Abstract: Scene classification enables the autonomous platform to understand the environmental information. For the scene classification task, randomization-based neural networks could quickly recognize the scene information and spend little time to train the weights. However, the shallow network structure of randomization-based neural networks limits the non-linear representation ability. Moreover, the fully connected method ineffectively extracts local feature information and introduces a large number of parameters. Ensemble architecture could effectively improve the accuracy. However, it introduces high computational complexity and lots of parameters, which will greatly slow down during inference. To tackle above problems, a multi-level convolutional randomization-based network ensemble architecture (E-MCRNet) was proposed for real-time scene classification tasks. Firstly, replacing the fully connected layer with multi-level convolutional layer, the randomization-based network was constructed a multi-level convolutional randomization-based network (MCRNet). Secondly, multiple MCRNets were combined to form an ensemble architecture named E-MCRNet. The E-MCRNet consists of one main-hidden layer and multiple sub-hidden layers. The main-hidden layer was concatenated with each sub-hidden layer to form component networks respectively. Testing results show that E-MCRNet can improve the accuracy and decrease model complexity. Moreover, it can be deployed on embedded equipment to deal with relevant tasks.