Reconstructing terrestrial water storage anomalies using convolution-based support vector machine

Study region: The Congo River basin in west-central Africa Study focus: Traditional machine learning algorithms are recently being replaced by integrated learning techniques in pattern recognitions and predictions. These updated tools (or techniques), which attempt to explore higher dimensions and uncover hidden patterns in considerably non-linear datasets are the new normal in small, medium and large dataset even in higher orders. This study investigates the performance of a convolution-based support vector machine in hydrological analysis to optimize forecast of terrestrial water storage anomalies (TWSA). The linear, polynomial, and radial basis function (RBF) kernels were explored in reconstructing the three variants of TWSA obtained from the Gravity Recovery and Climate Experiment (GRACE) level 3 products. Using the original input training datasets, we built and trained the convolution-neural network (CNN) which is composed of convolution and fully connected layers, and was integrated into the traditional support vector machine (SVM). The network was trained with twenty-two (22) datasets. The original predictor datasets are composed of hydrological fluxes (HF), climate indices (CI), and sea surface temperature (SST) datasets whose influence on GRACE-TWSA was analyzed in the body of our work. Our results show that the polynomial-kernel based on the convolution-based support vector machine (CSVM) outperformed the other regression models in the reconstruction of all variants of TWSA. The high accuracy achieved using the CSVM demonstrates its promising potential to fill the gaps in the missing GRACE observations through its refined reconstruction capabilities. New hydrological insight for the region: In this study, (i) SST variants or components (e.g., those of east tropical Atlantic) were leading predictors of TWSA, (ii) increasing polynomial order does not increase accuracy in all polynomial kernel operations, especially in cases of severe non-linearity between predictors and predictands, (iii) for the RBF kernels, intermediate gamma values ranging from 0.05 to 0.5 are ideal for climatic analysis, as very small or very large gamma values will either behave like a linear model or over fit, respectively, and (iv) overall, the polynomial kernels reconstructed and predicted TWSA better than the other kernels. The other conventional machine learning procedures used to compare the fit of the CSVM show significant insight for future TWSA reconstruction processes. Their robustness can be explored by varying the independent variables fed into the machine learning framework and tuning their hyper-parameters to result in better fitting index, which promises to be much more useful than traditional learning methods.