Search

Land cover segmentation is an important and challenging task in the field of remote sensing. Even though convolutional neural networks (CNNs) provide great support for semantic segmentation, standard models are still difficult to capture global information and long-range dependencies in remote sensing images. To overcome these limitations, we proposed an attention guided encoder-decoder network with multi-scale context aggregation to achieve more accurate segmentation of land cover. Based on the structure of the encoder-decoder network, we introduce a multi-scale feature fusion module with two attention modules to the top of the encoder. The multi-scale feature fusion module is employed to aggregate multi-scale features and capture global correlations. The attention modules are used to exploit the long-range dependencies and the interdependence between channels from the perspective of space and channel respectively. The experimental results on the GF-2 images show that our proposed method achieves state-of-the-art performance, with an OA of 84.1% and the mIoU of 62.3%. Compared with the baseline network, our method improves the OA by 3.3% and the mIoU by 4.4%. The comparative experiments also demonstrate that the proposed approach can significantly improve the accuracy of land cover segmentation than other compared methods.

Recovering missing values plays a significant role in time series tasks in practical applications. How to replace the missing data and build the dependency relations from the incomplete sample set is still a challenge. The previous research has found that residual network (ResNet) helps to form a deep network and cope with degradation problem by shortcut connection. Gated recurrent unit (GRU) can improve network model and reduce training parameters by update gate which takes the place of forgetting gate and output gate in long short-term memory (LSTM). Inspired by this finding, we observe that shortcut connection and mean of global revealed information can model the relationship among missing items, the previous and overall revealed information. Hence, we design an imputation network with decay factor for shortcut connection and mean of the global revealed information in GRU, called decay residual mean imputation GRU (DRMI-GRU). We introduce a decay residual mean unit (DRMU), which takes full advantage of the previous and global revealed information to model incomplete time series; and the decay factor is applied to balance the previous long-term dependencies and all non-missing values in the sample set. In addition, a mask unit is designed to check the missing data existing or not. An extensive body of empirical comparisons with other existing imputation algorithms over real-world data and public dataset with different ratio of missing data verifies the performance of our model.

Remote sensing scene classification is still a challenging task in remote sensing applications. How to effectively extract features from a dataset with limited scale is crucial for improvement of scene classification. Recently, convolutional neural network (CNN) performs impressively in different fields of computer vision and has been used for remote sensing. However, most works focus on the feature maps of the last convolution layer and pay little attention to the benefits of additional layers. In fact, the feature information hidden in different layers has potential for feature discrimination capacity. The most attention of this work is how to explore the potential of multiple layers from a CNN model. Therefore, this paper proposes multi-layers feature fusion based on CNN and designs a fusion module to solve relevant issues of fusion. In this module, firstly, all the feature maps are transformed to match sizes mutually due to infeasible fusion of feature maps with different scales; then, two fusion methods are introduced to integrate feature maps from different layers instead of the last convolution layer only; finally, the fusion of features are delivered to the next layer or classifier as the routine CNN does. The experimental results show that the suggested methods achieve promising performance on public datasets.