SqueezeNet with Attention for Remote Sensing Scene Classification

Scene classification is an important problem in remote sensing (RS) since it is a prerequisite to other more intelligent analysis operations. Given an RS scene, not all of its parts are important for classification. Thus, using an attention mechanism that directs the classification system to focus on the parts that are important and ignore the irrelevant background should enhance the system's accuracy. In this work we propose a deep CNN architecture based on the pre-trained SqueezeNet CNN. This CNN is composed of nine fire modules (fire 1 to fire 9) each consisting of Squeeze followed by expansion convolution layers. First, we improve the SqueezeNet CNN by introducing several modifications to the architecture. Then we introduce a separate branch to the network that implements an attention mechanism. Each neuron in this activation map of the fire 9 module covers a different receptive field in the original scene. An attention mechanism is applied to these neurons to learn the appropriate weighing scheme for merging the feature vectors corresponding to each neuron. Feature vectors that are assigned a higher weight indicate that the network has given more attention to the receptive field in the scene corresponding to that feature vector. Preliminary results are presenting on five popular scene datasets, namely UC Merced, KSA, AID, Whurs19, and Optimal31 datasets.