Abstract:Soil Moisture (SM) plays an important role in the global water, energy, and carbon cycle, and its spatial distribution is also one of the key components of global climate change. Although passive microwave remote sensing technology is the most effective way of monitoring SM distribution at global scale, these kind products are generally limited by their low spatial resolution, which further prevents them from meeting the requirements of regional applications. On this basis, spatial downscaling has gradually become an alternative way of improving the spatial resolution of passive microwave SM products and a research hotspot in the field of remote sensing. Therefore, this paper reviewed and summarized the progresses on passive microwave SM spatial downscaling in the past 20 years. In terms of the downscaling methods, they can be divided into three key categories: empirical, semi-empirical, and physical model-based downscaling methods. The empirical downscaling method is simple and can easily achieve large-scale downscaling but lacks in physical background on the downscaling process. However, empirical methods have been widely used in passive microwave SM spatial downscaling study due to their simplicity and practicability. Physical model-based methods usually use data assimilation or/and land surface process models as the downscaling relational model. Usually, the process is complex, resulting in the low applicability of the physical model-based method, but this method can often obtain the downscaling results with good accuracy. The semi-empirical downscaling method generally can ensure the accuracy of the downscaling results and the operability of the method itself. However, the applicability of the semi-empirical method is still limited by its uncertainty related to the linking model for soil moisture expression and also some downscaling factors. Although numerous passive microwave SM spatial downscaling methods exist, the available downscaled SM products with good accuracy are limited. Currently, few passive microwave downscaling SM products are continuously produced, including the SMOS L4V5 SM product produced by BEC and the active and passive fusion SM products generated by NASA SMAP/Sentinel-1. Although the two kinds of downscaling SM products have the same spatial resolution (1 km), both suffers from the poor spatial coverage. In general, there are still some problems and challenges should be considered for current spatial downscaling study for passive microwave SM product. Aiming at obtaining the downscaling results with high spatial resolution, good accuracy, seamless spatial coverage, and daily temporal resolution. The uncertainty in the downscaling model (the relationship between SM and the downscaling factors), the uncertainty in the original passive microwave SM products and their incomplete spatial coverage, and the uncertainty in the downscaling factors (such as the influence from cloud cover and topography) are the issues should be well addressed. Overall, the development of the spatial downscaling study for passive microwave SM products will also provide more references and opportunities for promoting the application of SM products based on remote sensing in various fields including agro-forestry management, water resource assessment, and natural disaster monitoring.
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.
Abstract:Remote sensing image fusion, as the most challenging work in the field of image processing, has been an academic research hotspot. SAR has various features, such as all-weather service, and penetrating clouds. However, the image is difficult to interpret due to the speckle noise problem. In contrast, optical images can reflect the spectral information of ground objects, which are easy to interpret, but interference by clouds and fog can easily occur, resulting in information loss. The fusion of optical and SAR image data can realize the complementary information between different types of sensor imaging, which can facilitate the subsequent image analysis and interpretation.In this paper, the research status and future development trend of optical and SAR remote sensing fusion is reviewed. The introduction part presents the motivation of this paper by explaining the importance of image fusion. The second part outlines the classification of optical and SAR image fusion from traditional methods to deep learning methods. The third part presents the datasets in terms of optical and SAR images and explains each dataset in detail. The fourth part summarizes the difficulties and some challenging problems of optical and SAR image fusion and highlights the future trends in the field of optical and SAR image fusion.The future development trend of fusion has four main aspects, namely, datasets, time series image fusion, fusion evaluation system, and algorithm lightweight. First, having well-targeted and sufficient datasets is an important part of training excellent fusion models, and the production of data sets is the key to improving the applicability of future models. Second, time series optical images are being used as supplementary information to SAR data with the increasing amount of satellite time series data, thus improving the utilization of image fusion. Third, the evaluation of the image fusion performance is an essential topic. No unified evaluation metric is available for objectively and comprehensively evaluating fusion algorithms, and an evaluation metric system in the field of optical and SAR image fusion must be developed. Finally, the lightweight of deep learning algorithms is an important future research direction. This paper provides a reference for researchers in optical and SAR image fusion.
