Rapid damage assessment after disasters is crucial for humanitarian relief and emergency response. The abrupt and unpredictable nature of disasters causes variations in the time, location, and sensors used for image collection, resulting in significant data disparities in satellite imagery. This poses a significant challenge for assessment tasks. To enable a rapid response, training models from scratch using a sufficient amount of on-site data is impractical due to time constraints. Thus, in practical applications, a model with robust adaptability and generalization is essential for autonomously adjusting to data variations. However, the majority of current models are trained and tested on a singular dataset, neglecting the aforementioned issues. To address these challenges, this study created datasets from the Turkey earthquake with large and dense building features and introduced datasets from the Louisiana hurricane with small and sparse building features. These datasets exhibit significant style differences and cover a broader range of building characteristics, providing a comprehensive evaluation of the model’s adaptability. In the building damage assessment task, the model is trained on public datasets and validated using the newly introduced scenario data. Compared with existing assessment models, the U-net model demonstrates the highest adaptability performance in objectively evaluating damage levels.

Landslide identification is an important task in the field of geologic disaster monitoring and early warning, which is of great significance for improving social safety and mitigating the impact of disasters. With the development of computer vision, deep learning is widely used in landslide recognition research. We focus on segmenting landslides from high-resolution optical satellite images using convolutional neural network. Currently, deep learning semantic segmentation models still face issues such as neglecting small objects and incorrectly segmenting terrain features with similar shapes and pixel characteristics. Considering the unbalanced distribution of categories and large differences in scene styles during the extraction of key feature information from remote sensing images, landslides have diverse and complex backgrounds. We propose a fusion DeepLabv3+ and completed local binary pattern (CLBP) landslide image semantic segmentation method (CLBP-DeepLabv3+), using the improved inverted residual block as the core structure of backbone to extract different levels of image information, and after backbone extracts landslide image features, it connects the improved DenseASPP to fuse the different levels of features to better pay comprehensive attention to local and global features and obtain contextual information at different scales. Then, the texture and edge features of the image are extracted using CLBP, and the multi-level features are merged by introducing the feature aggregation module, which constitutes the CLBP-DeepLabv3+ model. Through ablation experiments and comparative tests on a self-made dataset, the experimental results show that the proposed method performs the best on the validation set, with a mean intersection over union (mIoU) of 88.62%, mean pixel accuracy (mPA) of 94.17%, recall rate of 90.17%, and an intersection over union (IoU) for landslides of 80.53%. Compared with the original DeepLabv3+ model, the improved DeepLabv3+ increased the mIoU by 3.15%, mPA by 3.99%, recall by 4.93%, and IoU by 4.97%. Compared with other semantic segmentation models, the improved DeepLabv3+ also achieved better segmentation accuracy in extracting landslide features.

With advancements in unmanned aerial vehicle (UAV) technology, UAV applications are rapidly growing, and their operations are becoming increasingly intelligent. Localization of UAVs commonly relies on global navigation satellite systems combined with inertial navigation systems through sensor fusion. However, this approach is vulnerable to significant risks, such as signal spoofing. In military conflicts, signal spoofing by hackers poses a severe security threat with potentially catastrophic outcomes. To address this issue, we propose a two-stage vision-based UAV localization method. This approach utilizes multi-category semantic segmentation and template matching to establish a connection between heterogeneous sensors. Experimental results demonstrate the method’s effectiveness in accurately identifying the UAV’s location within extensive geographical areas captured in remote sensing images. In addition, it achieves high precision in aligning UAV locations with Baidu maps, offering robust and accurate localization capabilities.

The digital elevation model (DEM) serves as a vital data source for surface 3D modeling. Due to the limitations in sampling conditions and cost constraints, we usually obtain unevenly distributed and relatively sparse sampling points. To reconstruct a complete DEM of the sampling area, we need to utilize spatial interpolation algorithms. However, traditional spatial interpolation methods typically have lower model complexity and often involve a large number of iterative calculations to approximate the points to be interpolated. This often results in significant interpolation errors and low real-time performance. We propose a multi-scale conditional generative adversarial network (multi-scale cGAN) with adaptive joint loss weights. In addition, during the model training process, we design a joint loss function that incorporates generator adversarial loss, content loss, and perceptual loss, with the ability to adaptively adjust the weight coefficients of each component, thereby optimizing model training and further improving its generalization and generation ability. The experimental results demonstrate that compared with the traditional spatial interpolation algorithm and other typical deep learning–based models, the interpolation error on typical land DEM data (including slopes, valleys, and ridges) is smaller, and the interpolated image has the highest clarity and similarity compared with the original image. Overall, our approach demonstrates high robustness and low error when dealing with DEM spatial interpolation tasks in complex terrain environments while also possessing the potential for expansion into various terrain environment DEM reconstruction applications.

