Remote Sensing Scene Classification by Multi-granularity Fused CNN

Owing to the ability to capture deeper spatial feature representations, the Convolutional Neural Networks (CNN) has been successfully devoted in remote sensing scene classification currently. Nevertheless, the performance of deep models inevitably meets the bottleneck in the scene classification, due to the huge difference in a category and high similarity between different categories. To capture the latent ontological feature, we propose a novel Multi-Granularity Fused CNN (MGF-CNN) in this work. We firstly crop input images to learn multi-grained features progressively. And then, we learn the corresponding Gaussian covariance matrices of different granularities. Thirdly we fuse the granularities learned from the same images. Extensive experimentation and evaluations have demonstrated that the proposed network can achieve promising consequences in public remote sensing datasets.KeywordsRemote sensing scene classificationConvolutional Neural NetworkMulti-granularity fuseGaussian covariance matrix