Statistical Downscaling Model Research Articles

Research on the future climate change and runoff response in the mountainous area of Yongding watershed

Climate change has affected hydrological processes at global scale, and redistributed water resources in space and time. Anthropogenic activities even exacerbated the water resources problems. In this paper, Statistical DownScaling Model (SDSM) was used to downscale the predicted precipitation and temperature in the mountainous area of Yongding watershed under 4 SSP-RCP climate scenarios in 4 General Circulation Models (GCMs). Bayesian model averaging (BMA) method was employed for multi-model integration to reduce the impact of GCM uncertainty on climate change prediction. Then, the multi-model ensemble data were corrected by the Quantile mapping (QM) method. Soil and Water Assessment Tool (SWAT) was established in the study area considering the intensive human activities, including the water use in irrigation areas and regulation of the 8 large and medium-sized reservoirs. The SWAT model was applied to simulate the runoff response under the future climate change. The results showed that the maximum temperature (Tmax), minimum temperature (Tmin), precipitation (P) and runoff all show an upward trend under the four future climate scenarios. SSP585 scenario always showed the highest increase in temperature and precipitation. The simulated future annual runoff increases significantly compared with the reference period, indicating that the water shortage of the watershed may be alleviated in the future. The research results can provide a scientific basis for the decision-making of water resources management and protection in the watershed.

Development of statistical downscaling methods for the assessment of rainfall characteristics under climate change scenarios

Abstract The objective of this research was to develop a statistical downscaling approach in the Phetchaburi River Basin, Thailand, consisting of two main processes: predictor selection and relationship construction between predictors and local rainfall. Super predictor (SP) and stepwise regression (SR) were employed with principal component analysis (PCA) and the statistical downscaling model (SDSM) was applied later. Four statistical models were finally used: M1 (SP–PCA), M2 (SP–SDSM), M3 (SR–PCA), and M4 (SR–SDSM) with 26 large circulation indices generated by CanESM2 (CMIP5). Finally, the characteristics of extreme rainfall events were observed under three climate change scenarios during three different periods (2020–2040, 2041–2070, and 2071–2100). The results revealed that rainfall and geostrophic airflow velocity were the best predictors for rainfall downscaling, followed by divergence, meridional velocity, and relative humidity. Three objective functions (R2, NSE, and RMSE) were applied to evaluate model performance. The M4 model presented the highest performance while M1 showed the lowest skill. The average annual rainfall specifically increased compared with the historical rainfall for RCP2.6, RCP4.5, and RCP8.5 scenarios in future periods. Very to extremely wet years determined by the standardized anomaly index (SAI) occurred more often in the far-future while severe to extremely dry years frequently occurred in the mid- and far-future.

Comparison of machine learning statistical downscaling and regional climate models for temperature, precipitation, wind speed, humidity and radiation over Europe under present conditions

AbstractThere are two main approaches to downscale global climate projections: dynamical and statistical downscaling. Both families have been widely evaluated, but intercomparison studies between the two families are scarce, and usually limited to temperature and precipitation. In this work, we present a comparison between a statistical downscaling model (SDM) based on machine learning and six regional climate models (RCMs) from EURO‐CORDEX, for five variables of interest: temperature, precipitation, wind, humidity and solar radiation under present climatic conditions. The study is conducted at a continental scale over Europe, with a spatial resolution of 0.11° and daily data. Both the SDM and the RCMs are driven by the ERA‐Interim reanalysis, and observations are taken from the gridded dataset E‐OBS. Several aspects have been evaluated: daily series, mean values and extremes, spatial patterns and also temporal aspects. Additionally, a multivariable index (fire weather index) derived from the fundamental variables has been included. The SDM has better scores than the RCMs for all the evaluated metrics with only a few exceptions, mainly related to an underestimation of the variance. After bias correction, both the SDM and the six RCMs present similar results, with no significant differences among them. Results presented here, combined with the low computational expense of SDMs and the limited availability of RCMs over some CORDEX domains, should motivate the consideration of statistical downscaling at the same level as RCMs by official providers of regional information, and its inclusion in reference sites. Nonetheless, further analysis on crucial aspects such as the impact on long‐term trends or the sensitivity of different methods to being driven by global climate models, is needed.

