陆探一号干涉SAR系统应用
Theme Keywords： synthetic aperture radar interferometrySARremote sensingphase filteringlandslide susceptibility assessmentlandslide interpretationheterogeneous ensemble learningfull polarizationdynamic screeningdeep learning
- The Paper
- Abstract：Urban building change detection is an important part of land use, resource management, and urban planning. It plays an extremely important role in urban development, housing security and post-disaster reconstruction in earthquake areas. Optical image and SAR image are two commonly used data sources in urban building change detection. Optical images have the advantages of high resolution, high timeliness and convenient visual interpretation in change detection. However, optical images are difficult to use because of the influence of cloud, fog, and rain weather in southern China. SAR satellite is a side-looking imaging, and its imaging signal contains rich surface texture information, which is convenient for change detection and analysis. In SAR images, the change of ground objects can cause the change of amplitude information and coherence characteristics of SAR images, and buildings have certain imaging characteristics. China’s self-developed LT-1A/1B satellite has the advantages of high resolution, short return period, multipolarization, and all-day, all-weather operation. It can be used as an important supplement to optical images and has strong application potential in change detection. At present, the mining and utilization of SAR image timing information is not enough, and there are still some problems in the fusion of amplitude feature and coherence information. Therefore, it is of great significance to integrate the amplitude and coherence characteristics of multi-temporal SAR images to realize the change detection of urban buildings for the development of urban planning statistics. On the basis of LT-1 data, this study proposes an urban building change detection algorithm that combines multitemporal SAR amplitude and coherence information. The algorithm uses building recognition, color model conversion, and coherence change constraints to correlate multitemporal image pixel value changes to building changes, thereby obtaining urban building change areas and locating building change periods. Through the conversion of color model, the change characteristics of changing buildings in H component are distinguished, and the color gamut segmentation is classified and discussed to further determine the correspondence between building changes and H component sequences. The algorithm also analyzes the comparative changes of coherence before and after building changes, and obtains the corresponding relationship between building changes and coherence changes. With the amplitude and coherence information, the algorithm can locate the time period of change accurately and the misjudgment area is better eliminated. The local area of Hengqin Town, Zhuhai City, is employed as an example to conduct an experiment, and the results of building change detection in the study area are obtained using five scenes of LT-1 ascending images from June 23, 2023, to November 14, 2023. In addition, combined with the development of Hengqin, the results of building change detection in the northeast and central parts of Hengqin Town are verified against the construction periods of buildings. The findings show that the changes in 19 buildings in eight typical areas are consistent with the construction period of buildings, verifying the reliability of the algorithm. This work reveals the application ability and value of LT-1 data in urban building change detection. These data have a wide application prospect in urban planning and illegal construction investigation.Keywords：SAR;building change detection;Amplitude;Color model conversion;coherence893|2119|0
<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2025-12-30 - Abstract：The LuTan-1 (LT-1) 01A/B satellite constellation, China’s first civilian L-band distributed Synthetic Aperture Radar (SAR) satellite system, is expected to further promote the widespread application of domestic radar data products. While LT-1’s dual-polarization mode supports high-resolution and wide-swath imaging, it lacks certain polarization information compared to full-polarization observation systems, which limits its applications such as land cover classification. To address this, this study designs a multi-head multi-branch Convolutional Neural Network (CNN) and a tailored loss function for two commonly used dual-polarization modes of LT-1. The goal is to restore missing polarization information from dual-polarization inputs, thereby achieving full-polarization SAR data reconstruction. Considering the complementary nature of polarization and spatial features, as well as the distinct characteristics of amplitude and phase data, the network employs separate branches to extract pixel-level polarization information and local spatial features within receptive fields. Two independent heads predict amplitude and phase values to mitigate data crosstalk during output. Additionally, an amplitude/phase combined loss function incorporating phase periodicity is proposed to guide network training. Experiments conducted on two full-polarization SAR images acquired by LT-1 over urban areas in the UK demonstrate that, the proposed method achieves amplitude prediction errors of 1—2 dB and phase errors below 0.5 rad. These results surpass the reconstruction accuracy of a U-Net model with comparable parameter scale. The reconstructed data exhibit statistical distributions and polarization decomposition effects nearly identical to the ground-truth data, confirming effective preservation of full-polarization information. Furthermore, land cover classification experiments show that, classification maps using the reconstructed data achieve overall accuracy comparable to those using real full-polarization data, with a 5% improvement over dual-polarization based classification maps. This study preliminarily validates the feasibility of leveraging LT-1’s dual-polarization data to achieve high-resolution, wide-swath imaging while retaining full-polarization capabilities, providing richer input information for SAR land cover/use classification applications.Keywords：SAR;LT-1;Dual Polarization;full polarization;CNN;Terrain Classification1120|1205|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2025-12-30
