Regular inspections of sewer systems and pipelines are crucial for maintaining functionality. Since many structures are not directly accessible, a robotic platform equipped with a camera ring and laser scanner is employed. This paper presents a novel hybrid image stitching pipeline with a geometry-driven decision process that combines model-based and simplified geometric approaches to generate unwrapped pipe textures from homogeneous, featureless images. In the model-based method, each camera image is projected onto a 3D pipe surface reconstructed from the laser scan and then unwrapped. This enables geometrically consistent texturing for complex geometries with deformations, house connections or deposits, but is computationally expensive. The main contribution is a real-time capable hybrid pipeline that combines the model-based method with a simplified stitching approach and automatically selects between them based on the scanned geometry and a strict time budget on embedded hardware. If the laser scan confirms that the sensor head is approximately parallel to the pipe axis and the upcoming pipe segment is close to a perfect cylinder, the simplified stitching method is used. In this case, the overlap between camera images is computed directly from the measured distance between each camera and the pipe wall and the camera’s opening angle. Both stitching approaches use the same laser scanner. The appropriate method is selected automatically based on the scanned geometry. The simplified method reduces computation time by about 50
Extracting robust semantic representations in Siamese architectures is often accompanied by the degradation of fine-grained spatial details that are critical for boundary localization. In practical bi-temporal remote sensing imagery, pseudo changes caused by illumination variation, seasonal differences, shadows, viewpoint shifts, and registration errors may further interfere with reliable feature comparison. To alleviate these coupled challenges, we propose the Temporal Interaction and Detail-Aware Learning Network (TIDAL-Net), which reformulates bi-temporal representation learning as a progressive process of structural detail enhancement, temporal response purification, and residual multi-scale fusion. In the early feature learning stage, the Difference Detail Enhancement Mechanism (DDEM) models bi-temporal dynamics and strengthens high-frequency edge cues before they are weakened by deep semantic abstraction. Before cross-temporal interaction, the Temporospatial Interactive Attention Module (TIAM) recalibrates unreliable temporal responses, allowing feature interaction to focus more on change-related structural shifts. After these refined representations are obtained, the Multi-Scale Residual Fusion (MSRF) strategy integrates complementary detail-aware and semantic information while preserving residual structural cues. Extensive experiments demonstrate that TIDAL-Net achieves competitive performance on established optical benchmarks, with F1 scores of 92.62 https://github.com/Bluemirror3/TIDAL-Net.git .
Image-based 3D reconstruction is vital in many applications, such as digital twins, smart cities, machine vision, and autonomous driving. In recent years, it has undergone a paradigm shift, propelled by advancements in both conventional photogrammetry and deep learning. This review provides a comprehensive photogrammetric perspective on both conventional and learning-based techniques, a viewpoint that prioritizes geometric fidelity, robustness, handling of uncertainty, and suitability for real-world applications. We first systematically revisit the fundamentals of traditional pipelines: Structure from Motion (SfM), Multi-View Stereo (MVS), and surface reconstruction. The review then details recent progress in conventional methods, highlighting innovations in scalable and efficient SfM, specialized camera models for MVS, and robust surface reconstruction algorithms. Subsequently, we explore the transformative evolution brought by learning-based techniques, including deep SfM, learning-based MVS, differentiable rendering-based scene representation methods (NeRF, 3DGS), groundbreaking feed-forward 3D reconstruction models (e.g., DUSt3R, VGGT), and surface reconstruction including explicit and implicit methods. Emphasis is placed on evaluating whether learning-based approaches genuinely meet photogrammetric requirements such as metric accuracy and reliability, rather than optimizing solely for visual realism. Furthermore, we conclude by identifying key challenges and research frontiers including generalization across domains, scalability to high-resolution imagery, real-time performance, and uncertainty quantification. By bridging the gap between classical photogrammetry and data-driven 3D vision, this work aims to guide future research toward robust, accurate, and certifiable 3D reconstruction systems suitable for engineering, industrial, and geospatial applications.
