最新刊期
卷 26 ， 期 4 ， 2022
-
摘要：With the rapid development of artificial intelligence technology such as machine learning and deep learning in remote sensing, data-driven models have become a new research paradigm for automatic information retrieval from remote sensing imagery, calling for higher requirements for the quantity, quality, and diversity of sample datasets. Before the era of deep learning, because classical machine learning methods (e.g., support vector machine and random forest) do not require huge numbers of samples for model training, the previously published sample datasets usually have a relatively small size (i.e., less than 100). In recent years, with the rapid development of technologies such as big data, parallel computing, and deep learning, many scholars and research institutions have issued a series of sample datasets, laying a solid foundation for a wide range of research and applications such as scene understanding, semantic segmentation, and object detection from remote sensing images. However, comprehensive review of the recently published sample datasets for remote sensing image analysis under the context of big data and deep learning remains lacking. Therefore, the objective of this study is to summarize and analyze these datasets to provide a valuable data reference for relevant researchers.On the basis of literature retrieval and analysis, this paper summarized a total of 124 widely used, open access, and influential remote sensing image sample datasets that were published between 2001 and 2020.We reviewed and summarized the development of recently published sample datasets for remote sensing imagery based on metadata analysis from the following aspects, such as data sources, application fields, keywords, and data size. Afterward, we analyzed these sample datasets from the perspective of spatial, spectral, and temporal resolutions. We listed the commonly used deep learning models (e.g., convolutional neural networks, recurrent neural networks, and generative adversarial networks) in the remote sensing field to show how these sample datasets could be used. We also divided the remote sensing image sample datasets into eight categories based on the following application fields: scene recognition, land cover/land use classification, thematic information extraction, change detection, ground-object detection, semantic segmentation, quantitative remote sensing, and other applications. The typical datasets and related research progress were carefully reviewed for each application field. In addition, because deep learning models are data-hungry, how to train a model with good generalization capability under limited labeled data has become a significant issue, especially for remote sensing applications given that obtaining sufficient labeled samples is time-consuming. To address this issue, we discussed several methods that could increase the model’s generalization capability, including sample transfer between spatio-temporal domains, few-shot learning, and zero-shot learning, active learning, and semi-supervised learning for sample discovery, as well as sample generation through generative adversarial networks.By means of multi-dimensional analysis, we give a comprehensive overview of remote sensing image sample datasets. To the best of our knowledge, this paper is the first review of remote sensing image sample datasets for deep learning, potentially providing data reference for researchers in related fields.关键词：remote sensing imagery;sample datasets;machine learning;deep learning- 5500
- |
- 2392
- |
- 8
发布时间：2022-11-01 -
摘要：A geomorphological dataset is considered to be one of the most important data sources to realize automatic classification of geomorphology and deepening understanding of geomorphological morphology. At present, the datasets of high-precision geomorphologic origin are scarce, hindering the development of automatic geomorphological interpretation using remote sensing data techniques. In the Tianshan–Xingmeng orogenic system, which is dominated by the gully arc-basin system in northeast China, three scene datasets namely, tectonic geomorphology, volcanic lava geomorphology, and flowing geomorphology are made. These geomorphology types were formed by strong tectonic movement, volcanism from the Neozoic, and flowing water action from the Neozoic. The data set covers an area of approximately 5000 km2, including visible light remote sensing image of Sentinel-2, SRTM1 DEM, and seven geomorphological variables based on DEM extraction (hillshade, slope, DEM local average value, standard deviation, two components of aspect, and relative deviation from mean value). Each sample patch is 64×64 pixels with a spatial resolution of 10 m. A multi-modal deep learning model is proposed for classification, and the results show that the average test accuracy is 82.63%. The quality of the dataset is high. The dataset (available from https://pan.baidu.com/s/1Kzj04cU-TiofPk6pTEKENg, password: cug0) could provide fundamental data support for the automatic classification research of geomorphological causes by remote sensing and promote the development of intelligent interpretation in the geoscience community by remote sensing techniques.关键词：geomorphology datasets;geomorphological classification;deep learning;remote sensing scene classification;vegetation cover- 1598