Abstract:Fine particulate matter (PM2.5) is a dynamic and complex mixture of particle matter with an aerodynamic diameter equal or less than 2.5 µm that can seriously affect the air quality and public health. High spatial and temporal resolution PM2.5 data is a basic requirement for public health risk assessment and epidemiological research. Compared with ground-based datasets, satellite remote sensing provides continuous, wide space coverage and low-cost observation, and the PM2.5 mass concentrations retrieval based on the satellite aerosol optical depth (AOD) has become a popular topic. This paper systematically scrutinizes the research on the near-surface PM2.5 concentration retrieved based on satellite AOD products. The basic method of estimating the PM2.5 concentration based on satellite AOD products is introduced, and the main satellite AOD products used for PM2.5 retrieval and their accuracy are described in detail. The existing PM2.5 estimation methods and their pros and cons are also discussed. Finally, the problems identified in PM2.5 retrieval research and the development direction of PM2.5 retrieval research are presented in the future.The scale factor method and the physical mechanism and statistical models can accurately estimate the PM2.5 concentrations at different degrees in different periods, but the scale factor method and the physical mechanism model are less used than the statistical model because of their limitations. Statistical models have been widely used and improved due to their unique descriptive ability of temporal or spatiotemporal heterogeneity and strong nonlinear description ability. However, the current PM2.5 retrieve research has three main limitations: 1. the non-random missing problem of satellite AOD causes missing PM2.5 data; 2. inaccuracy of retrieval models, and 3. Poor chemical composition estimation of PM2.5. Therefore, to accurately reveal the spatial and temporal trends of near-ground PM2.5 and improve the accuracy of the near-ground PM2.5 calculated from satellite AOD products, we predict several future research directions. First, the AOD products of new high-spatial-resolution (such as FY-4 and GF-5) and high-temporal-resolution (HIMAWARI-8/-9) satellites could greatly promote the research on PM2.5 estimation, which is of great significance to the reconstruction of PM2.5 concentrations with high spatial-temporal resolutions. Second, with the development of atmospheric detection technology, satellite-based, airborne, and ground-based lidar can obtain vertical distribution information, and the particle matter sensor carried on UAVs can achieve the vertical monitoring of PM2.5, which can be combined with optical remote sensing satellite and ground monitoring data to achieve three-dimensional PM2.5 concentration retrieval. Finally, PM2.5 chemical component information is particularly important for analyzing the cause of pollution and exposure characteristics, and its space–time change trend research is an important development direction. However, the ground PM2.5 component observation network is still imperfect, and overcoming the dependence on ground station network in satellite remote sensing estimation and achieving the high-precision retrieval of chemical composition need further study.This study is helpful in further understanding the principles, advantages, and disadvantages of different PM2.5 estimation methods, providing inspiration for the new development direction of near-surface PM2.5 concentrations retrieval based on satellite AOD products, and improving the accuracy and spatial-temporal resolution of near-surface PM2.5 concentrations retrieval.
Abstract:The COVID-19 epidemic swept the world and continues to spread. Without effective medical treatments and vaccine during the early stage of the pandemic, local governments in various countries had to lock down cities and adopt non-pharmaceutical interventions (NPIs), such as the stay-at-home order and social distancing. The NPIs against the COVID-19 epidemic have significantly changed the socio-economic activities in cities. However, the characteristics and patterns of urban socio-economic activities under this influence are still unclear. Given the development of earth observation technologies, large-scale changes in socioeconomic activities can be captured by satellites through remotely sensed Night-Time Lights (NTLs). In this study, we selected 20 major cities in the United States, including Atlanta, Chicago, Dallas, Denver, Detroit, Houston, Kansas City, Los Angeles, Miami, Minneapolis, New York, Orlando, Philadelphia, Phoenix, Pittsburgh, Salt Lake City, San Francisco, Seattle, Saint Louis, Washington D.C., to analyze the spatio-temporal variations of NTLs caused by the lockdown