Remote sensing image scene classification has been widely applied and has attracted increasing attention. Recently, convolutional neural networks (CNNs) have achieved remarkable results in scene classification. However, scene images have complex semantic relationships between multi-scale ground objects, and the traditional stacked network structure lacks the ability to extract multi-scale and key features, resulting in limited feature representation capabilities. By simulating the way that humans understand and perceive images, attention mechanisms can be beneficial for quickly and accurately acquiring key features. In our study, we propose a channel-attention-based DenseNet (CAD) network for scene classification. Firstly, the lightweight DenseNet121 is selected as the backbone for the spatial relationship between multi-scale ground objects. In the spatial domain, densely connected CNN layers can extract spatial features at multiple scales and correlate with each other. Secondly, in the channel domain, a channel attention mechanism is introduced to strengthen the weights of the important feature channels adaptively and to suppress the secondary feature channels. Thirdly, the cross-entropy loss function based on label smoothing is used to reduce the impact of inter-class similarity upon feature representations. The proposed CAD network is evaluated on three public datasets. The experimental results demonstrate that the CAD network can achieve performance comparable to those of other state-of-the-art methods. The visualization through the Grad-CAM ++ algorithm also reflects the effectiveness of channel attention and the powerful feature representation capabilities of the CAD network.

For identification of forested landslides, most studies focus on knowledge-based and pixel-based analysis (PBA) of LiDar data, while few studies have examined (semi-) automated methods and object-based image analysis (OBIA). Moreover, most of them are focused on soil-covered areas with gentle hillslopes. In bedrock-covered mountains with steep and rugged terrain, it is so difficult to identify landslides that there is currently no research on whether combining semi-automated methods and OBIA with only LiDar derivatives could be more effective. In this study, a semi-automatic object-based landslide identification approach was developed and implemented in a forested area, the Three Gorges of China. Comparisons of OBIA and PBA, two different machine learning algorithms and their respective sensitivity to feature selection (FS), were first investigated. Based on the classification result, the landslide inventory was finally obtained according to (1) inclusion of holes encircled by the landslide body; (2) removal of isolated segments, and (3) delineation of closed envelope curves for landslide objects by manual digitizing operation. The proposed method achieved the following: (1) the filter features of surface roughness were first applied for calculating object features, and proved useful; (2) FS improved classification accuracy and reduced features; (3) the random forest algorithm achieved higher accuracy and was less sensitive to FS than a support vector machine; (4) compared to PBA, OBIA was more sensitive to FS, remarkably reduced computing time, and depicted more contiguous terrain segments; (5) based on the classification result with an overall accuracy of 89.11% ± 0.03%, the obtained inventory map was consistent with the referenced landslide inventory map, with a position mismatch value of 9%. The outlined approach would be helpful for forested landslide identification in steep and rugged terrain.

Road information from high-resolution remote sensing images is widely used in various fields, and deep-learning-based methods have effectively shown high road-extraction performance. However, for the detection of roads sealed with tarmac, or covered by trees in high-resolution remote sensing images, some challenges still limit the accuracy of extraction: 1) large intra-class differences between roads, and unclear inter-class differences between urban objects, especially roads and buildings; 2) roads occluded by trees, shadows, and buildings are difficult to extract; and 3) lack of high-precision remote-sensing datasets for roads. To increase the accuracy of road extraction from high-resolution remote-sensing images, we propose a split depth-wise (DW) separable graph convolutional network (SGCN). Firstly, we split DW-separable convolution to obtain channel and spatial features, to enhance the expression ability of road features. Thereafter, we present a graph convolutional network to capture global contextual road information in channel and spatial features. The Sobel gradient operator is used to construct an adjacency matrix of the feature graph. A total of thirteen deep-learning networks were used on the Massachusetts Roads Dataset and nine on our self-constructed mountain road dataset, for comparison with our proposed SGCN. Our model achieved a mean Intersection over Union (IOU) of 81.65% with an F1-score of 78.99% for the Massachusetts Roads dataset, and a mean IOU of 62.45% with an F1-score of 45.06% for our proposed dataset. The visualization results showed that SGCN performs better in extracting covered and tiny roads and is able to effectively extract roads from high-resolution remote-sensing images.

Land cover mapping (LCM) in complex surface-mined and agricultural landscapes could contribute greatly to regulating mine exploitation and protecting mine geo-environments. However, there are some special and spectrally similar land covers in these landscapes which increase the difficulty in LCM when employing high spatial resolution images. There is currently no research on these mixed complex landscapes. The present study focused on LCM in such a mixed complex landscape located in Wuhan City, China. A procedure combining ZiYuan-3 (ZY-3) stereo satellite imagery, the feature selection (FS) method, and machine learning algorithms (MLAs) (random forest, RF; support vector machine, SVM; artificial neural network, ANN) was proposed and first examined for both LCM of surface-mined and agricultural landscapes (MSMAL) and classification of surface-mined land (CSML), respectively. The mean and standard deviation filters of spectral bands and topographic features derived from ZY-3 stereo images were newly introduced. Comparisons of three MLAs, including their sensitivities to FS and whether FS resulted in significant influences, were conducted for the first time in the present study. The following conclusions are drawn. Textures were of little use, and the novel features contributed to improve classification accuracy. Regarding the influence of FS: FS substantially reduced feature set (by 68% for MSMAL and 87% for CSML), and often improved classification accuracies (with an average value of 4.48% for MSMAL using three MLAs, and 11.39% for CSML using RF and SVM); FS showed statistically significant improvements except for ANN-based MSMAL; SVM was most sensitive to FS, followed by ANN and RF. Regarding comparisons of MLAs: for MSMAL based on feature subset, RF achieved the greatest overall accuracy of 77.57%, followed by SVM and ANN; for CSML, SVM had the highest accuracies (87.34%), followed by RF and ANN; based on the feature subsets, significant differences were observed for MSMAL and CSML using any pair of MLAs. In general, the proposed approach can contribute to LCM in complex surface-mined and agricultural landscapes.

Deep convolutional neural networks have become an indispensable method in remote sensing image scene classification because of their powerful feature extraction capabilities. However, the ability of the models to extract multiscale features and global features on surface objects of complex scenes is currently insufficient. We propose a framework based on global context spatial attention (GCSA) and densely connected convolutional networks to extract multiscale global scene features, called GCSANet. The mixup operation is used to enhance the spatial mixed data of remote sensing images, and the discrete sample space is rendered continuous to improve the smoothness in the neighborhood of the data space. The characteristics of multiscale surface objects are extracted, and their internal dense connection is strengthened by the densely connected backbone network. GCSA is introduced into the densely connected backbone network to encode the context information of the remote sensing scene image into the local features. Experiments were performed on four remote sensing scene datasets to evaluate the performance of GCSANet. The GCSANet achieved the highest classification precision on AID and NWPU datasets and the second-best performance on the UC Merced dataset, indicating the GCSANet can effectively extract the global features of remote sensing images. In addition, the GCSANet presents the highest classification accuracy on the constructed mountain image scene dataset. These results reveal that the GCSANet can effectively extract multiscale global scene features on complex remote sensing scenes. The source codes of this method can be foundin https://github.com/ShubingOuyangcug/GCSANet .