Projection of hydrological responses to changing future climate of Upper Awash Basin using QSWAT model

BackgroundProjecting future streamflow variation or the hydrological impact of climate change plays a pivotal role in the sustainable implication of planning water resources management. Therefore, this study predicts the potential of climate change’s impact on hydrological components in the Upper Awash Basin (UAB). The study applied a statistical downscaling model (SDSM) to generate future high-resolution climate data from the climate model output of the Canadian Second Generation Earth System Model (CanESM2) and the National Centers for Environmental Prediction (NCEP) under the representative concentration pathways (RCP4.5 and RCP8.5) scenarios. To analyze the trend of future rainfall and temperature, non-parametric Mann-Kendall, Modified Mann-Kendall tests, Sen’s slope estimator, and changing point (Pettit) tests were used. The output of downscaled climate data is used as input to a calibrated and validated Soil and Water Assessment Tool (QSWAT) model to assess the impact of future climate change on UAB hydrology.ResultsThe results show that annual rainfall and temperature are significantly increased (p < 0.05) in the UAB under RCP4.5 and 8.5 for the model ensemble mean for both short- and long-term scenarios. The change in the rainfall, the maximum and minimum temperature is mostly visible in the second period (the 2060s). Climate change is likely to cause persistent decreases in surface runoff (SUR_Q) and increases in actual evapotranspiration (ET) under all climate scenarios in the three periods. Reduction in SUR_Q despite an increase in rainfall could be due to an increment in both temperature and ET. The study also identified inconsistent seasonal changes in projected future precipitation that considerably impact overall climatic conditions.ConclusionsThis research is essential to develop an interdisciplinary approach that integrates environmental policies for the coherent use and management of water resources for future climate change and ecological protection in the basin, including other similar basins.

Assessment of drought variability using SPEI under observed and projected climate scenarios over Uttar Pradesh, India

Drought is a widely occurring natural extreme that causes more severe negative impacts in the continuation of changing climate. This study presents the assessment of drought variability based on the standardized precipitation evapotranspiration index (SPEI) under historic and projected future climate scenarios. A probabilistic approach run theory is applied to define drought based on SPEI drought events, severity, intensity, and frequency according to the threshold of SPEI ≤ −1. The Statistical Downscaling Model (SDSM) was employed to downscale the meteorological variables. Drought variability under projected climate change is evaluated according to three scenarios: representative concentration pathways RCP 2.6, RCP 4.5, and RCP 8.5 using CanESM2 of the CMIP5 model for 32 years from 2019 to 2050. A drought event with maximum severity with a higher number of drought events associated with RCP 4.5 and RCP 8.5 among three scenarios, whereas synoptic stations Kanpur, Moradabad, Azamgarh, and Chitrakoot exhibit drought with higher maximum severity and frequency occurred under RCP 2.6. The Mann-Kendall (MK) test was used to assess the change in climatic variables and SPEI time series observed period. Results reveal a significant decrease in the trend of rainfall during monsoon season as well as annually observed at 16 locations except for Lucknow and Meerut and 83% of the stations exhibit a positive trend of SPEI at each timescale, which demonstrates most of the study regions exposed to the risk of drought, and locations namely Saharanpur, Meerut, and Aligarh, showed a positive trend, indicating a decrease in drying conditions over historic period.

Impact of Future Climate and Land Use Changes on Runoff in a Typical Karst Basin, Southwest China

