking action involves using adaptation as a key strategy. This approach centers on enhancing life adaptability and preparation to mitigate the severity of impacts from future changes (Prathom & Champrasert 2023). To reduce local impacts, an assessment of risks and impacts is conducted by comparing potential future climate scenarios with the present situation. This informs the development of adaptation plans for different sectors (Buras & Menzel 2019; Liu et al. 2019; Miao et al. 2020; Toot et al. 2020). Predicting future climate conditions involves using general circulation models (GCMs) under various scenarios (Goodarzi et al. 2022a, b). These predictions account for different scenarios, such as variations in greenhouse gas emissions, land use, and socioeconomic conditions, providing valuable insights for planning and decision-making. This study specifically concentrates on a future scenario known as ‘Shared Socioeconomic Pathways’ (SSPs), currently employed as an input scenario for the GCM. SSPs are scenarios that simulate how various socioeconomic developments can influence or pose challenges to strategies addressing climate change (Riahi et al. 2017). SSPs include five socioeconomic scenarios labeled SSP1–5. The two SSPs selected in this study are regarded to be a combination of moderate social vulnerability with a moderate emission range (SSP2-4.5) and higher emissions that impose high mitigation but low adaptation challenges (SSP5-8.5) (Zhu & Yang 2021; Ma et al. 2022). However, due to the coarse spatial resolution, GCM outputs might inadequately depict local data. To enhance spatial resolution, the downscaling process is crucial. GCMs are limited by their coarse spatial resolution (250–600 km), leading to discrepancies between their predictions and actual local and global climate variations. The output might not accurately represent information for a small local area. This poses a challenge for adaptation activities (Gebrechorkos et al. 2019). Therefore, to utilize the prediction output, a necessary step involves a process called ‘downscaling’ to enhance spatial resolution and calibrate the prediction output. Downscaling is a method employed to enhance the resolution of data by capturing the historical relationship between observed data and prediction output, with the goal of calibrating the prediction.

Downscaling methods include statistical and dynamical approaches. The former, which is enhanced through machine learning (ML) techniques, is particularly effective for simulating hydroclimatic variables that exhibit complex nonlinear relationships (Xu et al. 2020). In this context, some ML methods have been applied for statistical downscaling, such as genetic programming (GP) (Coulibaly 2004), artificial neural network (ANN) (Ahmed et al. 2015; Tripathi et al. 2006), multiple linear regression (MLR) (Sachindra et al. 2014), relevant vector machine (RVM) (Okkan & Inan 2015), K-nearest neighbors (KNN) (Liu et al. 2017), support vector machine (SVM) (Goly et al. 2014), gene expression programming (Hashmi et al. 2011), and generalized linear models (GLMs) (Beecham et al. 2014). Furthermore, SVM, ANN, and MLR have been used to simulate minimum and maximum monthly temperatures (Duhan & Pandey 2015). Another study highlighted the applicability of RVM, ANN, GP, and SVM in developing downscaling techniques for drought analysis (Sachindra et al. 2018). The better performance of linear regression compared to ANN with limited available data was recognized (Hatanaka 2022). Also, multi-gene GP was found to provide more valid results than ANN for downscaling daily temperature (Niazkar et al. 2023). In addition, a combination of interpolation approaches, such as inverse distance weighted (IDW) with ML models, e.g., IDW-ANN, was used for downscaling precipitation and temperature (Prathom & Champrasert 2023), while a combination of four ML algorithms (Random Forest, KNN, Extra Trees, and Gradient-Boosting Decision Tree) was employed to predict future precipitation (Wang et al. 2023). Nonetheless, several research gaps still remain. To be more specific, challenges in the current methods include improving spatial resolution, capturing extreme events, and better quantifying uncertainties. In addition, downscaling efforts remained underrepresented for hydroclimatic variables, e.g., extreme precipitation or small-scale phenomena, in certain regions. Moreover, many models lack thorough validation against real-world data, raising concerns about their reliability. Finally, issues related to scalability and generalization across different regions and future scenarios require the development of more robust and adaptable models.

