Scene classification of high-resolution remote sensing images (HRRSI) is one of the most important means of land-cover classification. Deep learning techniques, especially the convolutional neural network (CNN) have been widely applied to the scene classification of HRRSI due to the advancement of graphic processing units (GPU). However, they tend to extract features from the whole images rather than discriminative regions. The visual attention mechanism can force the CNN to focus on discriminative regions, but it may suffer from the influence of intra-class diversity and repeated texture. Motivated by these problems, we propose an attention-based deep feature fusion (ADFF) framework that constitutes three parts, namely attention maps generated by Gradient-weighted Class Activation Mapping (Grad-CAM), a multiplicative fusion of deep features and the center-based cross-entropy loss function. First of all, we propose to make attention maps generated by Grad-CAM as an explicit input in order to force the network to concentrate on discriminative regions. Then, deep features derived from original images and attention maps are proposed to be fused by multiplicative fusion in order to consider both improved abilities to distinguish scenes of repeated texture and the salient regions. Finally, the center-based cross-entropy loss function that utilizes both the cross-entropy loss and center loss function is proposed to backpropagate fused features so as to reduce the effect of intra-class diversity on feature representations. The proposed ADFF architecture is tested on three benchmark datasets to show its performance in scene classification. The experiments confirm that the proposed method outperforms most competitive scene classification methods with an average overall accuracy of 94% under different training ratios.

With the increase of the spatial resolution of aerial images, the shadow problem is more prominent. The shadows affect the applications such as object recognition, image dense matching, and object classification. Existing shadow detection methods can acquire results with high accuracy, but usually need much manual intervention and the traditional stretch models generally lead to color distortion and undercompensation. Therefore, we propose an object-oriented automatic shadow detection method without manual intervention and a shadow compensation method by regional matching. In the proposed method, pixel-based soft shadow detection, which uses Gaussian mixture model to simulate the gray distribution and refines soft shadow map with guiled filtering, is combined with image segmentation result to obtain accurate shadow regions with complete shape and no hole. Then shadow regions are compensated, with less loss of details and brightness imbalance, referring to their optimal homogeneous nonshadow region obtained by regional matching based on Bag-of-Words. The total variation model is used to decrease the noise amplified by the pixel-based stretch and boundary effect in the result. Experiments are performed on three publicly available high-resolution aerial images to demonstrate the superiority of our proposed methods. It shows that the proposed method can accurately detect shadows from urban high-resolution aerial images with an overall rate of over 88%. The compensation results display excellent visual effects compared with the state-of-the-art methods, consistent with the true color of ground objects.

Scene classification is usually based on supervised learning, but collecting instances is expensive and time-consuming. TrAdaBoost has achieved great success in transferring source instances to target images, but it has problems, such as the excessive focus on instances harder to classify, the rapid convergence speed of the source instances, and the weight mismatch caused by the big gap between the number of source and target instances, leading to decreased classification accuracy. In this paper, in order to address these problems, classical particle swarm optimization (PSO) is modified to select the optimal feature subspace for classifying the “harder” and “easier” instances by reducing unimportant dimensions. A modified correction factor is proposed by considering the classification accuracy of the instances from both domains, to decrease the convergence speed. Iterative selective TrAdaBoost is also proposed to reduce the weight mismatch by removing the indiscriminate source instances. The experimental results obtained with three benchmark data sets confirm that the proposed method outperforms most of the previous methods of scene classification with limited target samples. It is also proved that modified PSO for optimal feature subspace selection, the modified correction factor, and iterative selective TrAdaBoost are all effective in improving the classification accuracy, giving improvements of 3.6%, 4.3%, and 2.7%, and these three contributions together increase the classification accuracy by about 8% in total.

Artificial target detection is a key technology for Earth observation applications in remote sensing images, including environmental monitoring, urban planning, intelligence reconnaissance, and land mapping. Recently, transformer-based object detectors achieved great success in computer vision through the attention mechanism. However, these detectors lack prior information, which may limit their multi-scale, multi-shape, and densely distributed target detection capabilities. To address these problems, we propose a novel transformer-based object detection method with geometric and object co-occurrence prior knowledge for remote sensing images, which makes improvements based on the deformable detection transformer (DETR). First, we introduce dynamic anchor object queries with multi-patterns to detect objects with multi-scale and dense distribution. Second, we propose a novel distance with geometrical invariance to measure the position deviation of multi-shape objects. Lastly, we design a graph convolutional reference module with co-occurrence prior knowledge to improve the inferential ability of the detector. Experimental results confirm that the proposed method outperforms most of the state-of-the-art methods with mean average precisions (mAP) of 70.2% in the DIOR, 91.0% in the HRRSD datasets and 91.4% in the NWPU VHR-10 dataset, respectively. The mAP values are more than 5% higher compared with those obtained by deformable DETR in above three public datasets. The code will be made available to public at https://github.com/momo91nj/GOCPK .

In this paper, the problem of multi-scale geospatial object detection in High Resolution Remote Sensing Images (HRRSI) is tackled. The different flight heights, shooting angles and sizes of geographic objects in the HRRSI lead to large scale variance in geographic objects. The inappropriate anchor size to propose the objects and the indiscriminative ability of features for describing the objects are the main causes of missing detection and false detection in multi-scale geographic object detection. To address these challenges, we propose a class-specific anchor based and context-guided multi-class object detection method with a convolutional neural network (CNN), which can be divided into two parts: a class-specific anchor based region proposal network (RPN) and a discriminative feature with a context information classification network. A class-specific anchor block providing better initial values for RPN is proposed to generate the anchor of the most suitable scale for each category in order to increase the recall ratio. Meanwhile, we proposed to incorporate the context information into the original convolutional feature to improve the discriminative ability of the features and increase classification accuracy. Considering the quality of samples for classification, the soft filter is proposed to select effective boxes to improve the diversity of the samples for the classifier and avoid missing or false detection to some extent. We also introduced the focal loss in order to improve the classifier in classifying the hard samples. The proposed method is tested on a benchmark dataset of ten classes to prove the superiority. The proposed method outperforms some state-of-the-art methods with a mean average precision (mAP) of 90.4% and better detects the multi-scale objects, especially when objects show a minor shape change.