Urban Flood Disaster Mitigation through Image Classification Using Transfer Learning Method with MobileNet Fine-tuning

Floods severely threaten numerous countries worldwide, with Indonesia being particularly vulnerable. Floods can profoundly affect various aspects of life, underscoring the critical importance of flood mitigation measures. Leveraging advancements in artificial intelligence and computer vision, particularly in deep learning, offers novel avenues for flood monitoring, particularly in urban areas. This research encompasses several vital stages, encompassing data collection, data pre-processing, the application of fine-tuned transfer learning techniques, and evaluation using a confusion matrix. The dataset comprises 2000 images categorized into flood and non-flood classes. Following the training, validation, and testing phases, impressive outcomes emerged. Training accuracy soared to 99% with a minimal loss of 0.01, while validation accuracy reached 96% with a loss of 0.19. Test results were equally stellar, with the confusion matrix displaying accuracy of 95%, recall of 97%, and an F-1 score of 96%. With a computing time of just 48.9 seconds for testing 200 images and a 99% confidence level in classifying each image further underscored the model's exceptional performance. These findings attest to the model's optimal operation and ability to effectively detect and classify flood images. The information produced by this model is beneficial.