- |
- 344
- |
- 2
发布时间：2022-11-01 -
摘要：Synthetic Aperture Radar (SAR) is one of the important data sources for built-up area information acquisition and dynamic monitoring. In this paper, SARBuD1.0, a SAR image patch dataset of built-up area with GF-3 Fine strip-map mode, is introduced.The dataset is derived from 27 scenes of 10 m resolution GF-3 SAR images covering different regions in China. Approximately 60000 samples of built-up area are obtained from the SAR images. The dataset consists of built-up area SAR image patches and corresponding label images that are interpreted by experts with high-resolution optical images. The dataset contains built-up areas of different distribution types and different regions. The terrain scenes of the samples include plain area, mountain area, and plateau. The SARBuD1.0 dataset can support researchers to analyze the image features of built-up areas in different regions, assist in SAR image understanding, and provide training and test data for deep learning methods for built-up area segmentation in SAR images.In this paper, using built-up areas in mountain areas as an example, the traditional texture features and deep learning features of the built-up area are analyzed. Experiments show that a convolution neural network can provide deeper features of built-up areas compared with traditional texture features. Features in different scales can be obtained by the shallow convolutional layers of the network, and semantic information can be obtained by the deep layers. Therefore, the deep convolutional neural network classifier can detect and extract the built-up area better. Based on the dataset, three deep learning methods are applied to extract the built-up area in different terrain areas. The experimental results show that the deep learning models can achieve good results for built-up area extraction based on the dataset.The dataset can effectively support the deep learning method for big data processing. Based on the SARBuD1.0 dataset, scholars can carry out research on feature analysis and semantic segmentation of built-up areas with SAR images.关键词：remote sensing;synthetic aperture radar (SAR);building;dataset;deep learning;GF-3;semantic segmentation- 2269
- |
- 521
- |
- 0
发布时间：2022-11-01
Data Articles
-
摘要：Data fusion is an effective way to solve the limitation of hyperspectral satellites on temporal and spatial resolution. Discussing the fusion effects of different methods on GF-5 hyperspectral data is highly important for information mining and promotion application of GF-5 hyperspectral data.In this study, based on the principle that the algorithm is easy to use and suitable for generalization, six fusion methods, namely, GS (Gram-Schmidt), GSA (GS Adaptive), CNMF (Coupled Non-negative Matrix Factorization), CRISP-B, CRISP-W (Color Resolution Improvement Software Package with Butterworth or Wavelet transform), GLP (Generalized Laplacian Pyramid) are separately used to perform fusion experiments on GF-5 hyperspectral data and multispectral data from BJ-2, GF-2, and GF-1/1C/1D domestic satellites. Visual interpretation, five indicators (correlation coefficient, universal image quality index, spectral angle mapper, erreur relative globale adimensionnelle de synthèse, and peak signal-to-noise ratio), classification application, and time costs are used to comprehensively evaluate the fusion results.Results show that the fusion image series are the same and the smaller the spatial resolution difference, the better the fusion result. CRISP-B, CRISP-W, and GLP can achieve a good balance in improving spatial resolution and spectral fidelity. In terms of spatial reconstruction, GLP is slightly better and more stable, while CRISP-B and CRISP-W are more stable and effective in maintaining spectral information. The data source will have a certain effect on the fusion method. In the tasks that require high spectral fidelity, such as spectral feature information extraction and analysis, GLP is more suitable for the fusion of homologous data (such as GF-5 and GF-1/1C/1D/2). When the multi-source images (GF-5 and BJ-2) are merged, CRISP-W is preferred. CNMF has a certain degree of color distortion and takes a long time to run. GSA and GS have the worst fusion effect. The spectral retention and the spatial resolution improvement ability of GSA are more stable than those of GS. Based on a small sample, the classification effect of the CRISP-B fusion result is stable and highly accurate. The GSA fusion results are rich in spatial details. Although the spectral distortion is relatively serious, it also increases the spectral distinction of the ground objects, which is still suitable for accurately drawing buildings and roads.This study provides method decision support for the fusion of GF-5 hyperspectral data and other domestic satellite multispectral data, which is helpful for the application and promotion of GF-5 hyperspectral data.关键词：hyperspectral remote sensing;GF-5;domestic satellite;data fusion;fusion method evaluation- 1414
