遥感图像目标检测
遥感图像目标检测是开展环境监测、城市规划、精准农业和土地测绘等对地观测应用的关键技术，具有重要的科研价值与广泛的应用前景。在环境监测和资源探测方面，可以有效识别森林与植被覆盖率，辅助城市有效规划等；在农业调查方面，可以快速获取农作物种植分布和种植面积等农情信息；在海洋监测方面，可以识别海上船舰、港口等重要海况，对军事侦察和民用监测具有重要意义。小编整理了近年发表的遥感图像目标检测的最新文章供您赏析！
- The Paper
- Abstract：Oceanic internal waves are widely presented in all levels of the water column in deep oceans and marginal areas. These waves play an important role in seawater energy exchange. The study of oceanic internal waves has important academic values and practical significance in marine resources, marine engineering, and the marine military. The oceanic internal waves are distinct bright and dark stripes in Synthetic Aperture Radar (SAR) images. Those stripes can serve as clues to efficiently identify the oceanic internal waves from SAR images. The growing popularity of computer vision has led to the wide adoption of deep learning for the detection of oceanic features in remote sensing data. In this study, we intend to apply the faster R-CNN, a state-of-the-art deep learning method, to the automatic detection of oceanic internal waves.The faster R-CNN is the most widely used version of the R-CNN family. This algorithm depends on the region proposal algorithms to hypothesize object locations. Based on the bright and dark stripes in the SAR images, a Faster R-CNN-based method is developed for oceanic internal wave detection. First, the oceanic internal waves are manually labeled in the SAR images to serve as the training set. The training set for the detection method contains 5480 SAR images, which are in multi-band, multi-polarization mode, and multi-spatial scale. These images are collected in the South China Sea region from 2001 to 2020. Then, the Faster R-CNN network is trained based on the obtained training set. Meanwhile, the parameters (such as training epochs) are optimized. The transfer learning technic is applied in the training process to transfer information from the previously learned tasks for detecting oceanic internal waves to accelerate the training process and avoid overfitting. The well-trained Faster R-CNN network can be applied by a sliding window on the SAR images to detect the oceanic internal waves. When the boundaries are obscured, the waves may be detected multiple times. In this case, the detection results will be grouped and merged. Finally, the detection results of the oceanic internal waves are acquired and recorded.The evaluations are conducted on multi-source SAR data, showing that the accuracy rate (AP) and recall rate (AR) of the developed method are up to 95.7% and 92.3%, respectively. This method can achieve high accuracy while keeping the false alarm rate relatively low. The experiments on the SAR images with complex ocean conditions also show favorable results.This work developed the approach for the detection of oceanic internal waves. This approach successfully transferred techniques developed in the computer vision field to solve the issues in remote sensing problems. The establishment of this method provides a technical basis for the detection of oceanic internal waves from a large amount of SAR data and further promotes the research of internal wave parameter inversion and dynamic processes. In addition, the proposed method was initially designed for lunar research, but it could be applied to the detection of other oceanic features.Keywords：oceanic internal waves;automatic detection;SAR images;deep learning;Faster R-CNN1927|2564|7
<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-05-10 - Abstract：Ship detection by satellite remote sensing is of great significance for the safety of maritime navigation and the maintenance of maritime rights and interests. The traditional ship detection based on high spatial resolution Synthetic Aperture Radar (SAR) and optical satellite images cannot easily realize high-frequency monitoring application due to the long revisit period. The medium resolution Coastal Zone Imager (CZI) carried by China’s “Ocean-1” series satellites (HY-1) has a relatively low spatial resolution (50 m). However, HY-1C and HY-1D form a double satellite network observation in the morning and afternoon, which has the advantage of short revisit period and is of great value for marine vessel monitoring. We attempt to realize the ship automatic detection and orientation technology of medium-resolution CZI images, which will be of great value to the monitoring of ships at sea. In this study, a convolutional neural network is used for feature learning and target extraction, and an automatic ship detection method of HY-1/CZI image is established. Verification results show that this method has the advantages of not requiring threshold adjustment and strong adaptability, and the detection accuracy reaches 77.71%, which can be applied to the automatic monitoring of marine moving ships in the HY-1/CZI image. The algorithm in this work can directly detect the position and motion information of marine moving ships from the medium-resolution HY-1/CZI image without manual screening, realize the automatic extraction of wake, and overcome the problem of insufficient resolution of the medium-resolution optical image. Based on the detection results, this work further quantitatively describes the wake and obtains the information of the ship's position and movement direction.Keywords：Coastal Zone Imager;Vessel inspection;convolutional neural network;satellite remote sensing934|2217|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-05-10