of cities. The first round of the COVID-19 epidemic occurred in the United States in mid-March 2020. Since March 2020, American cities have successively issued stay-at-home orders, but differences in the time and strictness of policy implementation were noted. Large cities have higher population density and intensity of social activities, so they are more susceptible to infectious diseases. The diversity of lockdown dates and the strictness of lockdowns in the cities in the United States are conducive to investigating the spatio-temporal variations of NTL. We acquired monthly averaged NTLS products of February, March, and April 2020, which were from Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) onboard the Suomi National Polar-orbiting Platform NPP. We further analyzed the spatial pattern, distance decay, and disparities in land use types of changes in the NTLs. Results show that the NTLs generally dimmed by 5%—8% in U.S. cities because of the lockdown of cities. In six cities, the luminous brightness dropped by more than 10%: Chicago, Dallas, Denver, Detroit, Minneapolis, and St. Louis. Among them, Minneapolis showed the largest decrease in luminous brightness, with a decrease of approximately 40% in March. The spatial change of the NTLs shows an obvious “core-periphery” pattern, indicating that the reduction of the NTLs declined with the distance from the city center, mainly because the central area of the city is a concentrated commercial area. After the closure of the city, commercial activities dropped significantly, resulting in an obvious reduction in NTLs around the city centers. The reduction of the NTLs varied among the diverse urban land use types. In New York, the NTL decreased the most on land for residence and aviation facilities by 12% and 11%, respectively. In Chicago, the NTL generally decreased by 20% in all types of urban land, and NTL recovered after one month of the lockdown of the cities in other urban lands, except the sports facilities land. This study only analyzed the spatio-temporal changes of NTLs. In the future, the results of this study can be combined with multi-source data to explain the driving force of NTL changes. Night-time light remote sensing effectively reflects the urban socio-economic dynamics with an important application in monitoring and assessing the socio-economic impacts of emergencies.
Keywords:night-time light remote sensing;COVID-19;NPP-VIIRS;spatio-temporal variation;the United States
Abstract:In view of the insufficient research on the response law of urban thermal environment and the Impervious Surface Distribution Density (ISDD), this paper selected Beijing as a case study. On the basis of the MODIS data, the average temperature difference between the urban construction and marginal areas was calculated as the UHI intensity (UHII). In combination with the characteristics of the UHII level, six typical regions (ⅠXierqi, Xisanqi, Huilongguan, ⅡDongsi, Xi’anmen, Xinjiekou, Ⅲthe middle of Fengtai District, ⅣSihui, Dingfuzhuang, Dongba, ⅤShijingshan District, Ⅵnear the Sijiqing bridge on the North Fourth Ring Road) with a high expected UHII level were selected to study the response relationship between the UHII level and the ISDD. Finally, Landsat data was used to invert the surface temperature to study the critical point and optimal scale of surface temperature response to the ISDD in urban construction areas and typical regions. Results show that the following. (1) The frequency of occurrence of UHII levels, their variation degree during the day and at night have a certain correlation with the ISDD, and the variation characteristics of the UHII under similar ISDDs are the same. (2) The distribution characteristics of the surface mean temperature in urban construction areas and the ISDD are obviously consistent. As the scale radius r increases, the response is more obvious. Besides, there are critical points for the response of the surface temperature to ISDD. The response of critical points gradually increases with the decrease of scale radius r. When scale radius r=1000 m, the ISDD reaches 60%, and its influence tends to weaken; the critical points of r=500 m and r=300 m are 69% and 83%, respectively. (3) The optimal scales of different typical regions have differences. The optimal scale of Xierqi, Xisanqi, and Huilongguan is 150 m, and that of Dongsi, Di’anmen, Xinjiekou, and Qingqiao near the fourth ring is 60 m. However, seasonal changes have little impact on the optimal scale. The optimal scale can measure the degree of fragmentation of the impervious surface distribution in different regions and the single degree of surface type to some extent. This study can provide a reference for urban planning and UHI governance.