- Abstract：Deep learning, with its powerful feature learning and nonlinear modeling capabilities, has been widely applied in the field of phase filtering for Interferometric Synthetic Aperture Radar (InSAR). However, in high-noise and stripe-dense regions, existing methods still struggle to balance noise suppression with the preservation of detailed phase information. To address this, this paper proposes a multi-scale InSAR phase filtering method integrating an Adaptive Augmentation and dynamic Screening-based Technique Module (AASTM). This model constructs a multi-scale feature extraction and layer-by-layer fusion framework based on the U-Net architecture. The AASTM module is inserted at different scales to perform adaptive augmentation and dynamic screening of interferometric phase features, achieving a balance between phase detail preservation and noise suppression. Additionally, a rhombus-square grid method is employed to generate simulated training datasets covering high-noise and dense-stripe scenarios, further enhancing the network model’s robustness and generalization capability in complex environments. The filtering performance of the proposed method is experimentally validated using simulated data and LT-1 A/B dual-satellite SAR data, with comparisons against existing filtering techniques. Results demonstrate: On simulated data, the proposed method achieves an average reduction of approximately 20% in Root Mean Square Error (RMSE) compared to other filtering methods, an improvement of about 18% in Structural Similarity Index (SSI), and an increase of approximately 5% in peak signal-to-noise ratio (SNR). Particularly in high-noise and dense stripe regions, the proposed method exhibits superior phase edge and detail retention capabilities. On LT-1 A/B field data, the proposed method achieves a residual error removal rate of 90.42% while better preserving phase detail information. In summary, the proposed method demonstrates significantly superior filtering accuracy compared to other methods, along with enhanced phase resolution and detail retention in dense stripe regions. It provides more reliable technical support for the precise inversion of LT-1 A/B digital elevation models.Keywords：deep learning;synthetic aperture radar interferometry;phase filtering;adaptive enhancement;dynamic screening854|996|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2025-12-30
- Abstract：Landslide susceptibility assessment based on machine learning often faces challenges because of the precision of spatial geoinformation sample datasets and the model’s ability to fit disaster-causing mechanisms, resulting in misjudgment and omission of high-risk areas. In the aftermath of the Luding 9.5 earthquake, secondary landslides have occurred frequently, and highly developed vegetation limits the accuracy of remote sensing-based landslide cataloging, severely affecting the accurate assessment of post-disaster landslide susceptibility. Therefore, this study integrates LT-1 ascending/descending track time-series InSAR surface deformation monitoring results with existing historical landslide catalog data to enhance the timeliness and accuracy of landslide spatial distribution base data. Gradient boosting decision tree (GBDT) and extreme gradient boosting (XGBoost) are selected as base learners to construct a stacking (GBDT-XGBoost) heterogeneous ensemble learning model for landslide susceptibility assessment. Through a comparative analysis of prediction accuracy, the model and algorithm optimization are completed, ultimately achieving precise, reliable, timely landslide susceptibility assessment and mapping. The experiment utilizes newly acquired ascending/descending track LT-1 satellite time-series SAR imagery datasets from 2023 to 2024. By extracting the surface deformation rate field by using stacking InSAR technology and conducting comprehensive landslide interpretation with Gaofen-2 satellite imagery, 36 new landslides are identified, expanding the historical landslide catalog dataset. On this basis, landslide susceptibility prediction is performed using three existing machine learning models. Comparative analysis of the accuracy and performance of the proposed heterogeneous ensemble learning model shows that the prediction performance and accuracy of each model improve because of the support of the LT dataset detecting landslide information. The stacking (GBDT-XGBoost) model has higher predictive performance and accuracy (area under the curve = 0.981, accuracy = 93.13%, recall = 92.82%, and F1-score = 0.932) compared with existing machine learning models. Moreover, the identified spatial distribution of landslides aligns closely with the high-risk areas predicted by the heterogeneous ensemble learning model. The proposed landslide cataloging method and landslide susceptibility assessment model help improve the accuracy and timeliness of landslide risk assessment and can provide a reference for disaster prevention, mitigation, and scientific reconstruction planning in relevant areas.Keywords：remote sensing;Lutan-1;Luding earthquake;landslide interpretation;heterogeneous ensemble learning;landslide susceptibility assessment1550|773|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2025-12-30