- Abstract：The existence of a cloud reduces the application value of remote sensing images. Accurate and automatic cloud and cloud shadow detection and labeling for multispectral satellite images is conducive to the subsequent application of remote sensing images. China currently has a large number of high-resolution multispectral satellite images. However, standard data products rarely contain pixel-by-pixel cloud and cloud shadow tag data for quality analysis. Traditional cloud detection algorithms usually require parameters, such as satellite imaging geometry, imaging time, and calibration coefficients. However, many Chinese satellites’ images have lost parameter auxiliary files during multiple product iterations. Moreover, many military application satellite images are missing or do not provide parameter files. Multispectral satellite image cloud and cloud shadow detection with missing parameters requires special research.The present study investigates the cloud and cloud shadow detection method of domestic four-band multispectral satellite imagery with missing related parameters. This algorithm process is based on the classic spectral threshold cloud and cloud shadow detection algorithm. It also uses image processing and morphological algorithms to improve detection accuracy. A morphology-based method for estimating the azimuth and distance of the cloud shadow relative to the cloud area is proposed for the data with missing parameters.The experimental data in this study is from the GF-1 satellite Wide-Field View (WFV) sensor, and the 86 test images are from Dunhuang, Gansu, China. The experimental area contains a large area of a bright surface and snow-capped mountains that are easily misdetected in cloud and cloud shadow detection. The result of the cloud and cloud shadow detection experiment in this study for the case of missing parameters achieves accuracy similar to that achieved by normal algorithms. This study also analyzes the misdetection of the algorithm and clarifies the challenges of subsequent research.In this study, we propose a set of refined cloud and cloud shadow detection algorithms in case of missing parameters for domestic four-band multispectral satellite imagery. The algorithms are based on the classical spectral threshold cloud and cloud shadow detection algorithms. They also use image processing and morphological algorithms to improve accuracy further. Moreover, a morphology-based method for estimating the orientation and distance of cloud shadow relative to the cloud area is proposed for the data with missing parameters. The experimental results of GF-1 WFV data show that the detection results of this algorithm achieve an accuracy similar to that of the widely used MFC algorithm in the case of missing parameters.Keywords：cloud detection;cloud shadow detection;GF-1;multispectral satellite image;missing parameters4276|3326|3<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-04-18
- Abstract：The characteristics of remote sensing images, such as complex scenes, different sizes of targets and unbalanced distributions, increase the difficulty of target detection. However, feature pyramids that are suitable for detecting targets of different scales do not consider the importance of different feature maps when fusing the feature maps, let alone emphasize the features of target areas. For this purpose, this paper proposes a feature attention pyramid-based remote sensing image object detection method (namely, the feature attention pyramid network, FAPNet).First, the feature maps of different depths are fused by channel concatenation, and the features of different sized receptive fields are provided for the feature maps used for detection. The channel attention mechanism is used to recalibrate the fused feature maps in the channel dimension. The feature maps from different depths are adaptively adjusted according to the scale of the object to be detected to strengthen the feature that matches highly between the size of the receptive field and the object to be detected and weaken the feature with a low degree of matching. Second, the weakly supervised attention module uses the superimposed atrous spatial pyramid pooling structure and convolutional segmentation module to model spatial attention weights to adjust the feature distribution of the feature map used for prediction, strengthen the object area feature, and weaken the background area feature, which further improves the performance of object detection methods.The experimental results show that compared with RetinaNet, the proposed method improves the accuracy (AP) for car targets by 3.41% and 2.26% on the UCAS-AOD dataset and RSOD dataset, respectively, achieves better AP results on each target for multiclass targets and is superior to other comparative object detection methods on the mAP indicator for multitargets.A feature attention pyramid-based remote sensing image object detection method is proposed in this paper. Its contribution lies in the designed feature attention pyramid module and weakly supervised attention module. With the new modules, the proposed method can extract target features more accurately in complex scenes with targets of different sizes by channel attention and spatial attention, thus improving the performance of detection. The experimental results show that the proposed method is superior to the RetinaNet and FAN methods and is more suitable for remote sensing image object detection tasks with complex scenes and multiscale targets.Keywords：remote sensing image;object detection;weakly supervised segmentation;attention mechanism;feature pyramid1758|4280|13<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-06-30