The aim of this study is to employ some ML models to downscale outputs of Coupled Model Intercomparison Project 6 (CMIP6) climate models, i.e., Access-CM2, MPI-ESM1-2-HR, MPI-ESM1-2-LR, and MRI-ESM2-0. In this context, this study assesses the applicability of ANN, support vector regression (SVR), decision tree regression (DTR), KNN, Light Gradient-Boosting Machine (LightGBM), and Stochastic Gradient Descent Regressor (SGDRegressor) for downscaling hydroclimatic variables. Based on the literature review, the last three ML models have been rarely applied for statistical downscaling. In addition, the daily temperature for the future period of 2015–2045 is predicted under SSP2-4.5 and SSP5-8.5 climate scenarios.

The daily gridded temperature data obtained from four CMIP6 climate models, shown in Table 1, are for 1995–2015 for the historical period and for 2015–2045 for the future period. The data were collected from the Earth System Grid Association (https://esgf-node.llnl.gov/search/cmip6). Furthermore, SSP is a combination of representative concentration pathways and alternative pathways of socioeconomic development (O'Neill et al. 2016). The two SSPs selected in this study are SSP2-4.5 and SSP5-8.5.

SVR, ANN, DTR, KNN, LightGBM, and SGDRegressor are applied in this study to effectively downscale GCM outputs, each leveraging different strengths in capturing patterns and relationships within the data.

An SVM is an ML algorithm that handles both linear and nonlinear regressions. SVR is widely recognized as one of the most commonly used supervised ML techniques (He et al. 2022). Basically, it creates an optimal hyperplane to transform input data into a higher dimensional space, where data points become vectors (Bisong 2019). The optimum hyperplane, which divides data into two classes, aims to widen the area between the hyperplane and nearest data points, i.e., support vectors. Finally, the SVM prediction at each point is the hyperplane value at that point plus a bias to avoid overfitting. The performance of an SVM model is significantly influenced by kernel hyperparameters and a regularization term (Leong et al. 2021). After comparing linear, polynomial, sigmoid, and radial basis kernel functions, the radial basis function kernel was ultimately chosen for feature research.

The ANN, a method proficient in constructing nonlinear models between input and output samples (Piraei et al. 2023), derives inspiration from the functioning of the human brain's biological nervous system (Agatonovic-Kustrin & Beresford 2000). It is a mathematically modeled system mimicking biology, consisting of processing elements called neurons (or perceptrons) connected with parameters (weights) assigned to the connections, forming the neuronal structure. In comparison to traditional statistical regression techniques like MLR and GLM, ANN demonstrates superior performance, finding widespread application in hydrology and climatology (Kişi 2008; Okkan & Fistikoglu 2014).

In general, a common network in ANN comprises three layers: (i) an input layer including a set of neurons to carry input data, (ii) a hidden layer that enables data flow in the network, and (iii) an output layer holding neuron associated with the output data. Since the input data should be independent, neurons within a layer are not connected to one another, while they are exclusively in connection with the neurons in adjacent layers (Piraei et al. 2023). The connection between two neurons relates their values with a mathematical model, whose parameters can be optimized when data flow back and forth in the network. As a result, the layer-based structure provides a relationship between input and output variables regardless of any prior knowledge on the theoretical background of the problem in question.

The decision tree algorithm, a commonly used ML method for classification and regression tasks, exhibits a tendency to overfit, leading to less-than-optimal performance on testing datasets (Bisong 2019). With a structure resembling a flowchart composed of nodes and branches, the decision tree is widely embraced in data mining due to its simplicity and ease of understanding. In creating a treelike structure, DTR divides data based on feature values at branch nodes. Each data partition corresponds to the result of a splitting test on the training data. Branch nodes link to leaf nodes, where each leaf signifies a model outcome derived by averaging data points from the training data assigned to that node during the splitting process (Piraei et al. 2023).

The KNN technique, used for both classification and regression purposes, operates as a nonparametric method (Altman 1992). Its core function involves identifying the K closest data points to a designated testing data point and calculating the weighted average of their target values. Assessing similarity between the training and testing data points relies on a distance function. Common distance functions for continuous variables include Euclidean, Manhattan, and Minkowski (Nugrahaeni & Mutijarsa 2016).