Climate change and land use change are the two main factors affecting the regional water cycle and water resources management. However, runoff studies in the karst basin based on future scenario projections are still lacking. To fill this gap, this study proposes a framework consisting of a future land use simulation model (FLUS), an automated statistical downscaling model (ASD), a soil and water assessment tool (SWAT) and a multi-point calibration strategy. This frameword was used to investigate runoff changes under future climate and land use changes in karst watersheds. The Chengbi River basin, a typical karst region in southwest China, was selected as the study area. The ASD method was developed for climate change projections based on the CanESM5 climate model. Future land use scenarios were projected using the FLUS model and historical land use data. Finally, the SWAT model was calibrated using a multi-site calibration strategy and was used to predict future runoff from 2025–2100. The results show that: (1) the developed SWAT model obtained a Nash efficiency coefficient of 0.83, which can adequately capture the spatial heterogeneity characteristics of karst hydro-climate; (2) land use changes significantly in all three future scenarios, with the main phenomena being the interconversion of farmland and grassland in SSPs1-2.6, the interconversion of grassland, farmland and artificial surfaces in SSPs2-4.5 and the interconversion of woodland, grassland and artificial surfaces in SSPs5-8.5; (3) the average annual temperature will show an upward trend in the future, and the average annual precipitation will increase by 11.53–14.43% and (4) the future annual runoff will show a significant upward trend, with monthly runoff mainly concentrated in July–September. The variability and uncertainty of future runoff during the main-flood period may increase compared to the historical situation. The findings will benefit future water resources management and water security in the karst basin.

Evaluating the Feedback of the Reservoir Methane Cycle to Climate Warming under Hydrological Uncertainty

Freshwater reservoirs are widely recognized as methane (CH4) emission hotspots. Existing research has shown that temperature and hydrological conditions significantly affect wetland CH4 cycling processes. However, the feedback of the CH4 cycle to climate warming remains unclear for deep reservoirs where seasonal water thermal stratification exists. This study combined a reservoir CH4 cycling model and a Statistical DownScaling Model (SDSM) to evaluate reservoir CH4 cycling feedbacks under multiple climate change scenarios while accounting for hydrological uncertainty. Daily air temperatures in 2100 were predicted by the combination of the CanESM5 model and a SDSM. To address hydrological uncertainty, we selected three representative hydrological years (i.e., wet, normal, and dry) to create hydrological scenarios. Results showed that annual sediment CH4 production increased with warming, ranging 323.1–413.7 × 103 t C year−1 among multiple scenarios. Meanwhile, the CH4 oxidation percentage decreased with warming, which meant warming promoted sediment CH4 release non-linearly; 67.8–84.6% of sediment ebullient flux was ultimately emitted to the atmosphere (51.3–137.7 × 103 t C year−1), which showed ebullition was the dominant emission pathway. Higher air temperatures and drier conditions generally promote reservoir emissions. This study is helpful for predicting reservoir emissions while directing decision-making for reservoir sustainability.

Generating high-resolution climate maps from sparse and irregular observations using a novel hybrid RBF network

ABSTRACT Sparse and irregular climate observations in many developing countries are not enough to satisfy the need of assessing climate change risks and planning suitable mitigation strategies. The wide-used statistical downscaling model (SDSM) software tools use multi-linear regression to extract linear relations between large-scale and local climate variables and then produce high-resolution climate maps from sparse climate observations. The latest machine learning techniques (e.g. SRCNN, SRGAN) can extract nonlinear links, but they are only suitable for downscaling low-resolution grid data and cannot utilize the link to other climate variables to improve the downscaling performance. In this study, we proposed a novel hybrid RBF (Radial Basis Function) network by embedding several RBF networks into new RBF networks. Our model can well incorporate climate and topographical variables with different resolutions and extract their nonlinear relations for spatial downscaling. To test the performance of our model, we generated high-resolution precipitation, air temperature and humidity maps from 34 meteorological stations in Bangladesh. In terms of three statistical indicators, the accuracy of high-resolution climate maps generated by our hybrid RBF network clearly outperformed those using a multi-linear regression (MLR), Kriging interpolation or a pure RBF network.

