Evaluating Machine Learning Models for Flood Forecasting in Baladweyn City, Somalia

Baladweyne City is a city in the central part of Somalia that has experienced disruptions due to flooding over the years. The flooding has always affected the transport and general operations in the Shabelle River floodplain. The main aim of the research is to evaluate the performance of the Support Vector Regression approach in flood forecasting in Baladweyne, with specific emphasis on the dependence of the kernel, the possibility of overfitting and underfitting, and the overall improvement in the accuracy of the results over the benchmark models. The Support Vector Regression approach was implemented using the LIBSVM toolbox with the scaling and normalization of the data and the overall configuration of the input for the models. The results of the Support Vector Regression approach were compared with the results of the transfer function, trend, and naïve persistence models. The results of the research showed that the accuracy of the results of the benchmark models decreased significantly with the increase in the forecast period. For example, the best RMSE of the benchmark models for the first lead time was 170, while the RMSE for the sixth lead time was 1010, showing a deterioration of 494% in the accuracy of the results. On the other hand, the Support Vector Regression approach showed significant improvements in the accuracy of the results over the results of the benchmark models. For example, the Support Vector Regression approach showed improvements of 2.9% in the accuracy of the results for the first lead time and improvements of up to 17.3% for the longer lead times. The results of the Support Vector Regression approach showed that the linear kernels were more robust in the conditions where the rainfall for the future period is not available, while the RBF kernels showed better accuracy when the rainfall for the future period is available. The results of the Support Vector Regression approach showed that the approach is capable of capturing the time of the major flood peaks with minimal relative deviation, while the results of the rainfall-response experiment showed a scaling from low rainfall of 0-4 mm to high rainfall of 50-100 mm.