- Abstract：The Coast Zone Imager (CZI) onboard the Chinese first marine aqua-color satellite HY-1C started operational operations in June 2019. The data acquired by CZI have the characteristics of medium resolution, large width and high revisit period and taking into account the requirements of ocean water color, terrestrial ecology and polar glaciers. Therefore, the large amount of coastal, land, and ocean data acquired by CZI is of great significance for marine disaster and environmental monitoring research. However, related studies have shown that clouds cover an average of 68% of the earth's surface. CZI data is severely affected by cloud, which will then have a strong impact on its subsequent applications. The effective identification of clouds in remote sensing images is extremely important for the application of CZI images. Most of the existing cloud detection algorithms are based on RGB images or multi-spectral images including thermal infrared band. There are few researches on cloud detection algorithms for RGB-NIR four-band remote sensing images, such as HY-1C CZI. The objective of this paper is thus to propose an unsupervised cloud detection method for HY-1C CZI remote sensing images that makes full use of NIR band information. The method includes four processes: training samples selection, feature extraction, Support Vector Machine (SVM) classification, and post-processing. In the selection of training samples, combining dark channel reflectivity, normalized vegetation index and whiteness index of the image, this paper proposes an automatic training sample extraction algorithm, which uses the whiteness index to obtain detail information, and accurately extract cloud/non-cloud samples through a gradual refinement process. For feature extraction, the spatial spectrum feature information of CZI remote sensing image is selected, including reflectance, spectral index, texture and structure features, to characterize remote sensing image features, and maximize the feature difference between cloud and non-cloud regions. Based on the above automatically extracted sample and its feature description, SVM is used to initially classify the CZI remote sensing data, and then the guided filtering, hole filling and geometric judgment post-processing are performed to obtain the final high-precision cloud detection results. This paper applies the algorithm to four typical scenarios (vegetation, soil, wetland, and ice and snow scenarios), and compares and analyzes it with the currently popular unsupervised cloud detection algorithms. Compared with other cloud detection algorithms, the qualitative analysis results show that the cloud detection results in this paper are in good agreement with real cloud distribution image labeled by human. In addition, the most commonly used error rate metric is also used to quantitatively evaluate the cloud detection results. It shows that the error rates of the proposed algorithm in vegetation, soil, wetland, ice and snow scenes are 0.027, 0.064, 0.026, and 0.049, respectively; and that has the lowest error rate in four scenarios. Through the above comparative analysis, the detection results of the proposed algorithm in different scenarios are more accurate, which demonstrates the effectiveness of the proposed algorithm for cloud detection from HY-1C CZI data.Keywords：HY-1C;Coast Zone Imager (CZI);cloud detection;Whiteness Index;unsupervised2256|1327|1<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-02-17
- Abstract：Object detection of remote sensing image is the description of visual features of the object and the expression of the image prior knowledge, and the information obtained by the interpretation has a wide range of applications in both military and civilian fields. A refined multi-scale feature-oriented object detection of remote sensing image is proposed to address the problems of insufficient feature extraction capability of remote sensing image objects in complex scenes, large variations in object scales, arbitrary and closely arranged directions, and difficulties in the accurate orientation of horizontal frames used in traditional object detection.First, a contextual attention network based on dilated convolution is designed, which can capture local and global semantic information by using convolution kernels with different dilated rates and integrate semantic information into the original features utilizing an attention