最新刊期
卷 28 ， 期 2 ， 2024
-
摘要：Land-Use Scene Classification and change Detection (LUSCD) aim to recognize land-use types and monitor their changes by using Remote-Sensing (RS) images, which play an important role in urban planning and land-use optimization. In the era of RS big data, conventional hand-crafted feature-based methods are infeasible for LUSCD because the extracted features are not sufficiently discriminative for RS images with high complexity. As a novel data-driven paradigm for information extraction from RS images, deep learning provides a new solution for LUSCD. However, the existing publicly available datasets have limited samples and is thus unable to train a successful deep-learning model. Therefore, it has great significance in constructing an open and large-scale LUSCD benchmark.To advance the progress of LUSCD using deep-learning methods, this paper releases a large-scale scene classification and change-detection dataset termed Multi-temporal Scene Classification and Change Detection (MtSCCD). The RGB images in MtSCCD are cropped from large-size high-resolution RS images captured from the central areas of five China cities, namely, Hangzhou, Shanghai, Wuhan, Nanjing, and Hefei. The size of the cropped images is 300×300 pixels with the spatial resolution of around 1 m. MtSCCD has 10 land use classes, which are residential land, public service and commercial land, educational land, industrial land, transportation land, agricultural land, water body, green space, woodland, and woodland. Based on the cropped land-use images in MtSCCD, this paper constructs two sub-datasets termed MtSCCD_LUSC (MtSCCD Land Use Scene Classification) and MtSCCD_LUCD (MtSCCD Land Use Change Detection) for land-use scene classification (LUSC) and land-use change detection (LUCD), respectively. MtSCCD dataset has the following characteristics. (1) It is currently the largest publicly available LUSCD dataset, and both of the two sub-datasets (i.e., MtSCCD_LUSC and MtSCCD_LUCD) have 65548 images in total. (2) The images in MtSCCD are split into training set, validation set, and testing set according to the five cities. For example, images from three of the five cities are randomly split into training and validation set, whereas the rest remain to be the testing set. Therefore, MtSCCD has high extensibility, i.e., it can be easily extended to be a larger dataset. (3) For a deep-learning model, the training set and testing set are categorized from different cities, so it is beneficial to demonstrate the model’s generalization ability. (4) MtSCCD has high intra-class diversity, making it a challenging dataset.Based on MtSCCD_LUSC and MtSCCD_LUCD, this paper evaluates several deep-learning feature-based methods for LUSC and LUCD. Specifically, AlexNet, VGG networks (i.e., VGG16 and VGG19), GoogLeNet, and ResNet networks (i.e., ResNet18, ResNet50, and ResNet101) are selected to extract deep-learning features that are then fed into SVM for LUSC. We also evaluate DenseNet, EfficientNet, SENet, ViT, and SwinT for LUSC. Two kinds of LUCD approaches including conventional classification-based methods and current similarity-based methods have been evaluated. Experimental results show that the highest overall accuracy of MtSCCD_LUSC dataset is around 76%, indicating much room for improvement. Regarding LUCD, similarity-based methods particularly similarity learning-based ones outperform classification-based methods by a significant margin, providing a promising research direction for LUCD.This paper presents the currently largest scene classification and change-detection dataset MtSCCD based on high-resolution RS images of the central area of five China cities. MtSCCD contains two subsets MtSCCD_LUSC and MtSCCD_LUCD. Both had 10 land-use types and 65548 images in total. Based on the two sub-datasets, this paper evaluates the performance of several deep networks for scene classification and change detection, expecting to provide baseline results for related researchers. We hope that the MtSCCD dataset can promote this progress in land-use type recognition and monitoring.关键词：land use;scene classification;change detection;dataset;information extraction;feature extraction;deep learning;convolutional neural network- 1368
- |
- 722
- |
- 0
发布时间：2024-03-22 -
摘要：Sample collection is one of key research foundations for wetland mapping. It plays an important role in classifier training and accuracy validation. Generally, wetland samples are produced by visual interpretation based on high-spatial-resolution images or automatic generation based on multi-source existing dataset. The visual interpretation is time and labor consuming and cannot meet the demand for large-scale wetland classification. The automatic sample-generation method is unsuitable to detailed-type wetland mapping due to the diversity of wetlands and classification-scheme inconsistency of existing wetland datasets. Thus, an efficient and accurate sampling method is in demand for large-scale and detailed-type wetland mapping.In our study, we collected a series of auxiliary datasets and developed an efficient solution for continental-scale wetland sample generation by combining automatic sampling method and visual interpretation. In the first part, the samples of five wetland types can be automatically generated by rule filtering based on multi-source existing datasets. River, lake, and reservoir samples were created using the JRC Global Surface Water, Global River Widths from Landsat and HydroLAKES datasets. Coastal swamp (mangrove) samples were produced by using Global Mangrove Watch dataset. Tidal flat samples were generated using the Global Intertidal Change dataset. In the second part, by combining time series of MODIS NDVI images and existing auxiliary datasets, we first produced potential wetland samples for coarse wetland types (i.e., vegetated wetland samples and inundated wetland samples). Then, we identified them by visual interpretation based on the Google Earth Engine platform, Google Earth