Downscaling of GPM satellite precipitation products based on machine learning method in complex terrain and limited observation area

Precipitation products with high spatial and temporal resolutions are crucial in hydrological monitoring, meteorological forecasting, and disaster warning. However, obtaining such products in regions with complex terrain and limited observation conditions is a challenging task. To address this issue, a statistical downscaling model for satellite precipitation products using an optimized Back Propagation (BP) neural network was developed. The model employed key environmental variables selected by the Mean Impact Value (MIV) index, including enhanced vegetation (EVI), land surface temperature (LST_D, LST_N), evapotranspiration (ET), digital elevation model (DEM), slope, and latitude and longitude (Lon, Lat). The downscaled precipitation products were corrected twice using a combination of the model error correction and observation data merging methods. Finally, the satellite precipitation products with a resolution of 0.1° (∼10 km) were downscaled to 1 km. A comprehensive evaluation of the results showed that: (1) the accuracy of the precipitation products obtained using this method was better than that of the original satellite precipitation products; (2) the method effectively supplemented precipitation data in areas with few ground stations, with better results in areas with a dense distribution of ground stations; (3) this method demonstrated the efficacy of satellite precipitation products for downscaling in areas with complex terrain.

EVALUATION OF SDSM MODELS FOR CLIMATE PREDICTIONS IN BANGLADESH

As one of the poorest countries in the world, agriculture is Bangladesh’s economic pillar, leading to Bangladesh’s economy becoming vulnerable to global warming. The generation of high-resolution climate predictions in Bangladesh can help to reduce the huge damage and losses inflicted by disasters linked to climate. The statistical downscaling model (SDSM) is the most widely used software on robust climate downscaling and prediction analysis. In this study, by using the SDSM model, we established the statistical relationship between observed climate data in Bangladesh and the large-scale low-resolution NCEP data and used three statistical indicators to evaluate the prediction performance of the SDSM software. Our results show that the SDSM software is more suitable for forecasting humidity/temperature in Bangladesh than rainfall.

A Statistical Downscaling Model Based on Multiway Functional Principal Component Analysis for Southern Australia Winter Rainfall

Abstract We propose a statistical downscaling model based on multiway functional principal component analysis (FPCA) for rainfall prediction. The model mainly explains the relationship between the winter mean sea level pressure (MSLP) and rainfall in southern Australia from the perspective of functional data. In comparison with the traditional approach of feature extraction based on principal component analysis, the multiway FPCA needs fewer principal components not only to capture the most variance in MSLP but also to greatly avoid the loss of spatial information. A functional principal component (FPC) regression is further developed to simulate both current and future rainfall. The main results show that the first five leading FPCs are sufficient to capture the spatial characteristics of winter MSLP, achieving the purpose of efficient dimensionality reduction. Specifically, no more than three FPCs are required to develop the functional downscaling models for the winter rainfall over four studied regions. The functional downscaling model provides a good skill in terms of the correlation higher than 0.7 between the predictions and observations and the ratio of root-mean-square error to the climatology of winter rainfall below 20% over four regions. The developed downscaling models are further used to interpret the MSLP patterns from four CMIP5 climate models [ACCESS1.3, BCC_CSM1.1(m), CESM1(CAM5), and MPI-ESM-MR], which have been used to simulate both present-day and future climate. The resulting downscaled values based on ensemble MSLP provide 1) a closer representation of observed present-day rainfall than the raw climate model values and 2) alternative estimates of future changes in rainfall that arise from changes in MSLP.

Generation of climate change scenarios for rainfall and temperature using SDSM in a Mediterranean environment: a case study of Boufakrane river watershed, Morocco

Climate change is the global phenomenon of climate transformation characterized by changes in the planet’s habitual climate. The main objective of this work is to assess and predict climate change for the 2015–2100 period, in the Boufakrane river watershed, Morocco. In this study, the Statistical DownScaling Model (SDSM) method has been used to generate the climate scenarios for rainfall and temperature related to Representative Concentration Pathway (RCP) scenarios such as RCP 2.6, RCP 4.5, and RCP 8.5. For this purpose, the region’s annual, monthly, and daily rainfall records were analyzed. A rainfall index was calculated to differentiate and distinguish between dry and humid years of the climate series, and a non-parametric Pettitt’s test has been applied to determine the trend of precipitation series for the projection period. The application of Pettitt’s test with a confidence interval of 99% and a significance level of 0.05 shows a break in the 1979/80 rainfall series, with an average of 675.78 mm and 511.94 mm respectively before and after the break. The combined application of the two approaches of SDSM, and the non-parametric Pettitt test for the period 2015–2100, showed a downward trend of annual rainfall with 17.29% for RCP 2.6. While no trend was recorded for the two scenarios RCP 4.5, and RCP 8.5. Therefore, this study highlighted the importance and urgent need for developing strategies and decision-making plans for climate change adaptation for sustainable water resources management.Graphical