Keywords:urban heat island effect;impervious surface distribution density;urban construction area;critical point;response law
Abstract:China has been experiencing rapid urbanization at an unprecedented rate with the significantly changing urban internal spatial structure. As an inevitable byproduct, Urban Villages (UVs), which refer to informal living spaces with substandard living conditions, have emerged in many newly and quickly industrialized regions and cities. Although UVs provide plenty of living spaces for floating populations, their poor living environment has a negative impact on the urban landscape and public health. Thus, obtaining the spatial distribution and environmental quality information of UVs in a timely manner and accurately for optimizing urban spaces and improving human settlements has practical significance. High-resolution Remote Sensing Images (RSIs) and Street View Images (SVIs) have been employed to quickly extract UV information. However, the combination of RSIs and SVIs for retrieving UV information has received little attention. In this study, we took Yuexiu District in Guangzhou City as the study area and then propose a UV identification method based on the GF-2 high-resolution RSIs and SVIs released by Baidu Company. First, street space quality information was derived from the SVIs using support vector machine and random forest. Then, on the basis of the pre-extraction results on the GF-2 images, multi-scale segmentation was performed based on object-based image analysis, including the building instance and block levels. Twenty-three features were obtained, including the spectrum, shape, texture, building structure, and scene from the RSIs, and five indicators were obtained to measure the street space quality on the basis of the SVIs. Finally, the random forest algorithm was applied to combine the features of the two kinds of images to identify the UVs. Experimental results demonstrate that the UV recognition based on RSIs has an overall accuracy of 94.5% and a Kappa coefficient of 0.58, and the overall accuracy and Kappa coefficient of the UV identification based on SVIs are 85.7% and 0.31, respectively. An overall accuracy of 96.1% and a Kappa coefficient of 0.67 were achieved by the fusion model of the two kinds of images, exhibiting the best performance in UV recognition. Street space quality, textural, structural, and shape features play an import role in the UV recognition based on the fusion model of RSIs and SVIs. The five indicators that measure street space quality on the basis of SVIs contributed 31.6% in feature importance to the fusion model. The information provided by RSIs from the bird view and the SVIs from the human perspective could complement each other, creating an outstanding feature space and reducing the misclassification phenomenon of UVs. The key to this method is integrating the information provided by SVIs into the UV extraction process based on high-resolution RSIs to obtain a highly stable and reliable UV classification result in Yuexiu District. Multi-source data fusion is an important method in improving the ability of RSIs, and other data should be collected to enrich the existing coupling methods and the technical system. This paper reveals that the fusion of high-resolution RSIs and SVIs in the feature level could improve the recognition accuracy of UVs, and the extracted UV distribution data can be used in urban planning and other studies related to urban development. The information in SVIs could be integrated into high-resolution RSIs and other data sources to assist in identifying informal living spaces, such as UVs. Therefore, retrieving highly accurate UV information is feasible through the combination of RSIs and SVIs.
Abstract:Multi-year mean phenology reflects the average state of vegetation growth and development rhythm and is one of the key parameters for predicting vegetation phenology. As an important source of spatial multi-year mean phenology, remote sensing is widely used for phenology detection. Different methods of multi-year mean phenology calculation are based on remote sensing. One is determining the phenological point of the annual time series curve first and then calculating the average (referred as the average method), and another is gaining the multi-year mean time series curve first and then determining the phenological point (referred as the reference curve method). The results of the above methods may be different. However, the uncertainty and its impacts need further elucidation. Hence, this study used the remote sensing vegetation index from 2001 to 2016 to extract the multi-year mean dates of the start of the growing season () using two methods in forests in China and detected the differences between the derived from the two methods (△) and the spatial pattern. Furthermore, a commonly used indicator in phenological research, that is, the temperature (Preseason Duration (PD)) based on , was used to explore the potential impact of the derived from different methods on the phenology–climate relationship. Results show that (1) the derived from different methods was significant different. The of the average method was generally smaller than that of the reference curve method (-2.6±2.2 days, accounting for 88%). The pixels with △>7 between the dynamic average method and the reference method and that between the fixed average method and the reference method accounted for 8.0% and 6.0% of the effective pixels, respectively, which are mainly distributed in the southeastern hilly area. (2) A significant spatial heterogeneity of △ showed a decrease with the increase of the annual average temperature (Slope=0.07 days/℃, P<0.01) and the decrease of the average annual precipitation (Slope=-0.0005 days/mm, P<0.01). (3) The PD derived from different methods was distinct. Approximately 40% of the effective pixels show a difference with PD > 5 days, and a half of them show a difference with PD>15 days, which are mainly located in the southeast hills and the southwest mountains. Overall, the achievements of this study provide a beneficial reference for the spatial parameterization of satellite-based phenology for modeling.