Using a desirable distance function, distance metrics can be determined and then sorted to obtain the minimum distance that is nearest, which implies the maximum similarity among neighbor points. In addition, the KNN output can be achieved by the K-nearest data points based on sorting the calculated distances. For this purpose, selecting the K value is crucial on KNN performances (Piraei et al. 2023). In essence, a high value for K may incorporate outclass data, whereas a low value may end up in poor training of KNN. Therefore, the optimum number of neighbors (K) should be computed using a cross-validation approach.

LightGBM is a scalable gradient-boosting ML model developed by Microsoft. In comparison to traditional ML models, it exploits a histogram-based technique for enhancing the training process. As a result, it reduces computational efforts and the amount of memory, which enables it to handle large datasets efficiently (Xu et al. 2023). Furthermore, LightGBM can tackle not only classification but also regression tasks.

LightGBM employs a specific strategy for creating decision trees, called leaf-wise. Commonly, the decision tree model adopts the level-wise strategy that works on the leaves of the same layer. Conversely, the leaf-wise strategy first searches all leaves to determine the ones with the highest branching gain and then proceed with the branching cycle. Thus, it can not only obtain higher accuracy by conducting more important splits but also decrease the overfitting issue (Fan et al. 2019). Moreover, LightGBM can employ various optimization algorithms, like Gradient-based One-Side Sampling and Exclusive Feature Bundling, to improve its precision and run time. To be more specific, these optimization techniques make it a powerful tool for developing robust ML models across various domains.

SGDRegressor utilizes stochastic gradient descent (SGD) as its search engine to develop a linear ML model for regression tasks. To be more precise, SGD iteratively conducts parameter estimation for each training example, which helps it to adequately handle large-scale datasets efficiently (Kumar et al. 2023). SGDRegressor treats one sample at a time, making it efficient. Thus, the speed of SGDRegressor is one of its key features. Moreover, SGDRegressor incorporates various loss functions to providing flexibility in handling different types of regression problems. In addition, it has several regularization techniques to prevent overfitting (Erdal & Karakurt 2013). Finally, these features make it a versatile ML for a wide range of applications.

The Mann–Kendall test, which is a nonparametric test, was first presented by Mann and then the statistical distribution of the test was extracted by Kendall. This test is widely used to find trends in meteorological time series data (Modarres & da Silva 2007; Goodarzi et al. 2022a, b). Due to the capabilities of this test in revealing the changes in the time series of climatic variables, it has been widely used in climate change studies. This test examines the null hypothesis of no trend in time series. In this regard, it also identifies sudden change points.

In essence, it consists of two hypotheses: (i) null hypothesis and (ii) alternative hypothesis. The former denotes the lack of pattern or trend in the data series, while the latter suggests the presence of a trend in the data series. For more information on this trend analysis method, interested readers are referred to previous studies (Hırca et al. 2022; Niazkar et al. 2023).

The implementation of analysis in this test is attributed to the diagrams UI and U′I. Both graphs originate from a specific point. Moreover, when |U| is greater than 1.96, it indicates that the UI graph intersects the two lines Y = 1.96 and Y = −1.96, signifying a significant trend. If U is greater than 0, or the overall trend of the UI graph is upward, it discloses a meaningful upward trend. In the case where U is less than 0, the trend is both significant and decreasing. Furthermore, if the UI and U′I graphs intersect within the range of −1.96 to 1.96, it indicates a sudden change.

Various metrics were employed to assess the performance of ML-based downscaling models: root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), the coefficient of determination (R²), and bias percentage (PBIAS). The metrics are displayed in Table 2, where T_obs and T_est represent the daily observed and estimated temperatures, respectively, and N denotes the number of data points.

Table 2

Model evaluation metrics

Measure .	Equation .	Range .	Optimal value .
RMSE		[0, ∞]	0
NSE		[−∞, 1]	1
R2		[0, 1]	1
PBIAS		[−∞, ∞]	0

Measure .	Equation .	Range .	Optimal value .
RMSE		[0, ∞]	0
NSE		[−∞, 1]	1
R2		[0, 1]	1
PBIAS		[−∞, ∞]	0

View Large

Referring to the optimal values listed in Table 2 for each metric, a greater R² value indicates a stronger correlation between the simulation and observed data. Unlike R², a higher value of RMSE signifies a larger deviation between simulated and observed data. A higher NSE indicates more accurate model performance. Finally, the ideal PBIAS value is 0. Negative and positive PBIAS values signify underestimation and overestimation, respectively.