Floods have caused severe damage in Austria in recent years, and climate change is expected to increase flood risks in the future. While Austria has an advanced hydrological measurement network for flood monitoring and prediction, Synthetic Aperture Radar (SAR) data from satellites can provide valuable additional information. This study presents a well-established Bayesian flood mapping approach that automatically retrieves flood extents using Sentinel‑1 SAR data. By combining VV and VH polarizations, the algorithm aims to improve sensitivity for flood mapping. However, Austria’s complex topography and land cover, as well as flood dynamics present significant challenges for SAR-based flood mapping. To assess the suitability of SAR-based flood mapping and specifically our algorithm for Austria, we introduce two novel evaluation methods: (1) assessing temporal coverage through hydrological measurements and (2) evaluating sensitivity using flood risk zones. Additionally, we validate the results using independent reference data from local helicopter surveys and high-resolution optical satellite imagery. Our findings show that the approach can map flood extents up to 60.64
Remote sensing change detection (RSCD) requires effective modeling of long-range spatial dependencies and large receptive fields to accurately capture subtle land-cover variations between bi-temporal remote sensing images. However, existing RSCD methods often suffer from limited receptive fields or high computational complexity, constraining their ability to capture large-scale contextual dependencies. To address these challenges, this paper proposes a Large Kernel State Space Model for Change Detection (LKMamba-CD). Within LKMamba-CD, the Large Kernel SS2D Module (LKS) is designed to extract large-scale and long-range feature correlations. The Bi-temporal Change Information Fusion Module (BCIF) is utilized to fuse bi-temporal change information. Meanwhile, the Large Receptive Field Feature Aggregation Module (LRFA) is utilized to extract large-scale land cover feature correlation. Within BCIF and LRFA, LKS serves as the core component, leveraging re-parameterized large kernel convolution and the Mamba architecture to capture long-range dependencies with expansive receptive fields efficiently. Experiments on S2Looking, WHU-CD, and GVLM-CD show that LKMamba achieves the highest F1 of 62.02 https://github.com/kion-86/lkmamba .
Downed coarse woody debris (CWD), defined here as fallen dead woody material in all stages of decay, is a significant carbon pool in forest ecosystems, accounting for up to 8
Due to its photorealistic rendering quality and precise geometric representation, Neural Radiance Fields (NeRF) has achieved promising progress in novel view synthesis. However, due to the multi-temporal nature of satellite imagery, existing NeRF methods for satellite scenes struggle to achieve high-quality rendering in scenarios with drastic lighting variations, shadow discrepancies, and transient object interference. To address these challenges, this paper proposes a novel NeRF-based method that incorporates physical modeling of scene illumination and filters transient objects via semantic segmentation. Specifically, our approach models scene lighting based on the bidirectional reflectance distribution function (BRDF), enabling more accurate color reconstruction. Additionally, we introduce a semantic segmentation-based method for transient object removal. By generating masks for transient objects, our model effectively distinguishes between dynamic and static components in the scene, thereby eliminating the interference of transient objects during training. Furthermore, we improve the density field in NeRF by incorporating an occupancy field and the secant method to derive more precise digital surface models (DSM). We conducted extensive experiments on the DFC2019 dataset, and the results demonstrate that our method can produce more photo-realistic results for the task of novel view systhesis and digital surface modeling.
