Yue Xu

Remote sensing visual grounding (RSVG) aims to accurately localize specific targets in remote sensing (RS) images based on natural language descriptions. However, existing RSVG datasets often contain overly simplistic textual descriptions, exhibit imbalanced object size distributions, and lack semantic connections between targets and the surrounding contexts. Moreover, current approaches rely on global visual features for region-level localization, while the interaction between visual and textual modalities remains limited. To address these challenges, we propose improvements from both the dataset and algorithm perspectives. First, we explore the semantic correlation between remote sensing objects and their scene context. Based on high-resolution Gaofen satellite imagery, we expand several typical object categories and construct richer textual descriptions that reflect object-background associations. By refining and extending the existing DIOR-RSVG dataset, we build a new dataset named DGF-RSVG. Second, to enhance the semantic alignment between global visual features and textual features, we propose a novel global vision-language multimodal feature interaction enhancement module (GME). In parallel, we design a local attention enhancement module (LAE) to facilitate fine-grained interaction between object-related textual features and regional visual proposals. These two modules form the foundation of our newly developed detection framework: the global-local attention enhanced detector (GLAED). Extensive experiments show that GLAED achieves state-of-the-art performance on the DGF-RSVG dataset, outperforming the closest competitor by 4.3% in Pr@0.5. It also achieves highly competitive results on the DIOR-RSVG dataset, demonstrating the effectiveness of both our proposed dataset and model framework.

This paper addresses the challenges of sample scarcity and class imbalance in remote sensing image semantic segmentation by proposing a decoupled synthetic sample generation framework based on a latent diffusion model. The method consists of two stages. In the label generation stage, we fine-tune a pretrained latent diffusion model with LoRA to generate semantic label masks from textual descriptions. A novel proportion-aware loss function explicitly penalizes deviations from the desired class distribution in the generated mask. In the image generation stage, we use ControlNet to train a multi-condition image generation network that takes the synthesized mask, along with its text description, as input and produces a realistic remote sensing image. The base Stable Diffusion model’s weights remain frozen during this process, with the trainable ControlNet ensuring that outputs are structurally and semantically aligned with the input labels. This two-stage approach yields coherent image–mask pairs that are well-suited for training segmentation models. Experiments show that models trained on the synthetic samples produced by the proposed method achieve high visual quality and semantic consistency. The proportion-aware loss effectively mitigates the impact of minority classes, boosting segmentation performance on under-represented categories. Results also reveal that adding a suitable proportion of synthetic sample improves segmentation accuracy, whereas an excessive share can cause over-fitting or misclassification. Comparative tests across multiple models confirm the generality and robustness of the approach.

Urban trees have significant ecological and social functions, and generating Urban Tree Canopy (UTC) maps is an effective method for understanding their distribution. However, existing studies primarily rely on medium- to low-resolution imagery for large-scale extraction or high-resolution imagery for small-scale extraction, making it challenging to balance spatial coverage and accuracy. To comprehensively analyze the tree canopy characteristics of urban trees across large areas, this study selects Beijing as the research area, and employs highresolution remote sensing imagery with deep learning techniques to construct the UTC map of Beijing. Deep learning models typically require large amounts of labeled data for training, but manually annotating pixel-level labels is time-consuming and costly. To address this, we propose an iterative method based on Intersection over Union (IoU), integrating multi-scale segmentation and nearest-neighbor classification algorithms. By leveraging a limited number of labeled pixels, this approach efficiently generates a extensive labeled data, enabling the rapid construction of a high-precision dataset. Using this method, we established the Beijing Urban Tree Canopy Dataset (BUTCD). By Employing BUTCD and InternImage T+UPerNet as the tree canopy segmentation network, we generated the urban tree canopy map for Beijing. Accuracy assessments demonstrated an accuracy of 0.9806, precision of 0.9569, recall of 0.9702, mean IoU (mIoU) of 0.9296, and F1-score of 0.9635. We analyzed the results of Beijing's urban tree canopy, revealing that the tree canopy coverage rate is approximately 43.99%, with higher coverage in the northwest and lower coverage in the southeast. The per capita tree canopy area in Beijing was calculated as 330.29 m². Within Beijing's central urban area, trees outside forests account for 26.54% of the the total urban tree population, indicating that conventional statistics on tree-related data in this area tend to underestimate the true value by about onefourth. This underestimation may originate from the substantial contributions of trees outside forest areas being neglected during the statistical process, especially in the central urbanized districts. We anticipate that the urban tree canopy map of Beijing will facilitate research on the city's ecosystem services, advance sustainable urban development, and improve the quality of life for residents.