mechanism to enhance feature extraction. Second, a refined feature pyramid network is proposed to reduce the loss of channel information in the feature pyramid by pixel shuffling and strengthen the network’s ability to understand multi-scale object feature information with large variances. Finally, the study uses gliding vertices to regress the oriented rectangular box to represent the location of directed objects within remote sensing images.In this work, the effectiveness of the algorithm is verified by using Fast R-CNN OBB as a baseline on the object detection public datasets DOTA and HRSC2016. Results show that the algorithm in this work improves the mean average precision (mAP) by 22.65% on the DOTA dataset compared with the baseline. The final detection accuracy mAP reaches 76.78%. The final detection accuracy mAP on the HRSC2016 dataset reached 89.95%. In addition, the algorithm in this work has a better improvement compared with the various advanced algorithms.ConclusionFirst, the contextual attention network with dilated convolution is used to strengthen the object features, which enhances the discriminative ability of the convolutional neural network for objects and backgrounds in remote sensing images. Second, the refined feature pyramid is used to solve the problem of large variation of objects in remote sensing images. Finally, the direction factor of gliding vertices is introduced to represent the oriented objects, which reduces the regression boundedness problem that can be brought by angle regression.Keywords：remote sensing;deep learning;object detection;feature extraction;multi-scale feature pyramid;oriented bounding box2687|4592|4<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2024-03-27
- Abstract：Some existing target detection algorithms are insufficient for feature extraction in remote sensing images. They cannot solve the difficult problem of large target scale differences in remote sensing images, especially in detecting small targets, resulting in low average detection accuracy. In response to these problems, this paper uses the Faster Region Convolutional Neural Network algorithm as the basic algorithm. Furthermore, it combines the target characteristics in the remote sensing images to improve the basic algorithm. Finally, this paper proposes a new remote sensing image target detection algorithm. First, we use the Residual Network with more powerful feature extraction capabilities to replace the Visual Geometry Group network in the original algorithm. It can solve the shortcomings of the original algorithm’s insufficient feature extraction of the remote sensing images. The deep residual network adopts the identity mapping method, which not only ensures that the performance of the network will not degrade as the network deepens but also extracts deeper features. Second, we add a feature pyramid network to the algorithm to fully integrate feature maps of different scales. The feature map obtained in this way has high-level semantic and low-level detail information. Accordingly, it can take category and location information into account. This approach can greatly solve the difficult problem of large target scale differences in remote sensing images and improve the detection accuracy of small targets to a certain extent. In addition, we use the focal loss function to replace the cross entropy loss function in the original algorithm to solve the problem of the weight of the hard and easy samples to the total loss. Finally, given the problem that the used data set contains a small number of images, we use data augmentation to expand the dataset. This paper carries out two sets of comparative experiments to verify the effect of this algorithm. The first set of experiments is the ablation experiments on the NWPU VHR-10 dataset and RSOD-Dataset of the improved modules proposed in this paper. The second set of experiments is the comparison experiments of the algorithm in this paper and the other comparison algorithms on the NWPU VHR-10 dataset. The results of the first set of ablation experiments show that the various improved modules proposed in this paper can help improve the accuracy of target detection in remote sensing images. For the NWPU VHR-10 dataset, after adding the feature pyramid network, focal loss function, and data augmentation strategy, the algorithm in this paper improves mean Average Precision by 2.6%, 4.8%, and 0.8%, respectively. Furthermore, on the RSOD dataset, the algorithm in this paper improves the mean Average Precision by 0.6%, 1.6%, and 0.9%, respectively. Accordingly, the target detection accuracy rates of the algorithm in this paper can reach 93.4% and 93.0% on the NWPU VHR-10 dataset and RSOD-Dataset, respectively. The results of the second set of comparative experiments show that the target detection accuracy of the proposed algorithm is better than the comparison algorithm, further proving that the proposed algorithm has good performance in remote sensing image target detection. Finally, compared with BOW, COPD, RICNN, original Faster R-CNN, ODDP, and Mask R-CNN, the algorithm in this paper improves the mean Average Precision by 68.8%, 12.7%, 20.8%, 10.6%, 6.7%, and 9.5%, respectively. The remote sensing image target detection algorithm proposed in this paper can better solve the difficult problem of large differences in target scale in remote sensing images. It can improve the target detection accuracy of remote sensing images, especially the detection accuracy of small targets.Keywords：remote sensing image;object detection;feature extraction network;Feature Pyramid Network;loss function;data augmentation5104|4010|12<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2022-12-09
