Reliable localization under Global Positioning System-denied or visually degraded conditions remains a fundamental challenge for autonomous systems. Vision- and Light Detection and Ranging (LiDAR)-based approaches often degrade in low illumination, adverse weather, or appearance-changing environments, as they rely on stable surface texture or geometry. In contrast, ground-penetrating radar (GPR) captures subsurface electromagnetic reflections that remain relatively stable across lighting, seasonal, and weather variations, making it a promising complementary sensing modality for long-term localization. However, spatial variability in subsurface dielectric properties induces fluctuations in electromagnetic wave velocity, leading to geometric distortions in GPR echoes and unstable feature extraction. To address this challenge, we propose the Velocity-Invariant Feature Transform (VIFT), a physics-guided self-supervised learning framework for GPR keypoint detection and description. VIFT explicitly models wave-velocity-induced distortions through a continuous velocity space parameterized by a Beta distribution, and leverages velocity-conditioned wavefield migration as physically consistent data augmentation. A Siamese network is trained with velocity-consistency supervision to jointly learn repeatable keypoint score maps and discriminative local descriptors from unlabeled real GPR scans. To further enhance robustness, sparsity-aware, dispersion, distinctiveness, and orthogonality losses are incorporated to improve repeatability, spatial coverage, and descriptor discriminability. Extensive experiments on public benchmarks and large-scale real-world GPR datasets demonstrate that VIFT consistently outperforms traditional handcrafted methods and recent learning-based Vison and GPR methods, achieving a 5–10% improvement in keypoint repeatability over state-of-the-art methods, particularly under extremely sparse keypoint sampling regimes, while also improving matching accuracy and registration robustness under diverse subsurface conditions.

Ground penetrating radar (GPR) based localization has gained significant recognition in robotics due to its ability to detect stable subsurface features, offering advantages in environments where traditional sensors like cameras and LiDAR may struggle. However, existing methods are primarily focused on small-scale place recognition (PR), leaving the challenges of PR in large-scale maps unaddressed. These challenges include the inherent sparsity of underground features and the variability in underground dielectric constants, which complicate robust localization. In this work, we investigate the geometric relationship between GPR echo sequences and underground scenes, leveraging the robustness of directional features to inform our network design. We introduce learnable Gabor filters for the precise extraction of directional responses, coupled with a direction-aware attention mechanism for effective geometric encoding. To further enhance performance, we incorporate a shift-invariant unit and a multi-scale aggregation strategy to better accommodate variations in dielectric constants. Experiments conducted on public datasets demonstrate that our proposed EDENet not only surpasses existing solutions in terms of PR performance but also offers advantages in model size and computational efficiency.

As a promising technology for autonomous driving, localizing ground penetrating radar (LGPR) is a vehicle localization method that relies on prior maps and couples deeply with subsurface features. However, the unique characteristics of GPR data often lead to a significant number of mismatched candidates during localization. Previous learning-based GPR place recognition methods have primarily relied on 2D convolutional neural networks (CNNs), which struggle to effectively capture critical temporal information, limiting further performance improvements. To address this limitation, we propose a spatial-temporal U-shaped network (STU-Net) that leverages 3D convolutional neural networks to simultaneously extract spatial and temporal features from GPR image sequences. Additionally, residual dense blocks (RDBs) are integrated into the network to enable multi-scale feature extraction. Extensive experiments conducted on publicly available datasets demonstrate that our STU-Net achieves state-of-the-art performance, outperforming existing methods with significant improvements.