Detecting kinematic transitions, often termed trend changes, in Interferometric Synthetic Aperture Radar (InSAR) derived displacement time series is a critical task for geophysical hazard monitoring. This process is frequently impeded by nonlinear deformation patterns, irregular sampling, and significant temporal data gaps. To address this challenge, we introduce an Attention-based Time-Gated Long Short-Term Memory (ATGLSTM) architecture that synergistically integrates temporal gating with a self-attention mechanism to enhance Change Point Detection (CPD), specifically under conditions of data sparsity. The proposed architecture was rigorously trained and evaluated using a combination of realistic synthetic time series, calibrated with European Ground Motion Service (EGMS) data, and real-world observations from the geodynamically complex and data-sparse region of Iceland. Benchmarked against a conventional Convolutional Neural Network (CNN) and a standard Time-Gated LSTM (TGLSTM), the ATGLSTM demonstrated superior performance, achieving an F1-score of 82.40
Digital elevation models (DEMs) are essential for many governmental, scientific, and commercial applications. Recently, with the release of several new products, the availability and quality of global 30 m DEMs (GDEMs) have improved considerably. Here, we present a validation of five original GDEMs (e.g., Copernicus GLO-30) and four bare-earth GDEMs, including FathomDEM and GEDTM, two recently developed products with enhanced artefact removal that are expected to outperform existing GDEMs. We evaluated the accuracy of elevation, terrain slope, surface ruggedness, and delineated stream network. Our results showed differences between the two categories of GDEMs, with bare-earth GDEMs being substantially more accurate than original GDEMs over Czechia. The highest accuracy was achieved by the FathomDEM, followed by GEDTM and FABDEM. The difference was most pronounced in forests, where the FathomDEM yielded an RMSE of 2.7 m, approximately 2–4 m lower than other bare-earth GDEMs and 7–11 m lower than the original GDEMs. We also observed a noticeable effect of land cover, terrain slope, and their interaction on the accuracy of elevation and derived terrain characteristics, with all metrics exhibiting the lowest accuracy in forested areas. Bare-earth GDEMs performed better in estimating slope, ruggedness, and particularly in delineating stream networks; however, original GDEMs can still be effectively used to derive terrain characteristics, and perform comparably well, except in forests on gentle to moderate slopes. Overall, our findings show that FathomDEM maintains the highest and most consistent accuracy across diverse conditions, constituting a major milestone in global terrain mapping.
Digital Elevation Models (DEMs) are essential for a wide range of geospatial applications, including urban planning, environmental monitoring, and disaster management. Conventional methods for DEM generation, such as LiDAR and radar, are accurate but costly and constrained by regulatory and logistical challenges. This study explores the use of deep learning techniques for DEM generation from single satellite images, focusing on the integration of advanced attention mechanisms and novel normalization strategies. A custom dataset, featuring Sentinel‑2 and Landsat 9 imagery paired with high-resolution DEMs of the Iberian Peninsula, was developed and made publicly available to support transparency and reproducibility. Two normalization approaches were evaluated: Global normalization, which preserves global elevation relationships, and Global normalization with Shift, which emphasizes local terrain features. Additionally, state-of-the-art architectures, including U‑Net, Pix2Pix, and DRPAN, were adapted with attention mechanisms such as the Global Attention Mechanism (GAM) and gradient-based loss terms. The results indicate that the proposed GAM, when combined with Shift normalization, produces Mean Absolute Error (MAE) values of 65 m when evaluated under the corresponding local terrain reconstruction objective. While the method does not match the accuracy of traditional sensors, it offers a promising, scalable, and cost-effective alternative for DEM generation in contexts with limited data or resources, where relative elevation accuracy is sufficient.
Synthetic aperture radar (SAR) images frequently suffer from speckle noise, which can greatly hinder subsequent processing and analysis tasks. As a result, speckle suppression is a challenge. Current deep learning-based approaches for speckle suppression often struggle to effectively combine local and global image features, frequently leading to the disruption of spatial continuity during the denoising process. To tackle these challenges, a multi-layer feature enhanced frequency correlation network (MFEFC-Net) was designed for speckle suppression. First, a multi-layer feature extraction block (MFEB) is developed to hierarchically extract deep representations from noise-affected images, enabling a more accurate characterization of the noise patterns. Second, a frequency correlator (FC) module is built to model the image over long distances and improve the gradient continuity of the denoised output. The FC module utilizes the discrete wavelet transform (DWT) to decompose the extracted deep features into high-frequency details and low-frequency structural components. It then leveraged a Transformer to establish frequency relationships by distinguishing diverse noise patterns across different sub-bands, ultimately removing noise while retaining more image details through the image feature enhancement (IFE) module. Finally, a feature decoder was employed to reconstruct the denoised image from the deep features, and a residual refine block (RRB) was utilized to restore more image details. Both quantitative evaluations and visual comparisons demonstrated that the proposed MFEFC-Net achieves outstanding performance compared to existing state-of-the-art denoising methods in both quantitative and qualitative assessments.
