Development of Lr_multi- Cross-validation Model for Prediction of an Imbalanced Dataset in Flood Susceptible Area

Abstract

Climate change has a profound impact on human well-being and health. It threatens the fundamental aspects of a good quality of life if not effectively managed. Changes in the frequency and intensity of heavy rainfall events can lead to shifts in the scale and occurrence of river floods, altering how floods happen. This results from warmer temperatures causing increased water evaporation from both land and oceans. However, situations like floods, droughts, and famines raise global concerns. These complex alterations entail calamities and necessitate comprehensive analysis for effective prediction and counteraction. Machine learning algorithms and cross-validation techniques have been employed in the past for flood forecasting by identifying patterns from various indicators. While traditional K-FOLD is an effective and commonly used cross-validation technique, novel adaptations like Multi Predicting Cross-Validation of K-fold, Stratified K-fold, and Repeated K-Fold could address overfitting in imbalanced datasets. Therefore, this study hybridized the application of Logistic Regression to check the distribution of data points for each fold from the three k-fold techniques and builds a Random Forest model for flood prediction. The area under the precision-recall curve (AUPRC) was the critical metric due to data imbalance. The new hybridized model demonstrates a marked of improvement when the result was compared with the traditional KNN-based model. The Random Forest had 99% AUPRC, against the previous result of 84.96% from the traditional KNN model. This underscores the power of meticulous model validation in enhancing model selection.