Ground-penetrating radar (GPR) , as a sensor for mapping and localizing subsurface features, has gained significant attention in robotic localization for complex environments. However, varying weather conditions change the subsurface dielectric constant, which weakens the registration correlation between real-time images and maps, thereby compromising localization stability. We design an image registration framework for localizing ground penetrating radar (LGPR) in multi-weather. Specifically, we first propose an adaptive hyperbolic receptive field that aims to significantly enhance the robust features in the GPR images, while improving both discrimination capability in map and reliability for image registration. Then, an image alignment module is introduced to eliminate the time-delay blurring problem caused by the variation of dielectric constant under multi-weather conditions. The proposed method was evaluated on three distinct datasets (simulated, publicly available, and self-constructed), demonstrating significant improvements in registration and localization performance. The average correlation coefficient on simulated data achieved a enhancement from 0.3796 to 0.7406 compared with baseline, validating the effectiveness of feature enhancement. Furthermore, measured datasets exhibited 15% higher average registration and localization accuracy than baseline. These results demonstrate that the method provides a reliable guarantee for the stable operation of LGPR systems in complex environments. The datasets will be released at: https://github.com/Eazin-bz/dataset_AHFT.git.

Localizing ground penetrating radar (LGPR) has been proven to be a promising technology for robot localization in various dynamic environments. However, the extreme scarcity of underground features introduces false candidate matches and brings unique challenges to this task. In this paper, we propose a sequence-based framework for LGPR to address the aforementioned issues. Specifically, we first introduce a trainable strategy to extract robust underground features in multi-weather conditions. By further using sequential information, our LGPR system can observe richer underground scene contexts, and the associated multi-frame scans could also improve the performance of underground place recognition. We demonstrate the superiority of our proposed method by comparing it against several recent state-of-the-art baseline methods applied to GPR image tasks. Experimental results on large public and self-collected datasets show that our proposed framework significantly improves the performance of various baselines in different scenarios.

For airborne repeat-pass synthetic aperture radar interferometry (InSAR), residual motion errors (RMEs) exist between the true and measured trajectories due to the inaccuracy of current navigation systems. In addition, a global unknown phase ambiguity (PA) exists in the unwrapped phase due to the error of absolute phase estimation. The RME and PA are significant error sources in the repeat-pass InSAR, which can cause phase errors in the final interferograms. In general, an external digital elevation model (DEM) can be used to estimate the RME and PA. However, it is hard to find a high-precision external DEM matching with the airborne InSAR data. Compared with the traditional InSAR technique, the 360° aperture gives circular InSAR (In-CSAR) the capability to estimate the RME and PA without external DEMs. Unlike previous approaches, in this article, a multiangle observation model is established to estimate RME and PA simultaneously. This model registers the DEM images obtained from different observation angles. In the field of image registration, the Demons algorithm is often used for nonrigid registration. We use, for the first time, the Demons algorithm to register multiangle false DEM images containing height errors and obtain the reference DEMs. Due to the gradient of the DEM image being destroyed by the RME and PA, a two-step preprocessing is proposed to achieve subpixel registration of multiangle false DEM images. After obtaining the reference DEM, the least square (LS) or robust mixed integer linear programming (RMILP) can estimate the unknown RME and PA. The results show that, with respect to the true values, the root-mean-square error (RMSE) of the calibrated DEMs is 0.37 m for the simulated dataset and 1.17 m for the real dataset. After correcting the RME-induced and PA-induced height errors, the RMSE of the DEM is improved by at least 92.6% for the simulated dataset and by at least 70.8% for the real dataset.

Verifying the neighborhood consensus to remove false correspondence is a popular idea in feature matching. However, traditional neighborhood consensus only considers spatial neighborhoods, which is not robust in challenging remote sensing tasks. This paper extends the traditional neighborhood consensus for improving robustness to the two key issues — significant geometric transformation and repetitive patterns. First, we introduce a novel matching neighborhood that extends the one-to-one correspondence in traditional neighborhood consensus to one-to-multiple structure to address the repetitive patterns, where one-to-multiple means that multiple matching candidates are preserved in calculating descriptor similarity. Second, the traditional spatial neighborhood is also extended using affine correspondence, which can adaptively address the significant geometric transformations without multi-scale processing. On the two bases, we construct a novelextended neighborhoodby combining theextended spatial neighborhoodwith thematching neighborhood. And consequently, the false feature correspondences are filtered by measuring the consensus between the extended neighborhoods. Numerous experiments demonstrate that the proposed method is state-of-the-art in comparison with recent learning and traditional methods, especially for the UAV localization task. We also show that the proposed method is robust to the basic settings, such as the the pre-filtering threshold and the type of local features.

