Aggregrated Transformation and Attention Neural Network for Remote Sensing Scene Recognition

Intelligent and accurate recognition of remote sensing scenes is a huge challenge and has important application potential in land resources, marine and coastal zones, inland waters, ecological environment, agriculture, forestry and other fields. In this paper, we propose aggregrated transformation and attention neural network (ATANet) and present a new loss function for remote sensing scenes recognition. ATANet is based on aggregrated transformations and attention mechanism, which increases the ability of feature extraction and reduces the training parameters. The presented loss function in this paper can obviously increase the distance between classes and force the reduction of the distance within classes. We tested our model on Tiangong Remote Sensing dataset and the experimental results show that the accuracy of our proposed method is superior to the baseline methods.