Object-Based Land-Cover Supervised Classification for Very-High-Resolution UAV Images Using Stacked Denoising Autoencoders

Over the last decade, object-based image classification (OBIC) has become a mainstream method in remote sensing land-use/land-cover applications. Many supervised classification methods have been proposed in the OBIC framework. However, most did not use deep learning methods. In this paper, a new deep-learning-based OBIC framework is introduced. First, we segment the original image into objects by graph-based minimal-spanning-tree segmentation algorithm. Second, we extract the spectral, spatial, and texture features for each object. Then we put all features into stacked autoencoders (SAE) or stacked denoising autoencoders (SDAE) network, and trained the parameters of the network using training samples. Finally, all objects were classified by the network. Based on our SAE/SDAE OBIC framework, we achieved 97% overall accuracy when classifying an UAV image into five categories. In addition, our experiment shows that our framework increases overall accuracy by approximately 6% when compared to the linear support vector machine (linear SVM) and radial basis function kernel support vector machine (RBF SVM) algorithms when sufficient training samples are lacking.