Keywords:multi-year average phenology;phenological preseason duration;remote sensing surface phenology;time series;forest
Abstract:The temporal and spatial distribution of carbon emissions and their heterogeneity are important topics in the study of ecological environment protection and climate change monitoring.Based on the fine analysis of the spatial distribution of carbon emissions in the Pearl River Delta urban agglomeration, this paper studies the spatial-temporal differences of carbon emissions in the area from 2000 to 2013 based on DMSP/OLS nighttime light images and land use data. This paper also reveals the spatial-temporal distribution characteristics, growth trend, and intensity trend of carbon emissions in different cities and land types.Results show the following. (1) The total carbon emissions of the Pearl River Delta urban agglomerations from 2000 to 2013 have been in the growth stage, but due to the 2008 financial crisis, the growth has turned from a high-speed growth stage to a slow growth stage. (2) The growth rate of per capita carbon emission intensity slowed down after the 2008 financial crisis. (3) The carbon emission intensity per unit of GDP experienced a small growth stage from 2005 to 2008, and then the overall trend is decreasing. (4) In terms of carbon emission intensity per square kilometer, the average carbon emission intensity of industrial and mining land transited from the growth stage before the 2008 financial crisis to the post-crisis reduction phase, while the average carbon emission intensity of urban land has been in a continuous growth phase. The carbon emissions of the Pearl River Delta urban agglomeration have obvious temporal and spatial differences before and after the 2008 financial crisis, and the continuous growth of carbon emissions from urban land will become a key issue for carbon emission reduction.This study can provide a scientific reference for carbon emission estimation and prediction, energy conservation and emission reduction, and ecological environment protection.
Keywords:carbon emission;spatial and temporal distribution;DMSP/OLS nighttime light image;land use data;refinement
Abstract:Change monitoring and disaster assessment of hurricane-damaged forests are important applications of remote sensing technology. The extraction of feature information from remote sensing images is very important to forest remote sensing monitoring. The combination of diversity features can effectively improve the accuracy of forest change monitoring. However, the current spatial information acquisition algorithms, such as texture features, still retain the traditional fixed computing model and do not fully consider the diversity of the spatial distribution of ground objects. When extracting texture features, the accuracy of texture features is affected when the sliding window is extremely large or small. Therefore, the calculation method of this kind of texture feature is always faced with the problem of hard balance between the number of features and the neighborhood reference range. Therefore, this paper focuses on forest change monitoring technology. To solve the above problems, a remote sensing monitoring method of forest destruction before and after hurricanes are proposed based on diversity features collaborative technology. First, the difference between the normalized vegetation index (NDVI) and the Enhanced Vegetation Index (EVI) before and after forest remote sensing image change is calculated. Second, the compound window technology is proposed to extract the texture features. Then, the texture features extracted from the remote sensing image and the spectrum of the remote sensing image are used to build a diverse characteristic combination model. This model can enhance the diversity of features. Finally, an improved rotating forest algorithm based on feature separation is proposed to reduce the direct feature correlation and improve the accuracy of the classifier. The study area is the remote sensing images of the Nezer forest in France before and after the hurricane, which were obtained from the Formosat-2 satellite. In the experimental part, the classification performance of the proposed algorithm was compared with those of six other methods. Experimental results show that compared with the traditional change detection methods based on spectral and texture features, the overall accuracy of the proposed method, the detection accuracy of the changed area, and the unchanged area improved by 3.68%, 6.53% and 3.46%, respectively. In addition, the sensitivity of the features extracted by different methods to the number of training samples was tested. The results show that the proposed method maintains high classification accuracy on different training sample numbers, and its overall accuracy and Kappa coefficient are better than those of the comparison methods. After the number of training samples reaches 50, the accuracy of the proposed method tends to flatten. The proposed method can effectively improve the performance of forest change monitoring. This method can be used to monitor the change of and damage to forests in real time and efficiently obtain information on forest disaster areas, providing important reference data for the emergency decision-making of forest resource management departments. Therefore, this method has a high practical value.
