Oladapo Kayode Abiodun
Department of Physical and Computer Sciences, McPherson University, Seriki Sotayo, Ogun State, Nigeria.
Email: [email protected]
Alowolodu Olufunsho Dayo
Department of Cybersecurity Science, School of Computing, Federal University of Technology, Akure, Ondo State, Nigeria.
Email: [email protected]
Abstract
Pluvial flooding is a type of flood that has become a threat to human life and the global economy due to the variability in climate change, rapid urbanization and fast-growing spatial development. Studies have shown that Machine Learning (ML) applications can be used to reduce this risk. However, less attention has been paid to the use of a fuzzy rule-based classification to appraise the performance of ML applications, based on pluvial flood Conditioning Variables (CVs) for training a classifier. Lack of categorization for a universal guideline in selecting the CVs has also been a challenge. Therefore, the study developed a fuzzified risk assessment model for the detection and prediction of pluvial flood using ML techniques.
The application was developed by categorizing the pluvial flood CVs using the degree of interdependency index. A total of 144,401 records and eight CVs out of 53 were gathered from the United States Geological Survey and Copernicus Climate Data Store for five Local Government Areas (LGAs). The LGAs are the urban regions in Ibadan Metropolis of Oyo State, Southwest Nigeria, which were purposively selected out of 11 LGAs. The CVs were further categorized into three groups: hydrological, topographical and anthropological. Fuzzy logic was applied to the eight CVs using the triangular method, after which five ML algorithms were used to train the CVs: K-Nearest Neighbours (KNN), Random Forest (RF), Classification and Regression Trees (CART), Naïve Bayes (NB) and Artificial Neural Network (ANN) algorithms. The fuzzy logic re-classified the selected CVs into five classes: no flood, low, moderate, high and very high pluvial flood susceptibility. The performance of the application was evaluated using the 10-fold cross-validation and hold-out techniques, based on accuracy, sensitivity, specificity, precision and Area Under Receiver Operating Characteristics (AUROC) metrics.
The performance evaluation results for each algorithm, using hold-out techniques in respect of accuracy, sensitivity, specificity, precision, and AUROC for KNN were 95.3%, 95.3%, 92.7%, 93.8% and 92.2% respectively; for RF, 72.8%, 73.0%, 73.2%, 73.0% and 83.6% respectively; for NB, 71.0%, 77.0%, 73.7%, 84.7% and 72.7% respectively; for CART, 98.4%, 98.4%, 98.3%, 98.4% and 98.6% respectively; and for ANN, 83.6%, 84.0%, 96.9%, 74.0% and 87.9% respectively. In addition, results obtained for using 10-fold cross-validation method for KNN were 96.4%, 96.4%, 94.1%, 96.6% and 93.7% respectively; for RF, 95.2%, 95.2%, 93.7%, 94.3% and 94.6% respectively; for NB, 77.3%, 77.3%, 74.7%, 84.3% and 89.5% respectively; for CART, 95.5%, 99.5%, 99.4%, 99.5% and 97.6% respectively; and for ANN, 89.5%, 89.5%, 89.7%, 89.1% and 89.9% respectively.
The study concluded that the fuzzified ML application can be used in detecting and predicting pluvial floods. Consequently, CART which had the best results, when compared to the rest of the classifier models, is recommended for use by experts. This is because the application showed the capacity to assist experts to alert the public on expected risks. Also, it will further assist local authorities and government agencies for appropriate pluvial flood management plans.
Keywords: Risk assessment, Pluvial flood, Fuzzy logic, Machine learning algorithms, Degree of interdependency index