software, and Collect Earth software. We applied our sample method in our study area, and produced continental-scale and detailed-type wetland samples.Results indicated that the total wetland samples in our study area was 150688, among which 141412 points were inland wetland samples, 11563 were coastal wetland samples, and 17693 were human-made wetland samples. Among the 13 wetland sub-categories, lake accounted for the largest proportion (39.22%) and primarily distributed the northern and central of study area, whereas lagoon accounted for the smallest proportion (0.19%), mostly scattered in coastal region of the study area. Samples of river, reservoir, inland swamp, and inland marsh also shared a considerable amount, accounting for 16.93%, 9.86%, 7.16%, and 11.12% of total wetland samples, respectively. River samples were primarily distributed north and south of the study area, and reservoir samples were primarily scattered south of the study area. Meanwhile, inland swamp and inland marsh samples were mostly distributed northwest and south of the study area.This study successfully produced stable and high-quality wetland samples at continental scale. The generated samples shared sufficient quantities and reasonable spatial distribution, which can lay a good foundation for classifier training and accuracy validation. Meanwhile, by combining the multi-source thematic datasets and multiple platform, our designed sample solution can make full use of the existing database and greatly reduce manual workload. It can also create high-quality samples for complex wetland types, such marsh, swamp and floodplain. Overall, the designed sample method in our study was efficient and reliable, which has significance for large-scale wetland mapping.关键词：remote sensing;wetland;sample production;multi-source thematic data;visual interpretation- 1377
- |
- 370
- |
- 0
发布时间：2024-03-22
Data Articles
-
摘要：Flood disasters are a great threat to the national economy and people’s property along the lakes and rivers in China. Synthetic Aperture Radar (SAR) adopts active imaging methods that can realize all-weather imaging and ensure continuous observation of flood disaster areas under severe weather, such as heavy rains and clouds. The current flood-monitoring methods based on SAR images often have problems, such as difficulty in threshold selection, high computational cost, or inefficient use of time-series information. Aiming at the above problems, this paper makes full use of information from time-series SAR image sequence to design an effective and stable method of monitoring flood, which can be adapted to complex areas.Through preprocessing and statistical analysis of the image sequence, two normalized difference indices including submerged range extraction index and submerged range in vegetation area extraction index are designed and applied to calculate the candidate area of flood inundation. Then, the adaptive selection method of threshold for flood extraction is given based on the stability assumption of vegetation seasonal distribution in the same area. Finally, considering the characteristics of the surrounding features of the lakes in China, a post-processing process is designed. The process involves removing spots and holes, excluding areas with large slopes and filtering out fragmented areas with large rectangular degrees. Post-processing is conducted to optimize the extraction area for the final results of flood-inundation range.In the experiment, this paper takes the East Dongting Lake basin as the main research area to verify the effectiveness of the proposed method by comparing the extraction accuracy with the other four methods. Experimental results prove that the overall extraction accuracy of the proposed method is higher than that of all comparative methods. To achieve the purpose of flood-disaster monitoring and evaluation, an analysis of the flood-disaster situation in the East Dongting Lake Basin in 2020 and an analysis of flood submerged land cover types are conducted. The method has also been successfully applied to the data of the East Dongting Lake basin in previous years and the Poyang Lake basin in that the method can be applied across time and space.Based on the time-series SAR image sequence, this paper proposes an effective method, forms a detailed process flow, and constructs a general framework for flood monitoring. The proposed method has advantages of simple parameter setting and low user dependence on threshold determination. Experiments show that the method has high extraction accuracy of submerged areas with good robustness and versatility. The proposed method can be applied to different flood-monitoring scenarios across time and space and can preliminarily distinguish different attributes of submerged areas. Thus, a certain reference is provided for flood-disaster monitoring, assessment, and early warning in other regions.关键词：Flood disaster;synthetic aperture radar;time series monitoring;hydrological remote sensing;SAR;Dongting Lake- 3565
- |
- 624
- |
- 0
发布时间：2024-03-22 -
摘要：Leaf-Area Index (LAI) is an important canopy structural parameter that accounts for the qualities of the growth state of vegetation. MODIS LAI product is one of the most commonly used remote-sensing LAI products in the world. However, the quality of MODIS LAI products varies with different situations because of variations in surface heterogeneity, data quality, and model accuracy, among others. The LAINet instrument, which is based on the wireless sensor network, can automatically obtain the LAI measured data with more intensive time frequency. It can provide strong support for the validation of satellite remote-sensing LAI products.This article aims to validate the accuracy and evaluate the stability of MCD15A3H LAI products (Colletion 6) with time-series ground-observation data. The specific objectives include the following: (1) generation of reference products that meet MODIS LAI product validation based on ground network observation time-series data, (2) validation