This paper is concerned with the matching stability problem across different decoder layers in DEtection TRansformers (DETR). We point out that the unstable matching in DETR is caused by a multi-optimization path problem, which is highlighted by the one-to-one matching design in DETR. To address this problem, we show that the most important design is to use and only use positional metrics (like IOU) to supervise classification scores of positive examples. Under the principle, we propose two simple yet effective modifications by integrating positional metrics to DETR’s classification loss and matching cost, named position-supervised loss and position-modulated cost. We verify our methods on several DETR variants. Our methods show consistent improvements over baselines. By integrating our methods with DINO, we achieve 50.4 and 51.5 AP on the COCO detection benchmark using ResNet-50 backbones under 1× (12 epochs) and 2× (24 epochs) training settings, achieving a new record under the same setting. We achieve 63.8 AP on COCO detection test-dev with a Swin-Large backbone. Our code will be made available at https://github.com/IDEA-Research/Stable-DINO.

Semantic segmentation of remote sensing images plays a critical role in areas such as urban change detection, environmental protection, and geohazard identification. Convolutional Neural Networks (CNNs) have been excessively employed for semantic segmentation over the past few years; however, a limitation of the CNN is that there exists a challenge in extracting the global context of remote sensing images, which is vital for semantic segmentation, due to the locality of the convolution operation. It is informed that the recently developed Transformer is equipped with powerful global modeling capabilities. A network called TCNet is proposed in this article, and a parallel-in-branch architecture of the Transformer and the CNN is adopted in the TCNet. As such, the TCNet takes advantage of both Transformer and CNN, and both global context and low-level spatial details could be captured in a much shallower manner. In addition, a novel fusion technique called Interactive Self-attention is advanced to fuse the multilevel features extracted from both branches. To bridge the semantic gap between regions, a skip connection module called Windowed Self-attention Gating is further developed and added to the progressive upsampling network. Experiments on three public datasets (i.e., Bijie Landslide Dataset, WHU Building Dataset, and Massachusetts Buildings Dataset) depict that TCNet yields superior performance over state-of-the-art models. The IoU values obtained by TCNet for these three datasets are 75.34% (ranked first among 10 models compared), 91.16% (ranked first among 13 models compared), and 76.21% (ranked first among 13 models compared), respectively.

The technique of interferometric synthetic aperture radar (InSAR) is increasingly employed for landslide detection over large areas, even though the limitations of initial InSAR analysis results have been well acknowledged. Steep terrain in mountainous areas may cause geometric distortions of SAR images, which could affect the accuracy of InSAR analysis results. In addition, due to the existence of massive ground deformation points in the initial InSAR analysis results, accurate landslide recognition from the initial results is challenging. To efficiently identify potential landslide areas from the ascending–descending SAR datasets, this paper presents a novel interpretation approach to analyze the initial time-series InSAR analysis results. Within the context of the proposed approach, SAR visibility analysis, conversion analysis of deformation rates obtained from the time-series InSAR analysis, and spatial analysis and statistics tools for cluster extraction are incorporated. The effectiveness of the proposed approach is illustrated through a case study of landslide identification in Danba, a county in Sichuan, China. The potential landslide regions in the study area are identified based on the interpretation of small baseline subset InSAR (SBAS-InSAR) results, obtained with ascending–descending Sentinel-1A datasets. Finally, on the basis of the field survey results, a total of 21 landslides are detected in the potential landslide regions identified, through which the results obtained from the proposed interpretation approach are tested.

The interferometric synthetic aperture radar (InSAR) technique is widely adopted for detecting and monitoring landslides, but its effectiveness is often degraded in mountainous terrains, due to geometric distortions in the synthetic aperture radar (SAR) image input. To evaluate the terrain effect on the applicability of InSAR in landslide monitoring, a variety of visibility evaluation models have been developed, among which the R-index models are quite popular. In consideration of the poor performance of the existing R-index models in the passive layover region, this study presents an improved R-index model, in which a coefficient for improving the visibility evaluation in the far passive layover regions is incorporated. To demonstrate the applicability of the improved R-index model, the terrain visibility of SAR images in Fengjie, a county in the Three Gorges Reservoirs region, China, is studied. The effectiveness of the improved R-index model is demonstrated through comparing the visibility evaluation results with those obtained from the existing R-index models and P-NG method. Further, the effects of the line-of-sight (LOS) parameters of SAR images and the resolution of the digital elevation model (DEM) on the terrain visibility are discussed.

The ground deformation rate is an important index for evaluating the stability and degradation of permafrost. Due to limited accessibility, in-situ measurement of the ground deformation of permafrost areas on the Tibetan Plateau is a challenge. Thus, the technique of time-series interferometric synthetic aperture radar (InSAR) is often adopted for measuring the ground deformation rate of the permafrost area, the effectiveness of which is, however, degraded in areas with geometric distortions in synthetic aperture radar (SAR) images. In this study, a method that integrates InSAR and the random forest method is proposed for an improved permafrost stability mapping on the Tibetan Plateau; to demonstrate the application of the proposed method, the permafrost stability mapping in a small area located in the central region of the Tibetan Plateau is studied. First, the ground deformation rate in the concerned area is studied with InSAR, in which 67 Sentinel-1 scenes taken in the period from 2014 to 2020 are collected and analyzed. Second, the relationship between the environmental factors (i.e., topography, land cover, land surface temperature, and distance to road) and the permafrost stability is mapped with the random forest method based on the high-quality data extracted from the initial InSAR analysis. Third, the permafrost stability in the whole study area is mapped with the trained random forest model, and the issue of data scarcity in areas where the terrain visibility of SAR images is poor or InSAR results are not available in permafrost stability mapping can be overcome. Comparative analyses demonstrate that the integration of the InSAR and the random forest method yields a more effective permafrost stability mapping compared with the sole application of InSAR analysis.