The results indicate that LightGBM and KNN have little dispersion for training data and they are close to the observed data line, while DTR displays greater dispersion among all. According to the results of testing data, the lowest dispersion is related to LightGBM and KNN and the highest dispersion is related to DTR. Clearly, scattering of data reduces the accuracy of the results.

According to Figure 8(a), the annual changes for 1996–2001 and 2007–2011 showed a negative and decreasing trend, while a positive and increasing trend was observed for 2002–2007 and 2011–2014, with sudden changes occurring in 1996, 2006, and 2008. Based on Figure 8(b), the predicted annual average temperature changes under the SSP2-4.5 scenario indicate a positive and increasing trend, with sudden changes in 2016, 2019, 2021, 2022, and 2025. Similarly, Figure 8(c), associated with the SSP5-8.5 scenario, shows a positive and upward trend from 2015 to 2020 and from 2025 to 2045. Based on Kendall's test, the intersection of the UI and U′I graphs between −1.96 and 1.96 indicates sudden shifts during the period of 2015–2019. These statistical changes suggest that the trends during this period are distinct from others, requiring further investigation into the factors driving these fluctuations.

In this research, a comparison was made between the observed data and the estimated data under both climate scenarios during 2015–2022 for Nazmakan station. The evaluation was calculated based on R² and NSE, which is shown in Table 3. KNN obtained the most valid performance with R² and NSE of about 0.92 in both scenarios and in other models, between 0.88 and 0.91. In general, it can be concluded that ML-based methods were acceptably precise for downscaling daily temperature.

In this study, six ML methods, namely, KNN, SVR, DTR, ANNs, LightGBM, and SGDRegressor, were employed to downscale the daily temperature of the Nazmakan station for the period from 1995 to 2015. Training and testing of ML-based downscaling models were conducted based on the observed daily temperatures, allowing the outputs of four CMIP6 climate models to be effectively downscaled. The results, evaluated using Taylor diagrams, NSE, and PBIAS, demonstrated that all models performed similarly well for both training and testing data. Based on the study's metrics, all six ML models performed acceptably, with LightGBM achieving the best results in the training dataset, while both KNN and LightGBM performed the best in the testing dataset. The SA indicated that the models were more sensitive to the MPI-ESM1-2-HR climate model. Furthermore, the six ML models were utilized to predict daily temperatures for the future period of 2015–2045 under the SPP2-4.5 and SPP5-8.5 climate scenarios. The results of the Mann–Kendall test illustrated a decreasing trend in the early observed temperatures, followed by a positive and increasing trend. The estimated temperatures also showed an upward trend in most forthcoming years. According to the future projections, there will be a reduction in the daily temperature range compared to the historical period, leading to cooler summers and warmer winters. The evaluation of the performance of ML-based models was supported by the overlapping of available observed and estimated data from 2015 to 2022, which yielded reliable results. Overall, this study provides significant contributions to the field by demonstrating the effectiveness of ML methods in downscaling climate data. Specifically, the KNN and LightGBM methods exhibited the best performance in simulating daily temperatures and showed strong capabilities in performing nonlinear regression.

To build upon the findings of this study, future research should explore advanced ML techniques, such as ensemble learning and deep learning, to improve downscaling accuracy. Comprehensive validation against observational data is essential for ensuring reliability. Focusing on underrepresented regions and climate variables will provide an enhanced understanding of climate trends. In addition, models must efficiently handle larger datasets. Ensuring that these models can generalize across diverse climatic conditions will strengthen climate predictions and risk assessments.

MG, ZH, AS, and MN conceptualized the study and performed the validation. ZH, AS, and MN contributed to the methodology and formal analysis. AS and MN contributed to software analysis. ZH performed the investigation. ZH and AS wrote the original draft of the article. MG and MN contributed to reviewing and editing the writing of the article, supervision, and project administration.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.