A multimodel ensemble machine learning approach for CMIP6 climate model projections in an Indian River basin

Abstract

Multimodel ensemble (MME) approach would help modellers to know the advantages of individual global circulation models (GCMs) and to avoid the weaknesses associated with them, and it would help the river basin modellers to make appropriate modelling decisions. The study highlights the river basinscale development of MME as a convenient way to reduce the parameter and structural uncertainties in the Coupled Model Intercomparison Project Phase 6 (CMIP6) GCMs simulations after identifying the best five CMIP6 GCMs based on the rating metric calculations. Furthermore, the performance of the MME was enhanced by integrating three machine learning algorithms (artificial neural network [ANN], random forest [RF], support vector machine [SVM]). Subsequently, comparative assessment depicted the improved performance in MME-integrated ML algorithms compared to simple arithmetic mean (SAM) in simulating observed precipitation (P), maximum temperature (T max ), and minimum temperature (T min ) over the Damodar River basin (DRB), India. The statistical metrics indicate that the SVM and RF methods yielded better results than SAM and ANN methods, thus selected for future projections. The robustness of the MME-RF and MME-SVM approach has also been observed while capturing the spatial pattern as IMD-observed with well representation of climate indices for both wet and dry seasons. Future projections with MME-SVM and MME-RF suggested a possible rise in mean annual P in the range of 1.4-15% and 6.8-39% with an increasing trend in temperature (T max , T min ) under the SSP245 and SSP585 scenarios, respectively. Replicating the spatial pattern of the future climatic variables projections evinced a warmer and drier climate in the southwest part of the DRB for both SSP scenarios during wet and dry season and thence warned a probable drier condition on the southwest part of the DRB in future time slices.

The effectiveness of machine learning‐based multi‐model ensemble predictions of CMIP6 in Western Ghats of India

“…There have been limited studies on AdaBoost and XGBR ML‐based ensembles. Most of the previous studies endorse the use of RFR and SVR in the ensemble of precipitation and temperature (Ahmed et al, 2019, 2020; Dey et al, 2022; Jose et al, 2022; Li et al, 2021; Yang et al, 2022). Asadollah et al (2022) has proved the efficacy of the Gradient Boosting Regression Tree (GBRT) in the downscaling of GCMs over RFR and SVR, XGBR in the land degradation study (Saha et al, 2022), and drought‐maize yield dynamics study (Muthuvel et al, 2023) using CMIP6 GCMs.…”

“…Also, the least ranked GCMs obtained in this study are in consonance with the observation made by Aadhar and Mishra (2020) and Mitra (2021). Over the river basin in Eastern India, the MPI‐ESM‐1‐2‐HR is found as a top‐performing GCM for precipitation, maximum, and minimum temperatures (Dey et al, 2022). Though the finding of top‐performing models is consistent with previous studies, there is a wide variability in the performance of the GCMs over the study area.…”

“…The most commonly used models are the Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANNs), K‐nearest neighbour (KNN), Relevance Vector Machine (RVM), which can model complex, often nonlinear relationships in climate data. Multiple studies have accounted for the machine learning approach for the ensemble of precipitation, temperature and evaporation (Ahmed et al, 2020; Dey et al, 2022; Jose et al, 2022; Kadkhodazadeh et al, 2022; Wang et al, 2018) and found the best performance of RF and SVM in most of the studies over Southwest Asia, Australia (Wang et al, 2018), KNN and RVM over SVM and ANN in Pakistan (Ahmed et al, 2020), RF over decision tree, Adaptive boosting and linear regression over China (Li et al, 2021; Yang et al, 2022). Also, the studies have endorsed the best performance of RF and SVM over ANN and AM methods in a tropical monsoon climate in eastern India (Dey et al, 2022), and RF in the per‐humid river basin in south India (Jose et al, 2022).…”

“…Multiple studies have accounted for the machine learning approach for the ensemble of precipitation, temperature and evaporation (Ahmed et al, 2020; Dey et al, 2022; Jose et al, 2022; Kadkhodazadeh et al, 2022; Wang et al, 2018) and found the best performance of RF and SVM in most of the studies over Southwest Asia, Australia (Wang et al, 2018), KNN and RVM over SVM and ANN in Pakistan (Ahmed et al, 2020), RF over decision tree, Adaptive boosting and linear regression over China (Li et al, 2021; Yang et al, 2022). Also, the studies have endorsed the best performance of RF and SVM over ANN and AM methods in a tropical monsoon climate in eastern India (Dey et al, 2022), and RF in the per‐humid river basin in south India (Jose et al, 2022). The combination of multiple weak learners in ensemble techniques in machine learning has been theoretically and empirically expected to perform better than the single weak learner's algorithms in handling complex regression and classification problems (Caves & Caves, 1996).…”

Bias correction of CMIP6 simulations of precipitation over Indian monsoon core region using deep learning algorithms

“…It was reported that the machine learning techniques are useful in improving the predictions of multiple model ensemble daily precipitation and temperature. In a recent study, Dey et al (2022) have reported improvement of performance of Multimodel Ensembles by integrating machine learning algorithms, such as ANN, Random Forest, and Support Vector Machine. Zhang et al (2022) applied three machine learning algorithms, Ordinary Least Squares regression, Decision Tree, and Deep Neural Network, to the precipitation and temperature ensemble.…”