Optical remote sensing, particularly satellite-derived optical trapezoid model (OPTRAM), can be used as a proxy to monitor peatland water table (WT), a key determinant for peatland condition. So far, OPTRAM has been used only in temporal monitoring of WT at coarse spatial resolution while it has not been tested to detect spatial patterns of WT in spatially heterogeneous northern peatlands. To address the abovementioned gap, we downscale four differently parameterized Sentinel‑2 OPTRAMs with the help of optical, thermal, and topographic uncrewed aerial vehicle (UAV) variables and random forest modeling in two open peatlands in northern Finland covered by spatially extensive field measurements of WT (n = 95). We (1) assess how parameterization of OPTRAM affects OPTRAM-WT correlation, (2) test whether downscaled OPTRAM correlates stronger with WT than the original OPTRAM, and (3) compare OPTRAM to other remote sensing variables calculated from Sentinel‑2 and UAV data. Our results showed that OPTRAM parameterization strongly affects OPTRAM-WT correlation, with Spearman correlation (rs) ranging between 0.23–0.53. Random forest-based downscaling models had a relatively high explained variance (45.6–72.4
3D deep-learning based models have been proposed to assist in the digitalization process of as-is infrastructure. Despite their outstanding performance to process large amounts of data, research on interpreting their execution rather than improving their outcome has been scarce. To further characterize the execution of these models, in this study we seek to identify influential parts of the input that contribute the most to a given model decision in the context of 3D semantic segmentation. Considering the intrinsic nature of 3D data and the existing basic characterization techniques, we explore variations of the rules to backpropagate the model prediction signal back to the input. On the qualitative side, we have found that the techniques implemented point to learned patterns that vary depending on the model architecture. We show that the generated explanations serve also to identify spurious learned patterns that would otherwise be unnoticed if only analyzing the model predictions. Our results serve then as a baseline comparison for further research into 3D explanation techniques.
Water turbidity is an important indicator of water quality, as it allows for conclusions on the particles contained in the water. Suspended matter, sediments, algae and other particles in the water column influence the light transmission, oxygen content, nutrient supply and productivity of a water body. Information on water turbidity is therefore relevant for a wide range of limnological and oceanographic questions. In particular, the vertical turbidity stratification plays an important role in understanding the ecological relationships within a water body. Conventional in-situ measurement methods provide only isolated, point-based turbidity profiles, which fail to capture the overall turbidity conditions throughout the entire water body, including all small-scale lateral and vertical variations. Hence, the utilization of Light Detection and Ranging (LiDAR) bathymetry data for assessing area-wide depth-resolved water turbidity information shows great potential. By analysing the waveform of the backscattered LiDAR signal, information about the turbidity of the water can be derived. In this paper we present a signal analysis approach that allows to derive depth resolved turbidity parameter fields from LiDAR bathymetry data. The methodology was tested on a pre-alpine mountain lake using a UAV LiDAR bathymetry system. For validation purposes, a comprehensive set of in-situ water property measurements is available, which were collected at the same day as the bathymetric data. The results show that the turbidity parameters determined from the LiDAR data generally correspond to the trends observed in the conventional water property measurements, although local deviations are present.