of accuracy of MODIS LAI products based on reference products, (3) evaluation of the time-series stability of MODIS LAI products, and (4) analysis of the reasons for the difference between the MODIS LAI product and the measured LAI.This work adopts an indirect comparison method, that is, establishing an empirical regression model based on time-series ground-measured LAI and high-spatial-resolution satellite remote-sensing vegetation index to obtain a high-spatial-resolution satellite remote-sensing LAI reference map. The resolution of the reference map is upscaled to the same resolution as those of MODIS LAI products. Finally, we validate the accuracy and evaluate the stability of MODIS LAI products with the upscaled satellite remote-sensing LAI image.Results show that compared with the reference true value of Landsat 8, the quality of the growing stage (RMSE2018=1.17, RMSE2019=1.14) is better than that of the senescence stage (RMSE2018=1.39, RMSE2019=1.84), and MODIS LAI is generally underestimated to Landsat LAI, especially in the late growing stage. MODIS LAI products can portray the seasonal characteristics in the vegetation growth and falling stages in time series, but the instability in the early period of growth is stronger than that in the later period. The difference in observation methods is the main reason for the underestimation of MODIS LAI, that is, the LAI value of the remote-sensing sensor is affected by the decrease in chlorophyll in the late growing season because the sensor observes from the space platform in the downward direction. Conversely, the LAINet instrument observes from the bottom of the canopy in the upward direction, which is primarily affected by the canopy-gap fraction. However, it is insensitive to changes in the pigment of leaves.The accuracy validation and stability evaluation results of MODIS LAI products show that the time-series LAI can be retrieved using ground-based data and satellite remote-sensing data. However, considering the difference between the observation objects and algorithm principles of MODIS LAI and LAINet LAI is necessary when using the late-growing-season data of corn crops.关键词：MODIS;leaf area index;LAINet;accuracy validation;stability evaluation- 833
- |
- 167
- |
- 0
发布时间：2024-03-22 -
摘要：Using remote-sensing technology to obtain information about burned areas is important for ecological environment monitoring. High-resolution data are more suitable for extracting small-scale burned areas. To develop the fire monitoring ability of domestic remote-sensing data and improve the extraction efficiency and accuracy of a small-scale burned area, two GF-1 WFV images (before and after fires) and multi temporal FY-3D MERSI fire products are used to extract burned areas for two study areas, respectively, located in the Tibetan Autonomous County of Muli and Xichang City, Sichuan Province.The reference true values are obtained by human-computer interaction for verification. The results of burned areas extracted by neural network classification are compared with the result of the proposed method. Our results show that the accuracy of burned areas detected by the proposed method is higher than that by neural network classification, and the Kappa coefficients in two study areas are 0.82 and 0.87, respectively. The regions of commission and omission are usually distributed at the edge of the burned area patch. The distribution of burned area in Xichang is more compact than that in Muli, so the accuracy of burned area mapping in Xichang is higher.The method can fully combine the advantages of the two kinds of data, reduce the uncertainty and time cost caused by sample selection, and extract the small-scale burned area quickly and accurately. Fully exploiting the temporal, spatial, and spectral characteristics of fire points and burned areas can compensate for the shortcomings of GF-1 WFV images in temporal and spectral resolution. Meanwhile, the method can fully combine the two kinds of data and minimize the impact of the difference of spatial resolution. In the future, the method can be improved using a higher accuracy of fire-point products. The accuracy of the reference true value of the burned area can be improved through field investigation.The method is primarily divided into two partsrough extraction and fine extraction. In rough extraction, according to the relationship between fire points and the formation of burned areas, the fire-point pixels are selected and expanded into the rough range of burned areas by combining temporal, spatial, and spectral characteristics. Temporal characteristic refers to fire points with concentrated occurrence time that easily form burned areas; spatial characteristic refers to fire points with concentrated location that easily form burned areas, and burned pixels are usually adjacent to fire-point pixels; spectral characteristics refer to pixels with higher NDVI difference before and after fire, which may be burned pixels. In fine extraction, the land-cover types included in the burned area are determined according to the number of fire-point pixels. The segmentation threshold is determined using the iterative-threshold method for each land-cover type. Burned pixels and unburned pixels in each land-cover type are classified using the segmentation threshold. The small patches are removed to obtain the result of burned-area extraction.关键词：remote sensing;burned area;fire point product;FY-3D MERSI;GF-1 WFV;NDVI;segmentation threshold- 793
- |
- 177
- |
- 0
发布时间：2024-03-22
Ecology and Environment
-