- |
- 701
- |
- 1
发布时间：2022-11-01 -
摘要：The unified sample cloud detection method proposed by Sun et al. based on the AVIRIS hyperspectral sample database simulates the cloud and clear sky surface pixels of the sensor to be detected. The method also inputs the simulated multi-spectral sample data into BP neural network for pixel-by-pixel classification to generate cloud detection models. This method can realize high-precision cloud detection of Landsat 8 OLI and other wide spectrum sensors. This method, which simulates the sample pixel libraries, is suitable for cloud detection of various sensors. Given that the Landsat 8 OLI sensor has more bands, the spectrum covers a wide range, easily facilitating high-precision cloud identification.In this paper, an improved cloud detection algorithm is proposed. Owing to the narrow spectral range of GF-6 WFV data and the lack of cloud-sensitive bands such as 1.38 μm, 1.65 μm, and thermal infrared bands, cloud identification accuracy in high reflectivity areas is low. The cloud and clear sky image metadata simulated based on the unified sample pixel database have a weak ability to identify clouds and bright surfaces, and realizing stable and high-precision cloud identification of this type of satellite is impossible. To further improve the application precision of cloud detection on GF-6 WFV data with a narrow spectral range, GF-6 WFV data typical highlighted surface pixels were added into the simulated sample base to realize cloud detection of GF-6 WFV data with high precision.The main content of the improved unified sample cloud detection algorithm is as follows. (1) GF-6 WFV multi-spectral data simulation. This paper simulates the apparent reflectance of the corresponding band of GF-6 WFV data by the weighted synthesis of the band of AVIRIS hyperspectral data. (2) BP deep learning cloud detection model. In the simulated GF-6 WFV data sample base, typical highlighted surface samples, such as bare land, buildings, and snow, in the GF-6 WFV data are added. The apparent reflectance of each band of the improved data pixel in the sample base is taken as the input vector. The inductive ability of the neural network is used to learn cloud detection rules, generate a cloud detection model, and conduct cloud detection experiments.The accuracy of cloud detection results was verified by visual interpretation. The artificial labeled cloud area was taken as the reference truth value, and the cloud detection results of the algorithm were compared with the reference truth value on a per-pixel basis. By constructing an error matrix, the precision of cloud inspection results of the proposed algorithm is calculated and verified. With the improved cloud detection algorithm, the average correct rate of cloud pixels reaches 0.884 and 0.874 for cloud pixels over the highlighted land surface and 0.926 for cloud pixels over different land surface types.The results show that the algorithm can accurately identify the thick clouds, broken clouds, and thin clouds over vegetation, water, buildings, and bare land with high precision by using the remote sensing data of visible and near-infrared channels. The algorithm of adding high-reflectivity ground objects can use limited bands to achieve high-precision separation of clouds and ground.关键词：cloud detection;GF-6 WFV data;Hyperspectral pixel library;Bright surface;Deep learning algorithm- 873
- |
- 359
- |
- 1
发布时间：2022-11-01
Chinese Satellites
-
摘要：This study aims to comprehensively evaluate the four sets of pure satellite precipitation data of IMERG and GSMaP under Global Precipitation Measurement (GPM). The evaluation content is divided into three aspects: extreme precipitation index, extreme precipitation detection ability, and accuracy evaluation of extreme precipitation events with different durations.Based on the grid data set of automatic stations and CMORPH, the study area is divided into three regions according to the threshold of extreme precipitation events. Five evaluation indexes, namely, CC, BIAS, RMSE, POD, and FAR, are used to quantitatively study the performance of satellite precipitation in extreme precipitation. The precipitation products of IMERG and GSMaP are affected by factors such as terrain, precipitation intensity, and inversion algorithm, showing similar error characteristics and obvious accuracy differences. The correction of similar error characteristics will be the focus and direction of extreme precipitation inversion in the future. The discussion of accuracy difference can provide a reference for the improvement of the precipitation satellite inversion algorithm. (1) In the RX1 extreme precipitation index, IMERG and GSMaP data are obviously overestimated in the complex terrain area