A Multi-Stream Attention-Aware Convolutional Neural Network: Monitoring of Sand and Dust Storms from Ordinary Urban Surveillance Cameras

Abstract

Sand and dust storm (SDS) weather has caused several severe hazards in many regions worldwide, e.g., environmental pollution, traffic disruptions, and human casualties. Widespread surveillance cameras show great potential for high spatiotemporal resolution SDS observation. This study explores the possibility of employing the surveillance camera as an alternative SDS monitor. Based on SDS image feature analysis, a Multi-Stream Attention-aware Convolutional Neural Network (MA-CNN), which learns SDS image features at different scales through a multi-stream structure and employs an attention mechanism to enhance the detection performance, is constructed for an accurate SDS observation task. Moreover, a dataset with 13,216 images was built to train and test the MA-CNN. Eighteen algorithms, including nine well-known deep learning models and their variants built on an attention mechanism, were used for comparison. The experimental results showed that the MA-CNN achieved an accuracy performance of 0.857 on the training dataset, while this value changed to 0.945, 0.919, and 0.953 in three different real-world scenarios, which is the optimal performance among the compared algorithms. Therefore, surveillance camera-based monitors can effectively observe the occurrence of SDS disasters and provide valuable supplements to existing SDS observation networks.