Satellite products are widely used to understand and monitor the dynamics of the Peruvian coastal upwelling system (PCUS), one of the four major eastern boundary upwelling systems. They complement in situ observations obtained through surveys and oceanographic stations periodically operated by the Peruvian Marine Research Institute (IMARPE). Therefore, we mainly aim to assess the reliability of satellite products for sea surface temperature (SST) and surface chlorophyll-a (chl-a) using a unique in situ database from IMARPE. We evaluated SST data from MODIS (levels 2 and 3) and chl-a data from MODIS and SeaWiFS (both at level 3) due to their prior and continuous utilization. In the case of SST, our results demonstrated a strong correlation and good performance from MODIS compared with IMARPE data, although there is a tendency to slightly overestimate it, particularly during autumn. For surface chl-a, during their respective satellite periods, both sensors exhibited a favorable correlation and performance, with SeaWiFS outperforming MODIS. When analyzing matching points in terms of location and timing, both sensors tended to underestimate IMARPE data (autumn). Despite some differences between satellite and IMARPE data, the evaluated satellite products provide reliable information for understanding and monitoring the PCUS. However, for chl-a, it is important to consider the differences between sensors for retrospective studies.

To rapidly and accurately determine the spatial distribution of permanganate index (CODMn) concentrations in surface water, we employed multispectral unmanned aerial vehicle (UAV) remote sensing for water quality inversion monitoring. River extraction was conducted using the normalized difference water index (NDWI) in conjunction with a self-developed parameter a. Fifteen characteristic variables were derived through exponential calculation, and modeling was performed in Jiaobei River using both single-band variables and those with a correlation coefficient exceeding 0.3 and tested in Nanguan River. The model selection included the elastic net regression algorithm, random forest regression algorithm, and gradient boosting regression algorithm. The findings indicated that accurate river extraction is achieved when NDWI exceeds 0.2 and a remains below 0. The gradient boosting regression algorithm proved to be the superior method for constructing an inversion model of CODMn in the Jiaobei River. When applied to the Nanguan River, the model also yielded reliable inversion results. However, in the concentration range of 3 to 5mg/L, the inversion results were less accurate due to an insufficient sample size. A quantitative remote sensing monitoring model of water quality in Taizhou was established using a multispectral UAV, providing technical support for accurate water pollution control.

Using ground control points (GCPs) can improve positioning accuracy in the calibration process, but the acquisition of GCPs is always difficult and high cost, especially in mountain areas and abroad. Thus, we proposed an improved calibration algorithm assisted by the digital elevation model data to calibrate geometric parameters for the single-scene synthetic aperture radar (SAR) image with no need for GCPs by providing an alternative solution that introduces topographic slope as an index of applicability and reliability. The algorithm includes three main steps: SAR image simulation, registration parameters calculation between the real and simulated images using a very large search window, and time delay estimation in the range and azimuth directions based on the above registration results. Experiments on Gaofen-3 SAR images of different modes show that the proposed approach provides good performance without any GCPs, reducing the errors of checkpoints from 151.3 to 26.6 m. Besides, the orthoimages further demonstrate the effectiveness of the algorithm in the whole image, especially in mountainous areas where the layover is well compensated. Finally, the applicability of the proposed method is analyzed, proving its great suitability for SAR images with terrain slopes greater than 0.05.

In recent years, algorithms for fast and accurate recognition and detection of circular markers have become crucial in the field of high-speed videogrammetry. However, most existing techniques often necessitate a manually selected region of interest that encompasses the full information of the circular marker. This manual box selection method is inefficient and unsuitable for practical engineering applications. To address this issue, we propose a global automatic recognition and detection approach that employs multi-level constraints for identifying circular markers in high-speed videogrammetry. First, an edge detection method based on the Canny algorithm is employed to extract candidate edges containing all circular markers. Subsequently, two geometric constraints—general geometric condition and roundness metric constraint—are applied to eliminate a large number of non-circular marker edges. Finally, pseudo-edges of circular markers are removed, and the corresponding accurate edges are retained by applying an extrema point condition constraint. The performance of the proposed method is evaluated using several high-speed videogrammetry image datasets. Experimental results demonstrated that our method can accurately detect and recognize all circular markers, outperforming comparable methods. The proposed method holds promise for efficient and wide-ranging applications in the recognition and detection of circular markers in high-speed videogrammetry.