affected by complex terrain and underestimated in other areas. In the R95pTOT index, four sets of satellite data perform well and have a high correlation with ground-based datasets. (2) In terms of detection capability of extreme precipitation events, the performance of four pure satellite products in Northeast China is better than that of other regions. GSMaP performs better than does IMERG data with a lower false alarm rate, but the retrieval accuracy for extreme precipitation is low. (3) In the accuracy evaluation of extreme precipitation events with different durations, IMERG and GSMaP satellite precipitation products have better performance and higher precision in long-term extreme precipitation events. For extreme daily precipitation events, the error of satellite precipitation products under high rain intensity is very obvious, which is much higher than that of complex terrain on the accuracy of satellite precipitation retrieval, resulting in the performance of satellite precipitation in complex terrain area Ⅲ is better than other regions. Overall, the IMERG products have the better ability to monitor extreme precipitation in the study area than GSMaP products, and IMERG_Late data performs best. The retrieval error of extreme precipitation from IMERG and GSMaP satellite precipitation products has obvious regional characteristics in China, and the error characteristic of underestimation of high rainfall intensity is significant. The four types of satellite precipitation products can show the extreme precipitation region characteristics in study area but underestimate the precipitation in most parts of the study area. The error correction of satellite precipitation products for rainfall intensity remains one of the key and difficult points in future extreme precipitation retrieval.关键词：remote sensing;IMERG;GSMaP;satellite precipitation;extreme precipitation;GPM;error characteristics- 1229
- |
- 291
- |
- 3
发布时间：2022-11-01 -
摘要：Salt marshes are highly productive ecosystems in the mid-high latitude coastal zone, and the ecological service functions provided by different types of salt marsh vegetation are significantly different. The combination of human activities and natural factors such as reclamation, invasion of Spartina alterniflora, and sea level rise has led to rapid changes in the structure and spatial distribution of salt marshes in China’s coastal zone. Existing optical methods are subject to tidal and cloud interference in the coastal zone. Obtaining large-scale and high-efficiency salt marsh vegetation information using hyperspectral or LiDAR data is difficult.This study took the Yangtze River estuary as the research area and proposed a coastal salt marsh vegetation classification method based on vegetation phenology and multi-temporal radar backscatter feature. Sentinel-1 radar data were used to analyze the annual time-series characteristics of radar backscattering in salt marshes, intertidal forest swamps, mudflats, and water bodies. Based on the phenological characteristics of salt marsh vegetation, the separability between the monthly backscattering characteristics of a typical salt marsh was calculated based on the separation threshold method. According to the optimal time-series radar classification characteristics, the random forest method was used to obtain the species, structure, and spatial distribution of salt marsh vegetation.Results showed the following. (1) The average value of VH polarization backscattering can distinguish water bodies, light beaches, intertidal forest swamps, and salt marshes well. (2) The mean backscattering values of VV polarization in April, VH polarization in November, and VV polarization in March were the optimal characteristics of Scirpus × mariqueter, Spartina alterniflora, and Phragmites australis. (3) Obtained by the optimal characteristics and random forest classification algorithm, the general classification accuracy of salt marsh vegetation was 85% with a Kappa coefficient of 0.80.Compared with optical remote sensing, radar images can effectively obtain the inter-annual and inter-monthly time series backscattering characteristics of salt marsh vegetation, and accurately obtain the spatial dynamics of coastal salt marsh vegetation. This study has shown the application potential of radar images in coastal zone research and provides important technical means and data support for coastal biodiversity conservation, wetland ecosystem function enhancement, and ecological environment management.关键词：remote sensing;Sentinel-1;salt marsh;time series analysis;phenological characteristics;Yangtze River estuary- 1128
- |
- 348
- |
- 4
发布时间：2022-11-01 -