Abstract:Opencast coal mining activities would lead to negative impacts on regional ecological environment, ensuring efficient monitoring and regulation of mining activities would promote environmental protection and sustainable development. With the development of remote sensing technology and artificial intelligence, there is great potential in automatically detecting opencast coal mine areas from high spatial resolution remote sensing imagery. Aiming at the problem of low recognition rate of scene sub-region recognition via single-label learning algorithm, this paper proposes an opencast coal mine scene recognition method by integrating multi-label learning and the first law of geography. In order to distinguish the opencast coal mine scenes from their surrounding different scenes, six categories of mining labels and seven categories of non-mining labels are set, and 9768 sub-region images are annotated to create a multi-label dataset. The Inception-v3 model is trained using the created dataset to perform multi-label classification. For scene recognition, firstly, the remote sensing image covering the study area is divided into non-overlapping sub-regions of the same size and the multi-label classification is carried out on the divided sub-regions. Then, inspired by the first law of geography, the sub-regions containing the mining labels are assigned to the coal mine scene type or not according to the correlation between the labels and the completeness of the labels. Finally, all the sub-regions judged as the coal mine scene type constitute the opencast coal mine scene recognition result from the high spatial resolution remote sensing image covering the study area. The experimental results show that the recognition result of Shengli west opencast coal mine areas obtained by the proposed method is much closer to the ground truth than the comparative methods based on single-label learning. The F1 score of the proposed method reaches 0.857 in the multi-label classification, with an increase of 8 percentage points compared to the ResNet50 single-label learning method which has the best performance in the compared single-label learning methods. The proposed method can automatically extract the effective features of multiple labels in sub-regions and improve the performance of opencast coal mine scene recognition, its recognition results can provide data support for opencast mining management.
Abstract:Ship detection plays a crucial role in various applications and has drawn increasing attention in recent years. Deep learning methods based on CNNs, particularly SSD, have greatly improved detection performance due to their highly efficient feature extraction capability. However, SSD still has two problems. For instance, the detection network of arbitrarily arranged ship targets lacks a connection between high and low-level features and ignores contextual semantic information. Another problem is that natural factors such as light and clouds affect remote sensing images, thus ship detection may cause an imbalance of positive and negative samples.Aiming at solving the above issues, this paper proposes to achieve ship detection in remote sensing images by using a method based on Dense RFB and LSTM. This proposed method includes three elements. First, to enhance the detail feature extraction capability, this proposed method introduces a shallow feature enhancement module. This module draws on the idea of the human viewpoint, which uses Dense RFB feature reuse and expansion convolution to increase the receptive field. Second, to effectively extract deep semantic information and enhance the expressive ability of the network feature layer, a deep multi-scale feature pyramid fusion module (MFPF) is designed, as this proposed method draws on FPN and LSTM deconvolution and residual structure fuse deep multi-scale features. Finally, to solve the imbalance of positive and negative samples, the focal classification loss function is added, improving the accuracy of ship detection during training.The experiments were carried out on an optical remote sensing image dataset, in which only the ship dataset was used for training, validation, and testing. Results indicate that the proposed algorithm achieved an Average Precision (AP) of 81.98% and the detection speed reached 29.6 fps for ship targets, in which most ships were detected successfully. Moreover, for blurred, occluded, and partially-cropped ship targets, the algorithm’s detection effect is better than the traditional algorithm. Qualitative and quantitative results indicate that the generalization capability of the proposed method enhances ship detection.From this paper, we can draw three conclusions: (1) The proposed method can improve the extraction of detailed features and increase the receptive fields. (2) The focal loss function method shows good generalization capability. (3) The rotating box detection method is suitable for multi-scale and densely-arranged remote sensing images.
Abstract:With the rapid development of remote sensing satellite technology, the automatic extraction of high-resolution remote sensing images has become a popular research direction in the field of remote sensing. Deep learning methods have been applied in remote sensing image road information extraction and achieved significant results. However, due to convolution and pooling and other operations in network, road extraction methods based on the deep learning have some problems, such as the loss of spatial features and ground object details and the frequent occurrence of false extraction during road extraction. In order to solve these problems, this paper designs an improved road extraction semantic segmentation network model to mitigate the impact of the above network structure.The proposed method is based on ResNet and introduces coordinate convolution and a global information enhancement module before and after the coding structure, respectively. First, the network structure is mainly composed of residual units of ResNet, which has powerful feature extraction and multiplexing capabilities, and extracts road features of different scales and levels. Second, coordinate convolution reduces the spatial information loss and enhances the edge information. The coordinate convolution before the coding structure introduces spatial coordinate information, which is beneficial to enhancing the extraction of effective spatial information. Finally, global pooling can improve global context awareness. The global information enhancement module after the coding structure can effectively extract global context information through global pooling, thereby improving the accuracy of road classification and reducing the influence of natural scene factors, such as houses and tree shadows, to a certain extent.In this paper, the Massachusetts Roads dataset was used in the experiment, and the results obtained exhibited good accuracy. The Recall rate was 71.02%, the comprehensive evaluation index (F1 Score) was 76.35%, and the IoU reached 62.18%. The F1 Score and IoU indicators of the proposed method are approximately 1% higher than those in U-net and D-LinkNet and exceed those of DeeplabV3+ and Segnet, which are lower than D-LinkNet in the recall index only.The comparison of the experimental results indicates that the proposed method can effectively alleviate the spatial feature and context information losses on the basis of the deep learning road extraction method and completely extract the roads in remote sensing images. Moreover, the proposed method can effectively extract the road in the case of trees and building shadows, and multi-scale roads can also be accurately extracted.