With the rapid development of artificial intelligence technology, deep learning has achieved significant advantages in synthetic aperture radar automatic target recognition (SAR-ATR). However, previous research showed that the addition of small perturbations not easily detected by the human eye can lead to SAR-ATR model recognition errors; that is, they are affected by adversarial attacks. To solve the problem of long computation time in existing SAR sparse adversarial attack algorithms, we propose a SAR fast sparse adversarial attack (FSAA) algorithm. First, an end-to-end sparse adversarial attack framework is developed based on the lightweight generator ResNet model using two different upsampling modules to control the amplitude and position of the adversarial perturbation. A loss function for the generator is then constructed, which mainly consists of the linear addition of the attack loss, the amplitude distortion loss, and the sparsity loss. Finally, the SAR image is mapped through the trained generator model in a one-step process to generate sparse adversarial perturbations quickly and effectively. Compared with the existing SAR sparse adversarial attack algorithm, the experimental results show that the generation speed of the proposed method is at least 30 times higher when the perturbation is less than 0.05% of the pixels in the entire image, and the recognition rate of the model is >13%.

For the synthetic aperture radar (SAR) target recognition task, feature representation accuracy determines the recognition performance. The previous mainstream deep learning–based SAR target recognition methods required a large amount of labeled samples to learn feature representation and made superior performance. However, only utilizing unlabeled samples to solve SAR target recognition is still a challenging task. To solve this problem, we propose an end-to-end multi-view feature enhancement-based contrastive clustering framework (MvECC) for unsupervised SAR target recognition. It utilizes the view of internal information and complementary information among adjacent views from multiple-view SAR images to learn discriminative target features. MvECC first augments the multi-view image sequences to build sequence pairs and feeds them into a pair of weight-sharing multi-view Vision Transformers (ViT) to extract features. The designed multi-view ViT has an intra-view and an inter-view Transformer layer, and it can capture and fuse the feature within each view and among different views from multi-view image sequences. Then, contrastive learning is performed at the sequence and class levels, which aims at optimizing pairwise similarity to learn feature representations and obtain clustering assignment results. The clustering accuracy of our method outperforms the state-of-the-art method by 7.87% on the moving and stationary target acquisition and recognition dataset. More experimental results on the synthetic and measured paired labeled experiment dataset show that MvECC has good robustness.

Very high-resolution remote sensing images are used to detect the changes with finer results. Accompanying with plenty of details, the environment around targets also becomes more complex, which poses challenges for change detection. In recent years, hybrid methods based on transformer and CNN have been widely used. The two methods usually are adopted to extract the change information because they can take advantage of global semantic relations and long-range spatial dependencies at the same time. However, within some complex environments, there is still information loss and inaccurate detection because the features cannot be fully integrated. We proposed a new cross-level hybrid feature aggregation network for change detection to improve the performance of change detection, especially within a complex environment. Within the new network, a parallel hybrid CNN-Transformer structure is adopted to model globally and locally, which extracts the features of different levels and produces rich semantic features. Meanwhile, the multi-branch feature interaction is used to implement interaction and fusion for multiscale feature information. Furthermore, multiscale feature aggregation was applied to remove redundancy. Subsequently, CNN-Transformer change feature enhancement is used to enhance the representation. Compared with several state-of-the-art methods on three available datasets, the accuracy is increased by 0.09%, 1.12%, and 2.62%, respectively. The experiments indicate that the method proposed in this paper detects the changed targets as continuous and complete objects with clear edges. Within a complex environment, it suppresses pseudo-changes and extracts more small changed targets.

Multi-label scene classification (MLSC) in remote sensing (RS) plays a crucial role in recognizing the intricate contents of RS images. However, the MLSC task is challenged by the complexity of label combinations and the diversity of visual content. Multispectral (MS) data provide valuable information for scene classification. To achieve accurate classification results, it is important to fully utilize the diverse information contained in MS data. To this end, we propose a multispectral transformer network (MST-Net), which leverages transformer-based architectures to capture the diverse information within MS images and the complex relationships between labels and features. Specifically, MST-Net consists of a feature fusion encoder and a semantic query decoder. Within the encoder, we have developed an MS deformable attention module based on a sampling strategy that reduces focus on redundant spectral areas, allowing for better integration of complementary MS information. In the decoder, geographic information is introduced as an inductive bias, utilizing the unique spatiotemporal characteristics of RS images to learn better class-related features. Extensive experiments were conducted on two RS multi-label datasets, namely, LSCIDMRv2 and BigEarthNet. Comparisons with several state-of-the-art multi-label classification methods demonstrate the effectiveness and superiority of MST-Net.