Deep Learning Method for Flood Detection

Abstract

Floods, one of the problems, faced by low-land communities can occur practically everywhere. The impacts of the flood can be devastating, especially when it fills up an area in a short amount of time. However, some of the existing systems can be expensive and complex in their implementation in a real-world setting. The aim of this study is to propose a deep learning system for early flood detection using AlexNet. For this purpose, related images would undergo an image processing technique before they are used for model training. The pre-trained model is used to extract important information from these images. Based on the results, all the training accuracy with different parameters reached 100% except for hyperparameter 20 batch size and 10 epochs. while for the validation accuracy, the average value for 5 epochs is 95.74% and for 10 epochs is 95.03% The confusion matrix for binary classification has also been implemented for this system which the parameters are accuracy, specificity, sensitivity, and precision. The best result was the system with 30 batch size0 epochs and 157 sample images which the accuracy value is 0.9574, precision and sensitivity have the same value which is 0.9677 and the specificity is 0.9375. It is hypothesized that the epochs and mini-batch size affected the time taken to finish the classification process and the accuracy of the system.