摘要：Establishing a more accurate remote sensing ecological index is necessary to evaluate urban ecological quality and provide timely warnings. Taking the Beijing city as the study area, this paper used five indices (vegetation index, humidity, Land Surface Temperature (LST), Normalized Difference Build-up and bare Soil Index (NDBSI) and air quality) through the Principal Component Analysis (PCA) method to construct a Modified Remote Sensing Ecological Index (MRSEI). The Eco-environment Index (EI) was derived from the Pressure-State-Response model (PSR) combined with the entropy weight method to compare with MRSEI and RSEI. Moreover, the nuclear principal component analysis (KPCA) was applied to establish the Nonlinear Remote Sensing Ecological Index (NRSEI), which was integrated vegetation index, humidity, LST, and NDBSI. Finally, MRSEI and NRSEI were separately compared with the remote sensing ecological index (RSEI). The results showed that MRSEI could reflect the spatial distribution of air quality, and the correlation coefficients between MRSEI and EI were 0.829 in 2014 and 0.857 in 2017 (P<0.01), which were improved by 0.035 and 0.055 over that of RSEI, respectively. Compared with EI, the average absolute error, root mean square error, and average relative error of MRSEI in the main districts were all lower than that of RSEI. These results indicated that the MRSEI in evaluating urban ecological quality was better than RSEI and the air quality indicator was feasible to monitor the ecological environment of Beijing. The contribution rate of the first principal component from NRSEI was increased by 11.94%—21.45% than that of RSEI in the experiment areas. Compared with RSEI, the correlation coefficients between each indicator and NRSEI increased by 0.128—0.198. NRSEI could demonstrate the transition of different ecological levels. RSEI sometimes underestimated the areas with poor ecological environments, and it sometimes overestimated the areas with excellent ecological environments. NRSEI was more consistent with the ecological conditions reflected by remotely sensed images. MRSEI is more suitable than RSEI for monitoring the ecological quality of Beijing. NRSEI, taking into account the weak linear or nonlinear correlations of various indicators, is better than RSEI in assessing the ecological environment quality.关键词：remote sensing;modified remote sensing ecological index;nonlinear remote sensing ecological index;air quality index;kernel principal component analysis;Pressure-State-Response model- 1740
- |
- 634
- |
- 8
发布时间：2022-11-01
Ecology and Environment
-
摘要：Remote sensing image classification technology has been widely used in land resource monitoring, forest resource investigation, and other related fields. For the traditional classification methods, terrain effect is an unavoidable factor that restricts the improvement of classification accuracy in the classification of land cover types, and the influence can be weakened by an appropriate correction model, and topographic correction has been proven to play a positive role in improving classification accuracy. Compared with traditional classifiers, deep neural network classifiers based on deep learning theory have the advantages of deep feature learning and expression, which have emerged in the field of image classification and have been gradually applied to land cover classification with good results.Our study has some shortcomings. First, some errors remain in GlobeLand30 and national forest classification products compared with the actual surface cover types. Therefore, more detailed and accurate classification results need to be found as sample labels in the future. Second, the surface of sunny-shady slopes with different types of coverage is not considered in this study. Finally, U-Net should not be the sole focus. In subsequent research, we will attempt to select a variety of deep neural network classifiers to explore the classification accuracy changes before and after topographic correction of neural network models at a more precise scale to obtain further conclusions.This paper further explores the influence of topographic correction on the classification accuracy of land cover classification by deep neural network classifiers.Using Landsat 8 OLI image data and GDEM_V2 terrain data with 30 m resolution as data sources, and based on the classification results of GlobeLand30 and national forest types, this paper implemented the classification extraction of land cover types by the U-Net semantic segmentation network and made a comparative analysis of classification accuracy before and after topographic correction with different sample acquisition methods and different levels of classification systems.Classification results show the following. (1) With two training sample acquisition methods, namely, grid clipping and aspect auxiliary clipping, the classification accuracy after topographic correction is unchanged or slightly reduced compared with that before correction, and the reduction range is 0.9%—1.39%. (2) For the classification of more precise forest types, the classification accuracy after topographic correction decreased by 1.66% compared with that before correction.In this paper, we conduct a preliminary study and find that in the classification of land cover types by the U-Net model with different sample acquisition methods, namely, regular grid clipping and aspect