摘要：Compared with high-spectral thermal infrared data, ultra-spectral thermal infrared data contains enhanced atmospheric vertical information of ozone (O3) and methane (CH4). This finding indicates the possibility to improve the accuracy of retrieved O3 and CH4 profiles. Due to the narrow channel intervals of the ultra-spectral thermal infrared data, abundant special information and redundant information is induced. However, information cannot be detected by channel-selection methods for high-spectral thermal infrared data, thereby impeding the superiority of ultra-spectral data for the retrieval of trace-gas profiles. As such, a novel channel-selection method based on the gas-sensitivity and weighting-function characteristics (OWSP) has been promoted, aiming to enhance the retrieval efficiency and accuracy of O3 and CH4 profiles from ultra-spectral thermal infrared data. The method comprises two steps. First, the sensitivities of the channels to different gases are analyzed, and the signal-to-interference ratio (rSTI) are obtained. On this basis, channels with abundant information for retrieved gas and insensitivity to other gases can be detected, which are taken as the initial channel group. Second, a strategy of optimizing the distribution of the weighting function is promoted based on the features of Jacobians to O3 and CH4. The channel information content can then be quantified by the optimized weighting function. An iterative approach is applied to select the optimal channel group to retrieve atmospheric profiles. In this paper, the promotion effect of OWSP method for O3 and CH4 profile retrieval from ultra-spectral thermal infrared data is evaluated by applying in the winter and summer atmospheric situation of the regions of Alxa Desert (AL), Beijing Tianjin district (JJ), Yangtze River Basin (YRD), and Pearl River Basin (PRD). The optimal sensitivity profile (OSP) method, which suggests good performance for high-spectral thermal infrared data in literature, is used in the control group. By comparing with the channel selection results of OSP method, it shows that the OWSP method can effectively screen the correlated channels with similar information for the strong infrared radiation gas, O3. It can also select some channels with special information. Conversely, it has relatively low sensitivity for the weak infrared radiation gas, CH4, thereby ensuring the accuracy and efficiency of the subsequent retrieval process. The retrieval results of O3 and CH4 profiles with the channel group selected by the two methods further prove that the OWSP method can efficiently improve the accuracy of the retrieved profiles in most situations, and the mean retrieval accuracy of the O3 and CH4 profiles increase 9.30% and 4.90%, respectively. This research has important theoretical and application value, which can provide some essential technological support for the development and data application of ultra-spectral TIR sensor for our country in the future.关键词：remote sensing;thermal infrared data;ultra-spectral;Channel selection;Jacobians;Gas sensitivity;O3 and CH4 profile retrieval- 652
- |
- 137
- |
- 0
发布时间：2024-03-22 -
摘要：We comprehensively evaluated various Remote Sensing Precipitation Estimates (RSPEs) to identify the ones that can better capture the precipitation pattern in the Weihe River Basin. Our findings can serve as an important scientific reference for the evaluation and management of water resources in the basin and provide favorable support for the implementation of environmental protection and high-quality development planning in the Yellow River basin.Based on the gridded precipitation data of CMA and the five popular RSPEs (including CHIRPS v2.0, CMORPH v1.0, PERSIANN-CDR, TRMM 3B42, and MSWEP v2.0), we comprehensively evaluated the basic skill of precipitation products by using the four statistical metrics, namely, Pearson correlation coefficient (CORR), bias, Root-Mean-Square Error (RMSE), and Kling-Gupta Efficient (KGE). We further used three categorical metrics, namely, Probability Of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI). Then, the hydrological simulation skill of RSPEs was also assessed by the traditional lumped hydrological model of ABCD and nash efficiency coefficient (NSE).All five RMPEs can capture the spatial distribution of precipitation. Among them, MSWEP, based on multi-source weighted-ensemble precipitation, can better capture the spatial heterogeneous of precipitation with superior performance. The spatial distribution of PERSIANN smoothly performed and was underestimated in most areas, resulting in lower performance. At the interannual level, all RMPEs generally performed well in the upper Weihe River, and TRMM was excellent, followed by MSWEP, whereas PERSIANN performed poorly. MSWEP products performed well for basic statistical skills, with lower RMSE, higher CORR, and KGE. However, the CHIRPS and PERSIANN had poor basic statistical skills. For the categorical skill of precipitation, the PERSIANN exhibited good skill for light rain, followed by MSWEP. Additionally, all RMPEs performed better than the other three basins in the upstream of Weihe River. To detect moderate rain and heavy rain, the performance of the PERSIANN was degraded. The MSWEP product had a better POD for these two types of rainfall, but its FAR was also higher. In terms of hydrological simulation performance, the hydrological simulation performance of the TRMM was the best, indicating that the retrieval algorithm of RMPEs based on active microwave had a high hydrological application prospect, followed by the MSWEP and COMRPH. The poor performance was the infrared/near-infrared-based CHIRPS and PERSIANN products, whose retrieval algorithms required further improvement in the climate-sensitive transition region.TRMM products performed better in capturing the temporal and spatial patterns of precipitation. The multi-source integrated product of MSWEP was significantly better than the other four products. The predictability of each precipitation estimate to moderate and heavy rain was unsatisfactory and was especially poor for the latter one. Using TRMM precipitation as the input of ABCD model, the performance of simulating runoff was significantly better than other remote-sensing products in the four sub-basins of the Weihe River, followed by MSWEP and COMRPH.关键词：remotely sensed precipitation;Wei river basin;ABCD model;performance evaluation;MSWEP- 813
- |
- 220
- |
- 0
发布时间：2024-03-22 -
