Search

Rapid and accurate acquisition of glacier regional changes is of great significance to the study of glaciers. Among all satellite images, Synthetic Aperture Radar (SAR) data has a great advantage in monitoring the glaciers in harsh weather conditions. Conventionally, glacier boundaries are manually delineated on images. However, this is a time-consuming process, especially in the batch process of large-area data. In this paper, we propose a Multiscale Joint Deep Neural Network (MJ-DNN) for large-scale glaciers contour extraction using single-polarimetric SAR intensity images. Based on U-Net, the proposed method has been improved in three aspects. First, Atrous Separable Convolution is used instead of convolution with the down-sampling part. Second, we propose a multiscale joint convolution layer to obtain information at multiple scales. Third, we deepen the network with the residual connection structure for higher-level features. At the final layer, we optimize the network result with the conditional random field method. To validate our approach, we test it on three glaciers and we compare the segmentation results of four different methods in parallel. The results show that the intersection over the union of the proposed method is the most efficient.

Accurate land-use classification is essential for the management and supervision of urban development, land resources and environment sustainability. Feature extractor and classifier are the important modules of land-use classification. Deep convolutional neural network has been proved to be able to learn more robust and discriminative features from images. In this paper, we increase the diversity and discriminative of features by fusing features extracted by three deep convolutional neural networks with different architectures, which are obtained by fine-tuning the pre-trained models with land-use image dataset. In order to make the classification faster and have excellent generalization performance, we select constrained extreme learning machine instead of fully connected layer or support vector machine. Experimental results show that the proposed method can achieve a better performance with the overall classification accuracy of 98.35%, compared with other state-of-the-art methods.