Keywords:deep learning;remote sensing image;road extraction;coordinate convolution;global information enhancement module
Abstract:Waterlogging is the most serious meteorological disaster affecting crops in China besides drought. The occurrence of waterlogging has a great impact on the safety of people’s life and properties and the growth and development of crops. From July to August 2021, the precipitation in many places in northern China reached the historical observation extreme value, while the occurrence and development of surface waterlogging in the corresponding period and its temporal and spatial characteristics have not been effectively studied. In this study, the high-precision soil water data (0—10 cm) obtained from the daily soil water data of the soil water stations in Mainland China and the soil water daily products retrieved from passive microwave remote sensing satellite SMAP were used to calculate the soil surface relative water content combined with the soil field capacity data. The soil’s relative water content of greater than or equal to 90% for 10 consecutive days was taken as the standard. The spatial-temporal distribution of the waterlogging damage in Mainland China from July 1 to August 25 in 2021 was analyzed, and the results were comprehensively analyzed on the basis of the cultivated land distribution and precipitation data in Northeast China. The results show the following. (1) Compared with the original SMAP microwave soil moisture product, the accuracy of the fused soil moisture product is significantly improved. (2) The longest duration of soil relative water content greater than or equal to 90% in paddy fields in Northeast China was 56 days, indicating that the proposed method could accurately reflect the situation of relative soil water content. (3) Northeast and Northern China were severely affected, with the most extensive waterlogging in the west part of Heilongjiang Province and the entire area of the Hebei, Henan, and Shandong provinces. The arable area affected by waterlogging accounted for approximately half of the total arable land area in China, and the area of the worst-hit area was 1.940 ×105 km2. (4) The west part of the Heilongjiang province and the Hebei, Henan, and Shandong provinces received more precipitation than in previous years, which is consistent with the waterlogging disaster areas.
Keywords:waterlogged disaster;microwave remote sensing;relative soil water content;precipitation;spatial and temporal distribution
Abstract:The surface rupture investigation and disaster assessment of large earthquakes are important in earthquake emergency response. The rapid data acquisition and analysis of satellite images after earthquakes are of great importance for improving the timeliness of emergency response. After the Menyuan Ms6.9 earthquake, we quickly acquired GF-7 satellite images on January 8, 2022, and identified the overall distribution characteristics, structural style, and dislocation scale of the surface rupture, which provided an important reference for earthquake emergency response and field investigation. According to the deep analysis of high-resolution satellite images, the Menyuan Ms6.9 earthquake generated two co-seismic surface rupture zones, of which the north branch mainly developed along the western part of the Lenglongling fault, and the recognizable surface rupture zone is approximately 19 km-long. The south branch with a length of 2.3 km was distributed along the eastern part of the Tuolaishan fault. The co-seismic surface rupture produced a series of sinistral faulted geomorphic features such as gullies, diluvium, roads, ice layers, etc., and the maximum sinistral displacement was about 2.2 m. Deformation structures such as bulges, tension depressions and oblique tension fissures were formed along the surface fracture zone. The distribution of the latest co-seismic rupture coincided with the historical rupture, which reflected the in-situ rupture characteristics of strong earthquakes along the Lenglongling fault zone. Its seismic risk needed to be highly concerned and deserved in-depth study. This application work shows that the great potential and advantages of the China High-Resolution Satellites in major earthquake emergency.