- Abstract：As the largest land cover, forests play an important role in human living environment, biological habitat, and global carbon cycle. Forest health is directly related to global ecological security and sustainable development of human society. In recent years, urban construction, disasters, forest management and deforestation, and other factors have caused different degrees of disturbance to forests. It is important to determine the exact time point and spatial range of forest burned area for forest damage assessment, management, carbon accounting, and forest restoration management. Owing to the continuity of spatial distribution of forest burned areas, most of the existing methods of forest burned area extraction use the two-step treatment strategy of first classification and then post-processing to suppress the effect of false alarm pixels. In this paper, a spatiotemporal detection method, Stacked ConvLSTM, is proposed for the detection of forest fire tracks in time series. This method avoids subjective post-processing operations on the basis of maintaining better spatial continuity of the results, and achieves end-to-end extraction of forest burned area information, which improves the extraction accuracy of forest fire-burning land. This paper proposes to use Stacked ConvLSTM to detect forest disturbance in time and space. Combined with the characteristics of ConvLSTM in extracting temporal and spatial characteristics from long-term historical series, it can predict the change trend of vegetation in a period of time in the future, and accurately determine the time point and spatial range of forest disturbance. ConvLSTM is an LSTM variant proposed on the basis of LSTM. The full connection state from input layer to hidden layer and from hidden layer to hidden layer of LSTM is replaced by convolution connection, which can make full use of spatial information. Compared with single-pixel-based methods, ConvLSTM can extract the spatiotemporal structure information of time series images at the same time, which is better for spatiotemporal analysis. In this paper, Stacked ConvLSTM is used to detect the temporal and spatial distribution of forest burned areas, predict the change trend of vegetation in a period of time in the future, and determine the presence of forest burned areas by comparing with the newest time-series images. With MODIS long time series data, based on the historical time series of Yinanhe Forest Farm of Zhanhe Forestry Bureau in Heilongjiang Province and Beidahe Forest Farm of Bilahe Forestry Bureau in Inner Mongolia from 2001—2008 and 2001—2016, the extraction results of burned areas were compared with Stacked LSTM and bfast algorithm. The Stacked ConvLSTM, Stacked LSTM, and bfast algorithms were used to extract forest burned areas from MODIS time series in both regions, and to compare the detection results with the Fire_CCI 5.1 burned areas products released by ESA. Results show that, firstly, from the visual effect, in study area Ⅰ, the error detection of Stacked ConvLSTM is fewer than that of Stacked LSTM and bfast algorithm and maintains high continuity in spatial distribution. In study Area Ⅱ, Stacked ConvLSTM detected a more complete area of fire. Secondly, in study area Ⅰ , Stacked ConvLSTM was 0.120 and 0.405 more accurate than Stacked LSTM and bfast algorithms, respectively. Moreover, the recall rate, accuracy, and Fire_CCI 5.1 F1-score were higher. In study area Ⅰ , the accuracy of Stacked ConvLSTM is 0.924 had a higher recall rate, accuracy, and F1-score than Stacked LSTM, bfast algorithms, and Fire_CCI 5.1. The detection accuracy of ConvLSTM model in space is higher than that of the other two methods, and its continuity of detection results in space is better. The detection effect of ConvLSTM model is equivalent to that of Stacked LSTM in time, but both of them are closer to the real fire time point than bfast algorithm. Results show that Stacked ConvLSTM has advantages in obtaining the change trend of forest long-term historical series for spatiotemporal prediction, and improves the detection accuracy of forest fire to a certain extent.Keywords：Stacked ConvLSTM;time series;spatiotemporal prediction;forest burned area2067|2870|4<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2022-11-18
