Integration of mobile deep networks and machine learning methods for flood risk classification: 2D grayscale transformation of data, feature intersection

Abstract

Floods pose significant risks to human life and the economy. This study presents a hybrid artificial intelligence approach designed to assess flood risk and improve risk management. The proposed model minimizes dataset complexity and optimizes performance by using 2D deep learning networks, allowing for more detailed feature extraction. The approach converts 1D data into 2D grayscale images as a preprocessing step, enabling the use of deep learning models that operate on 2D data. To ensure efficiency, mobile deep networks, including MobileNetV2 and NasNetMobile, were selected due to their lower computational requirements, making them suitable for hardware deployment. Feature sets were extracted from the final layers of the models and combined into a set of 2000 features. Machine learning classifiers (kNN, LDA, SVM) were used to assess the level of flood risk. The best performing method, LDA, was selected for further analysis. Feature selection techniques, Chi2 and Kruskal-Wallis, were used to refine the feature set and identify the most relevant features. The intersection of the features selected by both methods resulted in 767 features, optimizing both performance and computational efficiency. The LDA method achieved an accuracy of 99.86%, and in cross-validation tests, an accuracy of 99.68%. These results demonstrate that the integration of deep learning models with feature selection techniques significantly improves the accuracy of flood risk prediction while reducing time and hardware costs.