One of the main challenges with optical remote sensing imagery is the lack of sufficient data and the difficulty in collecting large amounts of images. Since the remote sensing field depends on large datasets for accurate processing by machine learning models, it is essential to have a greater number of data samples to enable effective AI-based processing in this domain. To address this issue, we propose generative models that can synthesize realistic texture images from limited data. This paper presents a novel approach to enhance texture image synthesis across high-, medium-, and low-resolution remotely sensed images, which encompasses both regular and irregular texture types. The main idea is to integrate texture descriptors into the attention mechanisms of generative models, such as the Texture-based ViTGAN model (Tex-ViTGAN) and the Texture-based Diffusion model (Tex-Diffusion), to improve the extraction of texture-specific features while preserving spatial arrangement, structural patterns, and fine details. The Tex-ViTGAN model has limitations in preserving structural patterns in regular textures, so the Tex-Diffusion model overcomes this problem and significantly improves texture representation for regular and irregular textures at all resolution levels of remotely sensed images. To validated the performance of our proposed methods, we used standard evaluation metrics to assess the quality of generated images by calculating the Fréchet Inception Distance (FID), Inception Score (IS), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) between generated and real image samples. Our Tex-Diffusion model achieved FID of 5.1, IS of 7.94, SSIM of 0.90, and LPIPS of 0.177 on an irregular remotely sensed sample, demonstrating comparable results to the state-of-the-art models.
Topo-bathymetric LiDAR has become a standard 3D geodata acquisition technique in geosciences with a wide range of applications in terrestrial and aquatic landscapes. In particular, the usage of short laser pulses with medium-sized footprints for bathymetric applications has seen growing attention over the past decade. These LiDAR configurations have shown higher relative changes in echo pulse width in relation to angle of incidence compared to standard topographic LiDAR using more collimated near-infrared lasers. Although angle-dependent changes in amplitude have been well documented, quantification of such a relation for the echo pulse width is still not entirely solved. By focusing on LiDAR with short and broad pulses, we can therefore use the higher relative changes of such systems together with numerical simulations to quantify the relationship between the echo pulse width and the angle of incidence. The simulation developed in this study can be used to estimate neighborhood-independent angles of incidence from the recorded waveform, which enables the angle of incidence calculation during waveform processing. These waveform-derived angles are comparable to established methods based on the local point neighborhood, but generally display a higher variance leading to a mean absolute error of about 10° when compared to neighborhood-based angles of incidence. Using the developed simulation, we also explore angle-dependent shifts of the peak amplitude linked to potential ranging offsets. There, we were able to show ranging offsets of up to 12 cm for strongly asymmetric laser pulses at angles of 80° and no offsets for symmetric laser pulses, which provides new insights into the correctness of topo-bathymetric LiDAR systems. In conclusion, we present a detailed simulation framework which can be used to estimate incidence angles and quantify potential ranging offsets.
Landslides are an important category of geo-hazards which endanger millions of people annually, especially during rainy season. Conventional approaches to monitoring landslides are progressively being trumped by modern methods such as remote sensing and deep learning models. This technology not only supports for landslide identification but also classification. Automatic classification for landslide types has not attended dur to various challenges related to obtaining large, rich data and the overfitting. This study focuses on new process to detect and classify different landslide types using high-resolution remote-sensing images and deep learning models. This study classifies five major types of landslides: (1) Debris flow, (2) Transitional landslides, (3) Rotational landslides, (4) Shallow landslides, and (5) Human-induced landslides. Two architectures including U‑Net and BiSeNet were applied in classifying five types of landslides using WorldView‑2 images with various optimization techniques. These models demonstrate high accuracy for differentiating landslides from non-landslides as well as for classifying landslides into different categories where average accuracy is higher than 95
Deep learning methods such as Mask R‑CNN enable the precise delineation of single-tree crowns from remote sensing data. However, their segmentation performance still depends on local stand conditions. Using UAV multispectral imagery and lidar canopy height models (CHM), we assessed the influence of tree species composition, stand density, and foliage condition on the robustness of deep-learning-based single-tree segmentation. High-resolution laser data and multispectral data were collected over several hectares of forest area (Bavarian Forest National Park; DBU Natural Heritage, Schönau Foundation; Black Forest National Park, Kinzigtal) using a DJI 600 Pro drone. The Fraunhofer Lightweight Airborne Profiler collected a multispectral point cloud using a 905-nm laser and two integrated RGB cameras with 4112 × 3008 pixels. Another multispectral camera captured RGB imagery with 4112 × 3008 pixels and two monochrome bands (725 nm RE, 850 nm NIR; 2164 × 2056 pixels each). Flights were conducted at 80 m altitude with ≥ 50