Evolution of Drought Trends under Climate Change Scenarios in Karst Basin

Karst basins have a relatively low capacity for water retention, rendering them very vulnerable to drought hazards. However, karst geo-climatic features are highly spatially heterogeneous, making reliable drought assessment challenging. To account for geo-climatic heterogeneous features and to enhance the reliability of drought assessment, a framework methodology is proposed. Firstly, based on the history of climate (1963–2019) from the Global Climate Model (GCM) and station observations within the Chengbi River karst basin, a multi-station calibration-based automated statistical downscaling (ASD) model is developed, and the Kling–Gupta efficiency (KGE) and Nash–Sutcliffe efficiency (NSE) are selected as performance metrics. After that, future climate (2023–2100) under three GCM scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5) are obtained by using the ASD model. Finally, the Standardized Precipitation Evapotranspiration Index (SPEI), calculated by future climate is applied to assess drought conditions. The results indicate that the multi-station calibration-based ASD model has good performance and thus can be used for climate data downscaling in karst areas. Precipitation mainly shows a significant upward trend under all scenarios with the maximum variation (128.22%), while the temperature shows a slow upward trend with the maximum variation (3.44%). The drought condition in the 2040s is still relatively severe. In the 2060s and 2080s, the basin is wetter compared with the historical period. The percentage of drought duration decreases in most areas from the 2040s to the 2080s, demonstrating that the future drought condition is alleviated. From the SSP1-2.6 scenario to the SSP5-8.5 scenario, the trend of drought may also increase.

Impacts of climate and land use/cover change on mini-hydropower generation in River Kyambura watershed in South Western part of Uganda

Abstract This study explored the combined impacts of climate and land-use change on mini-hydropower generation in the Kyambura watershed. The soil and water assessment tool (SWAT) was used as a hydrological model whereas the statistical downscaling model (SDSM) was used to downscale meteorological data for the Kyambura watershed for the year 2050. The results show that there will be an increase in urban land by 11.89%, barren land by 25.78%, water by 0.49% and a reduction in percentage area coverage of vegetation by 38.17% by the year 2050. A 10.6 and 17.7% increase is anticipated in average annual hydropower generated by the year 2050. There is, therefore, a need to develop governing policies to regulate management practices to preserve the integrity of the watersheds and ensure the reliability of power production.

Rainfall Prediction in Statistical Downscaling Using Tweedie Compound Response and Lasso Penalty

Statistical Downscaling (SD) is a technique in climatology to analyze the relationship between large-scale (global) data and small-scale (local) data using statistical modeling. The SD technique is used to overcome the inability of global scale data in the form of the General Circulation Model (GCM) as a low resolution predictor to predict local scale climatic conditions in the form of high resolution rainfall as a direct response. Rainfall consists of two components, namely continuous and discrete. The continuous component describes the intensity of rainfall while the discrete component describes the occurrence of rain. both components have an important role in predicting rainfall so it is necessary to choose the right distribution. One distribution that is able to handle both rain components is the mixed Tweedie distribution, namely the Gamma and Poisson distribution, hereinafter referred to as the Tweedie compound. GCM generally has multicollinearity problems in SD modeling. This can be handled using the Lasso penalty. This study aims to predict rainfall and rainfall events by taking into account the multicollinearity problem in the model for locations on different plains. Based on the research results, it was found that Cigugur Station from the highland gets the smallest RMSEP value and the biggest r-correlation. This model is not good enough to use for moderate plains rainfall data.