auxiliary clipping, and under different classification systems, the classification accuracy of the deep neural network classifier-U-Net is not improved by the topographic correction process.Our study has some shortcomings. First, some errors remain in GlobeLand30 and national forest classification products compared with the actual surface cover types. Therefore, more detailed and accurate classification results need to be found as sample labels in the future. Second, the surface of sunny-shady slopes with different types of coverage is not considered in this study. Finally, U-Net should not be the sole focus. In subsequent research, we will attempt to select a variety of deep neural network classifiers to explore the classification accuracy changes before and after topographic correction of neural network models at a more precise scale to obtain further conclusions.关键词：remote sensing classification;deep neural network;U-Net model;topographic correction;land cover classification;classification accuracy- 903
- |
- 341
- |
- 1
发布时间：2022-11-01 -
摘要：The dimensionality reduction processing of multispectral data is of considerable importance to deep learning-based single-tree crown detection research. However, how to use the appropriate dimensionality reduction method to improve the accuracy of single-tree detection is rarely discussed. In this work, an unmanned aerial vehicle equipped with a multispectral camera was used for aerial photography to collect multispectral images of ginkgo tree species in the research area. The original multispectral images were used to generate five different data sets through feature band selection, feature extraction, and band combination method for training three classical deep learning networks: FPN-Faster-R-CNN, YOLOv3, and Faster R-CNN. Based on the characteristics of the band selection method, the red, green, and near-infrared bands combined with different types of target detection in the network have the best results. The FPN-Faster-R-CNN network detection accuracy is up to 88.4% for ginkgo trees. The blue, red, and near-infrared band combination obtained by the OIF index has the highest amount of information but the lowest average network accuracy at 79.3%. Experimental results show the following. In the different dimensionality reduction methods, if the color and background of the target object in the image after dimensionality reduction are obviously different, and the contour is clear, the deep learning network can obtain better results in tree crown detection. However, the information content of the image itself has a limited effect on the ability of the deep learning network to detect tree crowns. In this study, the dimensionality reduction method of multispectral images is analyzed, providing an important experimental reference for the deep learning-based single-tree crown detection.关键词：remote sensing;individual tree crown detection;deep learning;unmanned aerial vehicle;multispectral images;dimensionality reduction- 1082
- |
- 632
- |
- 2
发布时间：2022-11-01 -
摘要：Hyperspectral remote sensing can obtain both spatial images and continuous spectral data of land cover, thus realizing the classification and identification of ground objects. However, the high-dimensional characteristics of hyperspectral images pose great challenges to classification. Therefore, this paper discusses a hyperspectral image classification method based on Hash learning. Hash learning can switch high-dimensional information to low-dimensional binary encoding, and achieve classification by calculating encoding internal product and using hamming distance.To effectively express nonlinear data, one supervised hashing method is proposed. However, the disadvantage of Hash learning are that they run slowly and lack consideration for spatial neighborhood information. Therefore, this paper introduces RBF kernel to improve efficiency. In addition, it uses four-dimensional convolution to fully express spatial information, namely Supervised Hashing with RBF Kernel and Convolution (CKSH).Experiments were carried out on the international general test data. Experimental results show that the proposed method is superior to traditional classification methods and other hash learning. Under different percentage conditions of training samples, it has achieved high classification accuracy and reached 96.12% (Indian Pine, 10%) and 98.00% (University of Pavia, 5%), which verified the effectiveness of proposed method.In view of the two problems that loss function of KSH using the L2 norm causes low efficiency, and it does not consider spatial information. This paper uses four-dimensional convolution to introduce spatial information, and applies the RBF kernel instead of the L2 norm. Experiments on general test data sets have confirmed the advantages of CKSH in classification accuracy and runtime. The reason for high accuracy and efficiency has two hands. On the one hand, CKSH uses four-dimensional convolution to mine underlying structural information. Therefore, the obtained binary encoding conducive to improving the classification performance. On the other hand, using the RBF kernel as loss function can significantly reduce the arithmetic series.关键词：remote sensing;hyperspectral image;hash learning;RBF;four-dimensional convolution;feature extraction;spectral-spatial classification- 1054