With the advent of high-quality SAR images and the rapid development of computing technology, the object detection algorithms based on convolution neural network have attracted a lot of attention in the field of SAR object detection. At present, the main dataset for SAR target detection in China focus on ships, there is a lack of SAR vehicle detect datasets, and complex ground scenes can affect vehicle detection performance. To solve these problems, we propose a lightweight SAR vehicle detection algorithm, aiming to improve the vehicle detection accuracy and simplify the model complexity. First, we constructed a multi-band SAR vehicle detection dataset (SVDD) with annotations as the training dataset of the object detection model. Then, we introduce dual conv into the RT-DETR model. Dual conv uses group convolution technology to filter the convolutional network to reduce model parameters, so we can achieve a lightweight and real-time end-to-end detection. Finally, we use the mmdetection framework as a benchmark performance and test the robust performance under different conditions. Experimental results show that the AP₅₀ of we proposed method reaches 98.5%, achieving excellent detection performance.

Semantic segmentation in remote sensing is a fundamental task with crucial applications across various domains. Traditional approaches primarily utilize bottom-up discriminative methods, where network architectures learn image features to generate segmentation masks. However, the complexity of remote sensing images, characterized by diverse ground object types and intricate scenes, often results in information redundancy and confusion during feature extraction, impacting segmentation accuracy. To address these challenges, we introduce a novel segmentation framework, DiffRSS, based on the denoising model paradigm. This top-down generative approach learns the data distribution of sample labels and uses image features as guiding priors for generating segmentation masks. We conceptualize the semantic segmentation of remote sensing images as a conditional generation task and design a Multiscale Cyclic Denoising Module (MSCDM), which effectively leverages multiscale features of remote sensing images, leading to superior segmentation outcomes. Inspired by diffusion models, our denoising structure, MSCDM, can be reused multiple times during inference, enhancing the quality of segmentation masks. This method allows for more precise capture and utilization of image features, resulting in finer and more accurate segmentation masks. Extensive testing on three public remote sensing datasets the ISPRS Vaihingen, ISPRS Potsdam, and GID Fine Land Cover Classification datasets demonstrates that our method achieves competitive results.

Wind turbines are one of the important carriers of clean energy utilization. Accurately and rapidly detecting wind turbine objects in large-scale remote sensing images can effectively monitor the development activities and optimize energy utilization. Addressing the detection challenges posed by the complex distribution scenes and the slender, dispersed structural characteristics of wind turbines in remote sensing images, this paper proposes a remote sensing image wind turbine detector, RSWDet, based on neural networks. RSWDet comprises two innovative key modules. The first is a dual-branch structured point set detection head, which, through training, adapts to the unique features of wind turbines, enabling accurate detection in large-scale complex backgrounds. The second is the Low-level Feature Enhancement (LFE) module, which compensates for the loss of wind turbine feature information during sampling by leveraging rich low-level feature information. Experimental verification of RSWDet was conducted on datasets and real-world scenes. The results demonstrate that RSWDet exhibits significant advantages compared to other algorithms, achieving the highest average accuracy of 83.1%, precision of 97.8%, and recall of 99% on the validation set. In the actual multi-scene GF2 remote sensing image test, with a threshold of 0.4, the precision can reach 85.3%, and the recall can reach 89.9%.