Mapping and predicting of water erosion using RUSLE in Mediterranean context: Case of El Kharroub watershed (Western Rif, Northern Morocco)

Worldwide, soil erosion is one of the hazards posing a serious threat to soil and water resources, especially in the Mediterranean context. In Morocco, this phenomenon is a major problem in the natural territory, especially in the Rif Mountains in the north of the country, where fragile rocks dominate on steep slopes. Thus, this work aims to assess the soil erosion in the El Kharroub River watershed over the baseline period 2000-2020 and two future periods 2021-2030 and 2031-2050, using the Revised Universal Soil Loss Equation (RUSLE). For the future climate periods (2021-2030 and 2031-2050), precipitations were produced using a classical statistical downscaling model (SDSM). Over the current period, the results showed that the annual rainfall erosivity and the vegetation cover decreased from 2002 to 2020 by 34.3% and 28.6% respectively. The annual soil loss maps showed a decline of about 54.8% during the baseline period. The changes in rainfall and vegetation cover are largely due to climate changes effect and the deforestation/reforestation that the region has experienced, which subsequently leads to changes in soil erosion due to the important function of these two factors. Furthermore, projected scenarios revealed that the average annual erosivity could decrease to 268.4 and 267.1 for 2031-2050 and 2031-2050 scenarios compared with the current period. As a result, the average annual soil losses could decrease by 21.3 and 21.8% for the projected scenarios.

Rainfall-runoff modelling based on global climate model and tropical rainfall measuring mission (GCM -TRMM): A case study in Hulu Terengganu catchment, Malaysia

The hydropower Plant in Terengganu is one of the major hydroelectric dams currently operated in Malaysia. For better operating and scheduling, accurate modelling of natural inflow is vital for a hydroelectric dam. The rainfall-runoff model is among the most reliable models in predicting the inflow based on the rainfall events. Such a model's reliability depends entirely on the reliability and consistency of the rainfall events assessed. However, due to the hydropower plant's remote location, the cost associated with maintaining the installed rainfall stations became a burden. Therefore, the study aims to create a continuous set of rainfall data before, during, and after the construction of a hydropower plant and simulate a rainfall-runoff model for the area. It also examines the reliability of alternative methods by combining rainfall data from two sources: the general circulation model and tropical rainfall measuring mission. Rainfall data from ground stations and generated data using inverse distance weighted method will be compared. The statistical downscaling model will obtain regional rainfall from the general circulation model. The data will be divided into three stages to evaluate the accuracy of the models in capturing inflow changes. The results revealed that rainfall data from TRMM is more correlated to ground station data with R2 = 0.606, while SDSM data has R2 = 0.592. The proposed inflow model based on GCM-TRMM data showed higher precision compared to the model using ground station data. The proposed model consistently predicted inflow during three stages with R2 values ranging from 0.75 to 0.93.

STATISTICAL DOWNSCALING MODEL WITH PRINCIPAL COMPONENT REGRESSION AND LATENT ROOT REGRESSION TO FORECAST RAINFALL IN PANGKEP REGENCY

Climate information, especially rainfall, is needed by various sectors in Indonesia, including the marine and fisheries sectors. Estimating high-resolution climate models continues to develop by involving global-scale climate variables, one of which is the global circulation model (GCM) output precipitation. Statistical downscaling (SD) relates global scale climate variables to local scales. Principal component regression (PCR) and latent root regression (LRR) techniques are statistical methods used in the SD model to overcome the high correlation between GCM data grids. PCR focuses on the variability in the predictor variables, while the LRR focuses on the variability between the response variables and predictors. This method was applied to Pangkep Regency rainfall data as a local scale response variable and GCM precipitation as a predictor variable (January 1999 to December 2020). This study aimed to obtain the number of principal component (PC) in the SD model and the forecast value of the 2020 rainfall data. In addition, the dummy variable resulting from K-means was used as a predictor variable in PCR and LRR. The result is that using the first 11-15 PC has a cumulative diversity proportion of 98%. Furthermore, by using the data for the 1999-2019 period, adding a dummy variable to the PCR can increase the accuracy of the model (the coefficient of determination is 92.27%-92.43%). However, LRR with and without dummy variables produces relatively the same model accuracy. In general, the LRR model is better at explaining the diversity of the Pangkep District rainfall data than the PCR model. The prediction of rainfall for the 2020 period at LRR with 13 PC is an accurate prediction based on the highest correlation value (0.97) and the lowest root mean square error prediction (75.17).