- |
- 345
- |
- 1
发布时间：2022-11-01 -
摘要：Land Surface Temperature (LST) is a key parameter in the physical process of surface radiant energy balance and the water cycle. Obtaining LST data accurately and promptly, and mastering its temporal and spatial changes, are of considerable importance to climate change research. Thermal infrared (TIR) measurements are limited in practical applications due to cloud cover and other effects. Passive microwave (PMW) remote sensing measurements can penetrate clouds and are less affected by atmospheric interference, which has the advantage of obtaining all-weather surface radiation information. The microwave remote sensing data Advanced Microwave Scanning Radiometer for EOS (AMSR-E) can obtain all-weather LST, which can be used as a supplement to the missing LST information in thermal infrared (TIR) products under cloudy conditions. However, the AMSR-E data has the problem of lack of information due to the satellite orbit scanning gap of its own sensor, causing the obtained AMSR-E LST data to be greatly restricted in practical applications. Therefore, proposing an effective method for solving the problem is necessary.Based on the superiority of deep learning in solving non-linear problems and the high dynamic variability of LST, this paper proposes a Multi-Temporal Feature-Connected Convolutional Neural Network (MTFC-CNN) that uses specific input combinations of multi-temporal information and spatial fusion units. The network structure is based on the characteristics of the temporal and spatial distribution of missing track gaps in AMSR-E LST data, and the reconstruction of missing LST values is carried out from the timing information.In the simulation experiment, the 2010 annual data was divided into fight data subsets in four seasons and into day and night. The average root mean square error of the reconstructed LST value is approximately 1.0 K and the coefficient of determination R2 is above 0.88. Compared with the other two methods, namely, spline interpolation (Spline) and time multiple linear regression (Regress), the reconstruction effect of the MTFC-CNN method performs better regardless of seasons or day and night, proving that MTFC-CNN is better than the other two methods at mining the characteristics of temporal and spatial changes in LST. In real experiments, through comparison with MODIS LST products, the LST value reconstructed at the missing area is consistent with it at other areas in temporal and spatial distribution. The reconstruction results show that the LST in the mainland China region shows a gradually increasing trend from January to July and a gradually decreasing trend from August to December. This finding is consistent with the temperature changes in the four seasons. The change in LST during the day is more significant than that at night. In summer, the temperature in Northwest China is significantly higher than that in other regions. In winter, the temperature in Northeast China is generally lower than that in other regions. At night, the difference between summer and winter is more obvious. The difference in LST changes at night in autumn is relatively close.The experimental results show that the MTFC-CNN method proposed in this paper mines the spatio-temporal variation information of LSTs more effectively than two traditional methods and achieves better results in reconstructing the orbital gap missing in AMSR-E LST data. The method provides the possibility for the reconstruction of missing information from TIR LST data under the cloud.关键词：remote sensing;land surface temperature;AMSR-E LST;data reconstruction;deep learning;MTFC-CNN- 989
- |
- 543
- |
- 5
发布时间：2022-11-01 -
摘要：Automatic identification of docks can provide an important basis for the construction and development of ports, acquisition of coastal geographic information, and evaluation of maritime military strength. However, docks are characterized by small sizes, large quantities, and scattered distribution. Docks are also negatively affected by serious information interference of the surrounding environment, including ships and buildings. Traditional algorithms cannot easily meet the needs of accurate monitoring of rapidly developing docks. Accurate identification of dock targets has become an urgent problem to be solved. Based on the open remote sensing data sets and Google Earth high-resolution remote sensing images, the data sets of three types of docks are constructed, and the following improvements are made to the Faster R-CNN algorithm according to the size and spatial distribution characteristics of docks. (1) The K-means algorithm is used to preset the anchors, making the anchor sizes more suitable for the actual dock sizes. (2) Soft-NMS is