摘要：Satellite remote-sensing precipitation products are currently the main source for obtaining large-scale and continuous precipitation observations. However, currently available satellite remote-sensing precipitation products have coarse spatial resolution and suffer from certain systematic biases. Thus, this paper aims to downscale the precipitation data and remove its inherent systematic biases.This paper proposes a two-stage Spatial Random Forest (SRF) method (SRF-SRF) by fully considering the influence of high-resolution environmental variables (including topography, NDVI, surface temperature, latitude, and longitude) on the precipitation and the spatial correlation of neighboring remotely sensed precipitation (stations). Taking the Global Precipitation Measurement Mission (GPM) monthly precipitation data of Sichuan Province from 2015—2019 as an example, its quality is enhanced with the help of SRF-SRF. The calculation results are compared with those of seven existing methods, including Geo-Weighted Regression (GWR), Back-Propagation Neural Network (BPNN), Random Forest (RF), Kriging interpolation of station precipitation (Kriging), Geographic Difference Analysis correction after downscaling by SRF (SRF-GDA), SRF correction after downscaling by bilinear interpolation (Bi-SRF), and annual precipitation downscaled by SRF. Subsequently, the results are scaled by month and corrected using SRF (SRFDis).This paper proposes a two-stage satellite precipitation product-quality enhancement method that considers spatial correlation. The method takes into account the spatial autocorrelation between precipitation and combines downscaling and calibration while integrating environmental factors. Accordingly, the spatial resolution and accuracy of precipitation products improve. Experimental results show that the new method outperforms the other seven classical methods and is more applicable to the quality improvement of precipitation products in complex terrain.Experimental analysis shows the following(1) At the monthly scale, compared with the original GPM, the mean absolute error (MAE) of SRF-SRF is reduced by 19.51%, and the medium error (RMSE) is reduced by 16.35%. The accuracy is better than those of other methods. At the seasonal scale, SRF-SRF has the smallest error in winter and the largest error in summer, but its calculation accuracy is better than those of other methods. At the annual scale, the four SRF-based methods (including SRF-SRF, SRF-GDA, Bi-SRF, and SRFdis) outperform GWR, BPNN, and RF. The accuracy of SRF-SRF is higher than that of Bi-SRF and SRF-GDA. (2) The spatial-distribution continuity of SRF-SRF precipitation products is better, and the local precipitation details are significantly improved. (3) The spatial correlation of precipitation plays an important role in the improvement in GPM precipitation quality. (4) SRF-SRF based on the monthly scale is better than SRFdis based on the annual scale. This finding indicates that NDVI can be used for precipitation-quality enhancement at the monthly scale in Sichuan province.关键词：remote sensing;precipitation;downscaling;Point and surface fusion;Random Forest;GPM;machine learning- 1080
- |
- 206
- |
- 0
发布时间：2024-03-13 -
摘要：Sea level is an important parameter to ensure coastal safety, monitor marine climate, and maintain elevation data. In recent years, the remote-sensing method using ground-based GNSS reflection signal can be used for sea-level monitoring. Compared with the traditional sea-level measurement method, GNSS multipath reflectometry (GNSS-MR) technology has the advantages of low cost and continuous tracking and can make all-weather, all-day observations. However, GNSS-MR technology is limited by two problems: low accuracy and low time resolution. The time resolution of retrievals can be improved by acquiring more observation data from more satellite systems. In this work, a robust regression strategy based on the IGGIII scheme is proposed to address the two limitations. This method uses the SNR data of GPS, GLONASS, Galileo, and Beidou. The Lomb-Scargle periodogram method in the classical tide level-inversion principle is used to obtain the sea-level estimates of each frequency band from quad-constellation. Then, a specific time window is established. The state-transition equation set is established in each time window considering the sea-surface dynamic change and tropospheric delay. Finally, the sea-level time series is solved by a robust estimation model. To prove the feasibility and effectiveness of this method, BRST station in France and HKQT station in Hong Kong are selected to validate the performance of the proposed method. The Root-Mean-Square Errors (RMSEs) between sea-level combined retrievals of multi-GNSS signals and the tide gauge records are calculated. The RMSE of BRST station is 12.43 cm, which is about 40%—60% higher than the single-signal results of each system. The RMSE of HKQT station is 7.09 cm, which is about 72% higher than the results of the four systems. BRST and HKQT stations can formulate a 10-min sea-level time series, which greatly improves the time resolution of sea-level retrievals compared with single-signal retrievals. Comparing the inversion results of the two stations, we conclude that using robust regression strategy based on the IGGIII scheme can lead to a clear increase in precision and thus achieve a higher temporal sampling because of the more frequent GNSS retrievals and better retrieval combination strategy. The estimated value of sea level well agrees agreement with the data of tide-gauge records and can clearly describe the sea-level fluctuation. In essence, it is a method of quality control and optimal valuation for GNSS-MR that is theoretically suitable for different geographical environments.关键词：remote sensing;GNSS-MR;sea level estimation;multiple system;robust regression;IGGIII- 820
- |
- 275
- |
- 0
发布时间：2024-03-22
Atmosphere and Ocean
-