The study of few-shot SAR image generation is an effective way to expand the SAR dataset, which not only provides diversified data support for SAR target classification, but also provides a high-fidelity false image template for SAR deceptive jamming. In this paper, we have constructed a multi-frequency and multi-target type SAR vehicle imagery dataset that encompasses frequencies such as X, Ka, P, and S bands. The vehicle types are coaster, suv and cabin. Subsequently, we utilized various Generative Adversarial Networks for image generation from the SAR vehicle dataset. The experimental result indicates that the images generated by the DCGAN and the LSGAN models are of superior quality. Furthermore, we employed different recognition networks to evaluate the classification accuracy of the generated images. Of all the frequency bands, the Ka band generated images achieved the highest recognition rate, with an accuracy of up to 99%. Under conditions of a limited number of samples, the LSGAN model performed the best, reaching a classification recognition rate of 71.48% with a dataset of only 20 samples. Finally, we use a conditional network generation model to generate conditions based on target categories and frequency bands, providing high fidelity samples for SAR deception jamming.

Land-cover mapping plays a crucial role in resource detection, ecological environmental protection, and sustainable development planning. The existing large-scale land-cover products with coarse spatial resolution have a wide range of categories, but they suffer from low mapping accuracy. Conversely, land-cover products with fine spatial resolution tend to lack diversity in the types of land cover they encompass. Currently, there is a lack of large-scale land-cover products simultaneously possessing fine-grained classifications and high accuracy. Therefore, we propose a mapping framework for fine-grained land-cover classification. Firstly, we propose an iterative method for developing fine-grained classification systems, establishing a classification system suitable for Sentinel-2 data based on the target area. This system comprises 23 fine-grained land-cover types and achieves the most stable mapping results. Secondly, to address the challenges in large-scale scenes, such as varying scales of target features, imbalanced sample quantities, and the weak connectivity of slender features, we propose an improved network based on Swin-UNet. This network incorporates a pyramid pooling module and a weighted combination loss function based on class balance. Additionally, we independently trained models for roads and water. Guided by the natural spatial relationships, we used a voting algorithm to integrate predictions from these independent models with the full classification model. Based on this framework, we created the 2017 Beijing–Tianjin–Hebei regional fine-grained land-cover product JJJLC-10. Through validation using 4254 sample datasets, the results indicate that JJJLC-10 achieves an overall accuracy of 80.3% in the I-level validation system (covering seven land-cover types) and 72.2% in the II-level validation system (covering 23 land-cover types), with kappa coefficients of 0.7602 and 0.706, respectively. In comparison with widely used land-cover products, JJJLC-10 excels in accurately depicting the spatial distribution of various land-cover types and exhibits significant advantages in terms of classification quantity and accuracy.

Change detection from heterogeneous satellite and aerial images plays a progressively important role in many fields, including disaster assessment, urban construction, and land use monitoring. Currently, researchers have mainly devoted their attention to change detection using homologous image pairs and achieved many remarkable results. It is sometimes necessary to use heterogeneous images for change detection in practical scenarios due to missing images, emergency situations, and cloud and fog occlusion. However, heterogeneous change detection still faces great challenges, especially using satellite and aerial images. The main challenges in satellite and aerial image change detection are related to the resolution gap and blurred edge. Previous studies used interpolation or shallow feature alignment before traditional homologous change detection methods, which ignored the high-level feature interaction and edge information. Therefore, we propose a new heterogeneous change detection model based on multimodal transformers combined with edge guidance. In order to alleviate the resolution gap between satellite and aerial images, we design an improved spatially aligned transformer (SP-T) with a sub-pixel module to align the satellite features to the same size of the aerial ones supervised by a token loss. Moreover, we introduce an edge detection branch to guide change features using the object edge with an auxiliary edge-change loss. Finally, we conduct considerable experiments to verify the effectiveness and superiority of our proposed model (EGMT-CD) on a new satellite–aerial heterogeneous change dataset, named SACD. The experiments show that our method (EGMT-CD) outperforms many previously superior change detection methods and fully demonstrates its potential in heterogeneous change detection from satellite–aerial images.