- Abstract：Water facilities play an important role in water scheduling, ecological protection and restoration of natural wetlands, utilization of resources and functions, and development. The traditional methods of statistics on the location and count of water conservancy facilities rely on compiled data, which has disadvantages such as time-consuming, untimely data update, and unknown specific geographic locations. Remote sensing provides new possibilities for large-scale detection of water conservancy facilities. Aiming at the problem of detection of water conservancy facilities with remote sensing images, this study proposes a large-scale image detection algorithm. Based on the YOLO v3 network and the characteristics of water conservancy facilities, the study was divided into two main aspects: (1) We improved the YOLO algorithm and obtained the E-YOLO algorithm. We proposed a PPA feature fusion method and a four-feature map cross prediction method with proportional prediction box to optimize the problems of small samples. Besides, we improved the loss function by highlighting the loss of confidence. In addition, we used the transfer learning method to read part of the feature extraction parameters of the pre-trained model. (2) With the improved E-YOLO algorithm as the core, a large-area water conservancy facility detection algorithm combined with the water body index constraint was obtained. Aiming at the problem of large image size with a small target scale, we used the water body index to constrain the sliding step to reduce the missed detection rate and false detection rate at the same time. Then we combined the network output with the contour merging method to optimize the detection results. We used the GF-2 data for this study. The experimental results show that: the E-YOLO algorithm can significantly improve the detection effect of water conservancy facilities. Compared with YOLO v3, the average F2 score of E-YOLO is increased by 1.25% and the E-YOLO algorithm has a better stability. The large-area detection method constrained by the water index can improve the detection accuracy while ensuring efficiency. Compared with the large-step and small-step methods, its F2 accuracy is increased by 3.72% and 2.70%, respectively. Our method provides a good solution for the detection of water conservancy facilities.Keywords：water conservancy facilities;remote sensing detection;E-YOLO;large image;NDWI1806|2113|6<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2022-11-18
- Abstract：Object detection is a fundamental aspect of remote sensing image processing. With the development of remote sensing image acquisition and the breakthrough in deep learning, object detection in aerial imagery based on deep learning has attracted considerable interest. Although significant progress has been made, there are still numerous obstacles due to the large-scale and highly complex backgrounds of optical remote sensing images. In addition, approaches based on horizontal proposals for common object detection frequently suffer from the mismatch issue when detecting densely arranged and arbitrarily oriented objects in aerial imagery. Therefore, numerous domestic and international researchers have proposed tilting box object detection algorithms based on deep learning that enhances the object detection effect of remote sensing images. This paper systematically organizes and summarizes them for researchers in related fields to comprehensively understand the theory, process, and existing problems of deep learning-based remote sensing image tilting box object detection.In this paper, we first analyze the limitations of Horizontal Bounding Box (HBB) object detection algorithms applied to remote sensing images, namely, the introduction of background noise, inappropriate post-processing operation, Non-Maximum Suppression (NMS), and the inability to accurately determine the orientation of objects, which can be remedied by the tilting bounding box object detection method.Following this, we list the classical HBB object detection algorithms based on deep learning and briefly describe their underlying principles. Then, the development of the tilting bounding box object detection algorithm and the process of improving the two-stage tilting bounding box object detection algorithm is described from three perspectives: the feature extraction network, anchor boxes, and the proposed region design. Finally, the one-stage detection algorithm’s loss function has been studied infrequently, so the two algorithms are merely introduced.In the fourth section, the detection performance of existing tilt box object detection algorithms is demonstrated on two publicly available and challenging aerial datasets (i.e., DOTA and HRSC2016). The comparison results of the three tables indicate that a particular object feature enhancement module must be designed to account for the uniqueness of the objects in remote sensing images and that the RSE problem in the algorithm for detecting tilting bounding boxes requires additional consideration. Although the one-stage detection algorithm is marginally less accurate than the two-stage algorithm, it has clear advantages in terms of efficiency and therefore has some research value.The paper concludes with a six-point summary of the tilt box target detection algorithm’s existing problems and an outlook on its future development trend.Keywords：remote sensing image;deep learning;convolutional neural network;object detection in tilting bounding box2364|3964|14<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2023-04-21