摘要：Building change detection is essential to many applications, such as monitoring of urban areas, land use management, and illegal building detection. It has been seen as an effective means to detect building changes from remote-sensing images.This paper proposes an object-based Siamese neural network, labeled as Obj-SiamNet, to detect building changes from high-resolution remote-sensing images. We combine the advantages of object-based image analysis methods and Siamese neural networks to improve the geometric accuracies of detected boundaries. Moreover, we implement the Obj-SiamNet at multiple segmentation levels and automatically construct a set of fuzzy measures to fuse the obtained results at multi-levels. Furthermore, we use generative adversarial methods to generate target-like training samples from publicly available datasets and construct a relatively sufficient training dataset for the Obj-SiamNet model. Finally, we apply the proposed method into three high-resolution remote-sensing datasets, i.e., a GF-2 image-pair in Fuzhou City, and a GF2 image pair in Pucheng County, and a GF-2—GF-7 image pair in Quanzhou City. We also compare the proposed method with three other existing ones, namely, STANet, ChangeNet, and Siam-NestedUNet.Experimental results show that the proposed method performs better than the other three in terms of detection accuracy. (1) Compared with the detection results from single-scale segmentation, the detection results from multi-scale increases the recall rate by up to 32%, the F1-Score increases by up to 25%, and the Global Total Classification error (GTC) decreases by up to 7%. (2) When the number of available samples is limited, the adopted Generative Adversarial Network (GAN) is able to generate effective target-like samples for diverting samples. Compared with the detection without using GAN-generated samples, the proposed detection increases the recall rate by up to 16%, increases the F1-Score by up to 14%, and decreases GTC by 9%. (3) Compared with other change-detection methods, the proposed method improves the detection accuracies significantly, i.e., the F1-Score increases by up to 23%, and GTC decreases by up to 9%. Moreover, the boundaries of the detected changes by the proposed method have a high consistency with that of ground truth.We conclude that the proposed Obj-SiamNet method has a high potential for building change detection from high-resolution remote-sensing images.关键词：change detection of remote sensing;Siamese Neural Network;object-based multi-scale analysis;fuzzy sets fusion;Generative Adversarial Network;Very High Resolution remote sensing images- 1297
- |
- 386
- |
- 0
发布时间：2024-03-22 -
摘要：Although the Deep Semantic Segmentation Network (DSSN) has notably enhanced remote-sensing image semantic segmentation, it still falls short of human experts’ visual interpretation. Unlike DSSN’s data-driven, pixel-level optimization, human experts rely on visual features, semantic insight, and prior knowledge for remote-sensing image interpretation. DSSN’s pixel-level approach is constrained by spatial scale, lacking comprehensive target inference and struggling to bridge structured data and unstructured knowledge. In response to the two issues above, this paper proposes a geographic knowledge graph-guided deep semantic segmentation network for remote-sensing imagery. We use the ground-object semantic information and geoscience prior knowledge extracted from the geographic knowledge graph to construct loss constraints, thereby autonomously guiding the training process of DSSN.The essence of our approach lies in the intricately crafted design of loss constraints. These loss constraints include the entity-level connectivity constraint and the inter-entity symbiosis constraint. The former calculates the loss in the unit of connected domain entities instead of pixels to achieve overall constraints on the entity. The latter embeds the spatial symbiosis knowledge quantified by the symbiosis conditional probability into the data-driven DSSN to constrain the spatial distribution of segmented entities. The entity-level connectivity constraint guides DSSN to autonomously learn entity-level feature representations during training. Accordingly, the segmentation results become more holistic and suppresses blurry boundaries and random noise. The inter-entity symbiosis constraint adjusts the spatial distribution of entities according to the spatial semantic information and the prior geoscience knowledge. This adjustment realizes the automatic optimization of the spatial distribution of segmented entities.Extensive experiments show that under the guidance of the entity-level connectivity constraint and the inter-entity symbiosis constraint, DSSN can complete the learning of entity-level features. It can also automatically optimize the spatial distribution of ground objects based on spatial symbiosis knowledge, thereby effectively improving the performance of remote-sensing image semantic segmentation.Our novel geographic knowledge graph-guided approach to deep semantic segmentation in remote-sensing imagery has successfully addressed the challenges posed by DSSN’s pixel-level optimization. By incorporating entity-level connectivity and inter-entity symbiosis constraints, we have enabled DSSN to autonomously learn comprehensive feature representations and optimize spatial distribution. The resulting improvements in semantic segmentation performance showcase the potential of merging domain-specific knowledge with data-driven techniques, bridging the gap between automated methods and human interpretation in remote-sensing image analysis.关键词：Geographic knowledge graph;deep semantic segmentation network;entity-level connectivity constraint;spatial symbiosis knowledge constraint;geographic knowledge embedding optimization- 578
- |
- 344
- |
- 0
发布时间：2024-03-22 -
