Scene Classification Based on a Deep Random-Scale Stretched Convolutional Neural Network.

With the large number of high-resolution images now being acquired, high spatial resolution (HSR) remote sensing imagery scene classification has drawn great attention but is still a challenging task due to the complex arrangements of the ground objects in HSR imagery, which leads to the semantic gap between low-level features and high-level semantic concepts. As a feature representation method for automatically learning essential features from image data, convolutional neural networks (CNNs) have been introduced for HSR remote sensing image scene classification due to their excellent performance in natural image classification. However, some scene classes of remote sensing images are object-centered, i.e., the scene class of an image is decided by the objects it contains. Although previous methods based on CNNs have achieved comparatively high classification accuracies compared with the traditional methods with handcrafted features, they do not consider the scale variation of the objects in the scenes. This makes it difficult to directly utilize CNNs on those remote sensing images belonging to object-centered classes to extract features that are robust to scale variation, leading to wrongly classified scene images. To solve this problem, scene classification based on a deep random-scale stretched convolutional neural network (SRSCNN) for HSR remote sensing imagery is proposed in this paper. In the proposed method, patches with a random scale are cropped from the image and stretched to the specified scale as the input to train the CNN. This forces the CNN to extract features that are robust to the scale variation. Furthermore, to further improve the performance of the CNN, a robust scene classification strategy is adopted, i.e., multi-perspective fusion. The experimental results obtained using three datasets-the UC Merced dataset, the Google dataset of SIRI-WHU, and the Wuhan IKONOS dataset-confirm that the proposed method performs better than the traditional scene classification methods.