Ground penetrating radar (GPR) images typically depict underground targets as hyperbolas, which pose a challenging detection task due to their low amplitude and resolution. To address this, we propose a robust and efficient feature descriptor based on a modified phase symmetry (PS) model. Specifically, we enhance the PS model to better represent hyperbolas in GPR images and introduce a weighted phase symmetry histogram descriptor (WPSHD) as a local structure descriptor. The proposed descriptor is used as the feature input to the classifier to realize the hyperbola recognition. The proposed method is compared with two baselines and state-of-the-art (SOTA) methods, such as histogram of oriented gradient (HOG), edge histogram descriptor (EHD), and histogram of oriented vector phase symmetry (HOVPS). Our validation experiments on both public datasets and real-world data show that our proposed algorithm improves hyperbola detection in GPR images, as demonstrated by qualitative and quantitative analyses.

Feature matching refers to the establishment of reliable correspondence between two sets of local features, which is an essential approach in remote sensing applications such as image registration and mosaicking. In this paper, a simple yet effective method, called frame-based locality preservation matching, is proposed for robust remote sensing image matching. We primarily focus on those images pairs that involve large-scale geometric transformations (e.g., extreme zoom). The key idea of our approach is to dig up the frame knowledge, such as the feature orientation and scale implied by common features like SIFT. The frame knowledge is free to obtain, and we find it to be of great significance in feature matching, especially for our focus -- large-scale geometric transformations. The proposed method can easily handle the geometric challenges and high outlier proportions, and significantly improves the performance compared to other state-of-the-art methods.

Feature matching is a fundamental technique in remote sensing image processing. This article proposes a new formulation of affine covariant feature matching for remote sensing images, where we suggest matching features by matching two sets of triplets. Compared with previous works, the formulation exploits the whole feature frame rather than the 2-D location to reject outliers. Besides, we also develop a new latent variable model to combine the feature frame and the SIFT ratio values, to enhance the convergence speed and success rate in challenging cases. We evaluate our model on three challenging datasets in terms of both qualitative and quantitative experiments. We also study the robustness to outliers since remote sensing images are typically affected by mismatches. The results demonstrate that the proposed method provides excellent matching performance with satisfying runtime and shows good robustness to outliers.

Hyperbola detection is an important application field of ground-penetrating radar (GPR) systems as underground threats and targets detected by these systems are presented in the form of a hyperbola. However, the low-amplitude hyperbola detection of deeply-buried targets and targets with low metal content has remained a major challenge due to the increase in the attenuation of radar echo with an increase in the detection depth and the features with low resolution extracted by existing methods. In this study, first, a high-level visual feature, namely, phase symmetry, is proposed to effectively improve the feature resolution and the robustness of amplitude change in GPR images. Then, we propose a handcrafted feature descriptor based on phase symmetry, namely, histogram of oriented vector phase symmetry (HOVPS) to improve the detection of hyperbola in GPR images. In constructing HOVPS, we first cite our previous work to enhance the descriptive ability of hyperbola in GPR images. Subsequently, we extend a phase symmetry model to develop a vector feature model [namely, vector phase symmetry (VPS)]. Finally, HOVPS is developed based on the VPS to extract structural information from GPR images. The proposed HOVPS describes the shape features in GPR images using a symmetrical structure, and it is used as the feature input of the classifier for hyperbola detection. The qualitative analysis of the proposed method is performed by comparing the performance of the proposed method for the extraction of features on different GPR data with those of other methods. In addition, we also provide quantitative analysis on different signal-to-noise ratio (SNR) test sets, and the results reveal that the proposed method outperforms various state-of-the-art methods (e.g., GPR histogram of oriented gradient (gprHOG), edge histogram descriptor (EHD), and log-Gabor (LG) feature).