摘要：To address data acquisition and labeling data in the training process of SAR ship-detection network based on deep convolutional neural network, we propose a SAR ship-detection framework via generative knowledge transfer of a knowledge transfer network for SAR image generation and a SAR ship-detection network. The knowledge transfer network consists of three parts: a cycle consistency GAN to synthesize virtual features which have spatial distribution of optical image domain and feature distribution of SAR image domain as well; We further use an identity loss to encourage pseudo-SAR images generated by the knowledge transfer networks to have more of the intrinsic features of SAR images. To alleviate the SAR feature confusion issue, we introduce a feature boundary decision loss to maximize the decision boundary of real SAR features and the pseudo ones. Therefore The knowledge transfer network generates pseudo-SAR images consistent with the spatial distribution of labeled optical remote-sensing images and has a feature distribution similar to those of SAR images. Our proposed method is evaluated from three aspects: (1) The evaluation on the generated pseudo-SAR images. When the object detection network is trained on 70% of SSDD and the pseudo-SAR images, The remaining 30% of SSDD is test set, the AP can reaches 97.50%. As for 0%, 10%, 20%, 30%, and 50% of SSDD, the AP is 64.55%, 91.14%, 94.69%, 96.21%, and 96.84%, respectively. When there is no real SAR images involved in the training process, Ap can still reach 64.55%. (2) Ablation study on loss functions. On the basis of using cycle consistency loss in knowledge transfer network，the best performance comes when applying both the identity loss and the feature boundary decision loss, the AP reaches 64.55%. (3) The evaluation on the ship detection network. The generated pseudo-SAR images in this paper are used in the training process of SSD, Faster R-CNN and YOLOv3 detection networks, which can increase the object detection network to learn more parameters suitable for SAR images, thus improving the detection effect of the network. Experiments in the above three aspects prove the effectiveness of the proposed method.关键词：SAR;object detection;deep learning;image generation;generative adversarial networks- 2634
- |
- 315
- |
- 0
发布时间：2024-03-22 -
摘要：With the rapid development of spatial technology, the resolution of remote sensing images gradually improves. The detailed information and spatial information contained in remote-sensing images are also richer. The ensuing problems are that the difference between various categories becomes and the difference between the same categories becomes larger, i.e., the phenomenon of the same spectrum of foreign objects and the different spectrum of the same objects is serious. However, the existing dual-modal segmentation methods do not extract the dual-modal feature information of remote-sensing images separately, and the fusion features are insufficient. The details of upsampling recovery are also insufficient, resulting in the inability to accurately and efficiently learn remote-sensing image information, thereby resulting in segmentation errors, edge blur, and other problems.This study proposes a high resolution remote-sensing image segmentation based on dual-modal efficient feature learning. The algorithm designs appropriate encoders for different modal remote sensing images, efficiently extracts dual-modal features, and reduces the differences between different path features through interactive reinforcement modules. Then, the dual-modal feature aggregation module and the deep feature-extraction module are proposed to further fuse and extract the dual-modal features. As a result, the network can fully learn the complementary information of the dual-modal. Finally, a multi-layer feature upsampling module is proposed, which uses high-level features with rich semantic information to weight the low-level features with rich detail information. Gradual upsampling is then conducted to achieve efficient feature recovery and improve segmentation performance.In this paper, experiments on the Potsdam and Vaihingen datasets demonstrate that the overall accuracy reaches 94.52% and 90.45%, respectively. Experimental results show that the segmentation effect of the proposed algorithm is better than that of existing algorithms. The proposed algorithm can efficiently extract and fuse the multi-modal complementary features of high resolution remote-sensing images and improve the segmentation accuracy of remote-sensing images.This study proposes a high-resolution remote-sensing image segmentation based on dual-modal efficient feature learning. Experiments on the ISPRS Potsdam and Vaihingen datasets show that the proposed model is more suitable for segmenting low vegetation and trees, buildings, and roads with very similar spectral features. It can also achieve the accurate segmentation of small targets, such as cars. However, the complexity of the model needs to be further reduced, and much room for improvement in accuracy remains. In the future, a better segmentation network will be designed to fuse more than two modal features and thus obtain more feature information to achieve more accurate remote sensing image segmentation.关键词：remote sensing image segmentation;efficient feature extraction;integration;dual-modal feature aggregation;deep feature extraction;multilayer feature upsampling- 216
- |
- 261
- |
- 0
发布时间：2024-03-22 -