used instead of NMS to reduce the rates of mistaken deletion and missed detection of dock borders in densely distributed areas. The experimental results show that the accuracy of the improved FKSN algorithm reached 92.6%, which is 6% higher than that of the Faster R-CNN algorithm. The final result of dock target recognition is compared with the ones of the traditional classification methods such as ISODATA, SSD, Faster R-CNN, and Faster R-CNN+K-Means. Among these approaches, the method suggested in this paper performs best in the evaluation indices of false alarm rate and omission rate, which are 3.2% and 7.6%, respectively. Thus, the proposed method has a better effect on the identification of various dock targets. The automatic dock identification algorithm based on the improved Faster R-CNN can provide technical support for reasonable construction, planning, and management of docks and provide effective approaches for efficient utilization and military strength analysis of docks.关键词：Faster R-CNN;automatic dock identification;K-Means algorithm;Soft-NMS algorithm;high resolution remote sensing- 922
- |
- 491
- |
- 1
发布时间：2022-11-01
Intelligent Remote Sensing Processing and Analysis
-
摘要：Support Vector Regression (SVR) method as a new idea in LAI inversion has certain application value and prospect. However, the value of penalty coefficient C, width parameter g of kernel function and insensitive loss function parameter- 890
- |
- 338
- |
- 2
发布时间：2022-11-01 -
Small water body extraction method based on Sentinel-2 satellite multi-spectral remote sensing image
摘要：The water body extraction algorithm based on satellite remote sensing is mainly aimed at large- and medium-sized lakes or large rivers. When applied to small water bodies, misjudging often occurs. The multi-spectral remote sensing data of the Sentinel 2 satellite has a spatial resolution of 10, 20, and 60 meters; a dual-star time resolution of 5 days; and a high temporal and spatial resolution. Therefore, this paper uses the sentinel-2 green light band (560 nm), red-edge band (705 nm), near-infrared band (842 nm, 865 nm), and short wave-infrared band (2190 nm) for remote sensing reflectance. A new water index algorithm (red edge-based water index, RWI for short) is proposed for the extraction of fine water. The normalized remote sensing reflectance of vegetation, shadow, building, mixed pixel, bare soil, and water body is compared and analyzed. The mechanism explains why RWI has a better effect of extracting fine water bodies compared with other water body indexes.To quantitatively evaluate the advantages of the RWI water body extraction algorithm proposed in this paper, this paper compares several current water body extraction algorithms, including the improved normalized difference water index MNDWI (modified normalized difference water index), multi-band water body index MBWI (multi-band water index), and AWEI (automatic water extraction index). The water body extraction results obtained by several algorithms and the area statistics of the artificial water interpretation of the water body results are calculated using the error formula to obtain RWI. The errors of MNDWI, MBWI, AWEIsh, and AWEInsh are 3.6%, 4.2%, 12.2%, 8.8%, and 19.8%, respectively. Results show that the RWI algorithm has the highest accuracy. At the same time, through analysis, the advantages and disadvantages of several water extraction results are quantitatively evaluated. From the results of image extraction, this paper proposes that the water boundary extraction method extracted by the RWI water body extraction algorithm is better, and it can eliminate mountains, building shadows, and clouds. The effect of shadows can eliminate the effects of mixed pixels.At the same time, this paper carried out multi-temporal water extraction in Xiong’an New District, Shendong Mining Area, and Yongcheng Mining Area. In the time range from January 2016 to December 2018, a total of 43 Sentinel-2 images without clouds were screened, and the small broken water bodies in the three areas of Xiong’an New District, Shendong Mining Area, and Yongcheng Mining Area were selected by the algorithm proposed in this paper. The distribution carried out multi-temporal analysis. Results from statistical area analysis indicate that Xiong’an New District has the largest water body area in spring, followed by summer and autumn. No significant change in water area was observed in the Shendong and Yongcheng mining areas. After careful observation, the results of each phase are very good, the boundary of the small water body is highly differentiated, and no data are misleading or missing, and the algorithm has good applicability and stability.关键词：water extraction;water body index;small water body;Sentinel-2;RWI;MNDWI;MBWI;AWEI- 1713
- |
- 722
- |
- 9
发布时间：2022-11-01
Remote Sensing Technology and Method
- Postal code：100190
- Tel：010-58887052 Email：nrsb@aircas.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备20021838号-9
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0