Downscaling long lead time daily rainfall ensemble forecasts through deep learning

Skilful and localised daily weather forecasts for upcoming seasons are desired by climate-sensitive sectors. Various General circulation models routinely provide such long lead time ensemble forecasts, also known as seasonal climate forecasts (SCF), but require downscaling techniques to enhance their skills from historical observations. Traditional downscaling techniques, like quantile mapping (QM), learn empirical relationships from pre-engineered predictors. Deep-learning-based downscaling techniques automatically generate and select predictors but almost all of them focus on simplified situations where low-resolution images match well with high-resolution ones, which is not the case in ensemble forecasts. To downscale ensemble rainfall forecasts, we take a two-step procedure. We first choose a suitable deep learning model, very deep super-resolution (VDSR), from several outstanding candidates, based on an ensemble forecast skill metric, continuous ranked probability score (CRPS). Secondly, via incorporating other climate variables as extra input, we develop and finalise a very deep statistical downscaling (VDSD) model based on CRPS. Both VDSR and VDSD are tested on downscaling 60 km rainfall forecasts from the Australian Community Climate and Earth-System Simulator Seasonal model version 1 (ACCESS-S1) to 12 km with lead times up to 217 days. Leave-one-year-out testing results illustrate that VDSD has normally higher forecast accuracy and skill, measured by mean absolute error and CRPS respectively, than VDSR and QM. VDSD substantially improves ACCESS-S1 raw forecasts but does not always outperform climatology, a benchmark for SCFs. Many more research efforts are required on downscaling and climate modelling for skilful SCFs.

A Comparison of the Statistical Downscaling and Long-Short-Term-Memory Artificial Neural Network Models for Long-Term Temperature and Precipitations Forecasting

General circulation models (GCMs) run at regional resolution or at a continental scale. Therefore, these results cannot be used directly for local temperatures and precipitation prediction. Downscaling techniques are required to calibrate GCMs. Statistical downscaling models (SDSM) are the most widely used for bias correction of GCMs. However, few studies have compared SDSM with multi-layer perceptron artificial neural networks and in most of these studies, results indicate that SDSM outperform other approaches. This paper investigates an alternative architecture of neural networks, namely the long-short-term memory (LSTM), to forecast two critical climate variables, namely temperature and precipitation, with an application to five climate gauging stations in the Lake Chad Basin. Lake Chad is a data scarce area which has been impacted by severe drought, where water resources have been influenced by climate change and recent agricultural expansion. SDSM was used as the benchmark in this paper for temperature and precipitation downscaling for monthly time–scales weather prediction, using grid resolution GCM output at a 5 degrees latitude × 5 degrees longitude global grid. Three performance indicators were used in this study, namely the root mean square error (RMSE), to measure the sensitivity of the model to outliers, the mean absolute percentage error (MAPE), to estimate the overall performance of the predictions, as well as the Nash Sutcliffe Efficiency (NSE), which is a standard measure used in the field of climate forecasting. Results on the validation set for SDSM and test set for LSTM indicated that LSTM produced better accuracy on average compared to SDSM. For precipitation forecasting, the average RMSE and MAPE for LSTM were 33.21 mm and 24.82% respectively, while the average RMSE and MAPE for SDSM were 53.32 mm and 34.62% respectively. In terms of three year ahead minimum temperature forecasts, LSTM presents an average RMSE of 4.96 degree celsius and an average MAPE of 27.16%, while SDSM presents an average RMSE of 8.58 degree celsius and an average MAPE of 12.83%. For maximum temperatures forecast, LSTM presents an average RMSE of 4.27 degree celsius and an average MAPE of 11.09 percent, while SDSM presents an average RMSE of 9.93 degree celsius and an average RMSE of 12.07%. Given the results, LSTM may be a suitable alternative approach to downscale global climate simulation models’ output, to improve water management and long-term temperature and precipitations forecasting at local level.