This article offers a comprehensive AI-centric review of deep learning in exploring landslides with remote-sensing techniques, breaking new ground beyond traditional methodologies. We categorize deep learning tasks into five key frameworks—classification, detection, segmentation, sequence, and the hybrid framework—and analyze their specific applications in landslide-related tasks. Following the presented frameworks, we review state-or-art studies and provide clear insights into the powerful capability of deep learning models for landslide detection, mapping, susceptibility mapping, and displacement prediction. We then discuss current challenges and future research directions, emphasizing areas like model generalizability and advanced network architectures. Aimed at serving both newcomers and experts on remote sensing and engineering geology, this review highlights the potential of deep learning in advancing landslide risk management and preservation.

Decorrelation is one of the main limitations for synthetic aperture radar interferometry. Masking decorrelated pixels is crucial for retrieving information from SAR interferograms. However, for traditional masking methods, manually drawing masks is time-consuming and may be unfeasible when decorrelation areas are with complicated and blurred boundaries. Setting a single coherence threshold is also difficult, if not impossible, to mask out all decorrelated pixels without losing valid phases. Here, we propose a deep-learning segmentation network (Mask Net) based on Selective Kernel Res-Attention UNet, for generating decorrelation masks with applications to TanDEM-X interferograms. We conduct several experiments to determine the training strategy and parameters, including sample size, batch size, loss function, and downsampling scheme, to optimize network performance. Afterwards, we compare the performance of Mask Net with other classical segmentation networks. Our evaluation metrics show that Mask Net outperforms the best performance of other segmentation networks by IoU of 6.32% and F1 Score of 3.97%, respectively. It also possesses the fastest inferring speed, 0.4505 s on sample size of 1024-by-1024 pixels, which is at least ∼50% faster than other segmentation networks. We applied Mask Net to three TanDEM-X interferograms of Ki $\bar{\imath}$ lauea crater in Hawaii, metropolitan region of Wuhan, and Muztagata Glacier in China. Our results show that comparing with coherence threshold method, Mask Net can clearly mask out all decorrelation regions, rarely causing loss of valid phases. It also exhibits better segmentation performance than other deep-learning segmentation networks, especially for those complex decorrelation boundaries, with less computational time.

Interferogram denoising is critical in interferometric synthetic aperture radar (InSAR) data processing, directly impacting the accuracy and reliability of final products. Current data-driven denoising models rely on training samples with noises simulated from certain statistic models. However, such simulations may not fully capture the complexity of real-world data, resulting in degraded performance in practical applications. Here, we introduce a learning method for InSAR phase denoising, marking the first instance of training a model with real-world noise. Our method consists of two phases: (1) model excitation, where the model is initially trained with simulated noise to develop basic denoising capabilities; (2) refinement boosting, an unsupervised, iterative process where the model gradually refines itself using real-world noise through noise extraction, noise purification, data augmentation, and model enhancement. We identified the cost-optimal denoising model by conducting experiments across network architectures of varying complexity, using identical training datasets and experimental settings. Experiments on synthetic interferograms with varying coherence levels show that our method achieves the lowest denoising errors, outperforming state-of-the-art methods such as NL-InSAR and InSAR-BM3D, while maintaining fast inference times comparable to the BoxCar method. Applied to real Sentinel-1 interferograms, our approach demonstrates superior denoising performance, as evidenced by the fewest residues and the smoothest unwrapping results. By analyzing noises simulated from the coherence-guided statistical model and extracted from Sentinel-1 interferograms, we find significant discrepancies between simulated and real-world noise, underscoring the importance of preparing training samples with real-world noise for InSAR data-driven models, e.g., denoising, unwrapping, and other applications.

This paper investigates the joint authentication of topology and geometry information of 2D engineering computer-aided design graphics, which focus more on topological modeling than geometric modeling of objects. A robust hashing scheme is proposed for joint topology and geometry authentication. The covariance matrices of descriptors are explored to fuse and encode both topology and geometry features of different types into a compact representation. First, a normalized binary shape texture is rendered for each geometric object through the render-to-texture technique. Then, for each geometric object, geometry features are computed based on statistical features that are extracted from image rings. Additionally, topology features are generated according to the topological relations among joint objects. To generate hash codes of the graphic, all geometric objects are first grouped according to their geometry features. Then, for each group, the covariance matrices of descriptors are applied to fuse both the topology and geometry features of all objects, and the intermediate hash codes of each group are computed based on the covariance matrices. The final hash sequence is formed by concatenating the intermediate hash codes that correspond to each group. Secret keys are introduced into both feature extraction and hash construction. The hashes are robust against topology-preserving graphic manipulations and sensitive to malicious attacks. By decomposing the hashes, the locations of tampered objects can be determined. Experimental results are presented to evaluate the performance and show the effectiveness of the method.