Accurate and Fast Classification of Natural Disasters using CNN-LSTM and Inference Acceleration

Abstract

Catastrophic occurrences induced by disasters often lead to fatalities, extensive damage, and societal disruptions. In pursuit of realizing disaster-resilient smart cities, video surveillance systems incorporating artificial intelligence (AI) can automatically process and classify the disaster content in real-time. This advancement is fueled by the recent progress in computer vision and AI algorithms, specifically deep learning neural networks, which can be leveraged for disaster categorization tasks. However, minimizing the complexity of AI models while preserving accurate disaster classification remains a formidable challenge. In this paper, we propose a convolutional neural network-long short-term memory (CNN-LSTM) model capable of discerning four types of natural disasters and a non-disaster event. Contrary to prior research that treats input video as a sequence of independent frames, we demonstrate the significance of spatio-temporal characteristics in reaping high prediction accuracy. Furthermore, conventional methods rely on resource-intensive hardware to boost AI model performance, which may not suit real-time monitoring. To facilitate real-time disaster monitoring applications, the trained model is further optimized by utilizing a neural network acceleration platform known as OpenVINO. Our findings reveal that the optimized version of the proposed CNN-LSTM model sustains 100% accuracy while boosting throughput by 25% in terms of frames per second (FPS).

Keywords