摘要：Hyperspectral image (HSI) classification is one of the fundamental tasks in the field of applied remote sensing. As technological advances have increased the spatial and spectral resolutions available for data acquisition, the problem of achieving accurate HSI classification is becoming more challenging. This problem is especially true for the HSI data with small labeled training samples and insufficient utilization of spatial-spectral information in HSI classification models. Aiming at these problems, this paper proposes a new HSI classification method (expressed as SKERW_SVM) by combining the Superpixel Dimension Reduction (SDR) with post-processing optimization.First, we develop a Superpixel Sparse Linear Discriminant Analysis (SSLDA) method by combining Regional Clustering (RC) with SLDA. In the SSLDA method, the RC is applied to construct a homogeneous local neighborhood set with high spatial correlation and spectral similarity for each pixel of the HSI. The SLDA is used to extract superpixel sparse mixture features that can fully characterize spatial-spectral information and related change information of the HSI based on the constructed homogeneous regions. Then, the extracted sparse mixture features are inputted into the support vector machine to generate the class probabilities of all pixels. Finally, the original class probabilities are optimized in the post-processing step by the extended random walker that can express the spatial relationship among adjacent pixels quantitatively. The classification map is obtained according to the maximum probability.To assess the performance of the proposed method, a series of experiments is conducted on three small-scale HSI datasets, including Indian Pines, University of Pavia, and Salinas, as well as a large-scale HSI dataset HoustonU. The proposed SKERW_SVM obtains overall accuracies of 98.58%, 96.88%, 98.54%, and 91.01% on Indian Pines, University of Pavia, Salinas, and HoustonU, respectively. Experimental results demonstrate that our SKERW_SVM can fully mine the joint spatial-spectral features of HSI and achieve higher classification accuracy under the case of small labeled training samples compared with several related advanced methods. Moreover, the operation time consumed by SKERW_SVM is more appropriate than that by other methods.Under the lack of the labeled HSI pixel condition, the proposed HSI classification method by combining the SDR with post-processing optimization can efficiently extract the high-discrimination mixture feature information of HSI and significantly enhance classification performance. The SDR based on the homogeneous local regions, one of the components of the SKERW_SVM classification model, can greatly reduce the data redundancy and fully extract the information of spatial and spectral signatures compared with pixel-wise dimension-reduction methods. Meanwhile, the extended random walker in the post-processing step can fully use the spatial information of HSI by constructing a relationship graph to optimize the original class probabilities, thereby further improving the classification performance.关键词：remote sensing;hyperspectral image classification;superpixel dimension reduction;mixture feature extraction;post-processing optimization;support vector machine- 1017
- |
- 165
- |
- 0
发布时间：2024-03-22 -
摘要：As an important technique in image interpretation, hyperspectral image (HSI) classification is extensively used in many fields, such as remote-sensing observation and intelligent medical service. HSI classification may comprise label prediction based on feature extraction and based on classifiers. Although deep learning can directly obtain the classification results by one step, which is achieved by the end-to-end network structure from data input to classification result output, they are actually viewed as a direct cascade of both feature extraction based on deep networks (such as deep autoencoder and convolutional neural network) and classifiers (such as softmax and logistic regression). Most current classification approaches do not consider the influence of classifiers on feature extraction, which may cause the incompatibility between the extracted features and the used classifier. This incompatibility is reflected in the poor matching relationship between the classifier model and its input feature data, leading to poor prediction results.Method To remedy such deficiency, this paper presents a novel kind of HSI feature-extraction methods embedded by the classifier mechanism, which can ensure the compatibility between feature extraction and the used classifier. Thus, the features can be more easily calculated by classifier accurately, and classification prediction results can be improved. Two specific forms are given in this paper. 1) The sparse representation (SR)feature-extraction model compatible with Support Vector Machine (SVM) classifier is built, which embeds the SVM property into the SR. 2) The deep autoencoder (DAE) feature-extraction model compatible with softmax classifier is constructed, which integrates the softmax function into DAE network. We also provide the optimization strategy to obtain the optimal solutions of the SR and DAE models.Results The proposed SR and DAE models are experimentally evaluated on the remote-sensing HSI data and medical HSI data. The experiments consist of parameter analysis, algorithm comparison, ablation study, and convergence analysis. According to the parameter analysis, we validate that the values of important parameters have obvious impact on the performance of our methods and successfully select the best values of these parameters. As suggested by the algorithm comparison, the proposed methods achieve better classification performance than some state-of-the-art approaches, which have obvious effectiveness and superiority. The overall accuracy, average accuracy, and Kappa indices in the HSI classification task are, on average, higher by 5.03%, 5.13%, and 7.30%, respectively. An ablation study is conducted to demonstrate the effectiveness of the compatibility between feature extraction and the bedded classifiers for the performance improvement of HSI classification. Convergence analysis indicates that the designed optimization-solution strategy can meet the application requirements of reliability and rapidity.Conclusion The proposed SR and DAE methods realize good compatibility between feature extraction and classifiers. Accordingly, the extracted features can be better calculated by classifiers, and more competitive classification performance can be achieved.关键词：hyperspectral image classification;feature extraction;classifier mechanism;sparse representation;deep autoencoder network- 248
- |
- 328
- |
- 0
发布时间：2024-03-22
Remote Sensing Intelligent Interpretation
- 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