The detection and monitoring of changes in urban buildings, as a major place for human activities, have been considered profoundly in the field of remote sensing. In recent years, comparing with other traditional methods, the deep learning-based methods have become the mainstream methods for urban building change detection due to their strong learning ability and robustness. Unfortunately, often, it is difficult and costly to obtain sufficient samples for the change detection method development. As a result, the application of the deep learning-based building change detection methods is limited in practice. In our work, we proposed a novel multi-task network based on the idea of transfer learning, which is less dependent on change detection samples by appropriately selecting high-dimensional features for sharing and a unique decoding module. Different from other multi-task change detection networks, with the help of a high-accuracy building mask, our network can fully utilize the prior information from building detection branches and further improve the change detection result through the proposed object-level refinement algorithm. To evaluate the proposed method, experiments on the publicly available WHU Building Change Dataset were conducted. The experimental results show that the proposed method achieves F1 values of 0.8939, 0.9037, and 0.9212, respectively, when 10%, 25%, and 50% of change detection training samples are used for network training under the same conditions, thus, outperforming other methods.

Cloud, one of the poor atmospheric conditions, significantly reduces the usability of optical remote-sensing data and hampers follow-up applications. Thus, the identification of cloud remains a priority for various remote-sensing activities, such as product retrieval, land-use/cover classification, object detection, and especially for change detection. However, the complexity of clouds themselves make it difficult to detect thin clouds and small isolated clouds. To accurately detect clouds in satellite imagery, we propose a novel neural network named the Pyramid Contextual Network (PCNet). Considering the limited applicability of a regular convolution kernel, we employed a Dilated Residual Block (DRB) to extend the receptive field of the network, which contains a dilated convolution and residual connection. To improve the detection ability for thin clouds, the proposed new model, pyramid contextual block (PCB), was used to generate global information at different scales. FengYun-3D MERSI-II remote-sensing images covering China with 14,165 × 24,659 pixels, acquired on 17 July 2019, are processed to conduct cloud-detection experiments. Experimental results show that the overall precision rates of the trained network reach $97.1\%$ and the overall recall rates reach $93.2\%$, which performs better both in quantity and quality than U-Net, UNet++, UNet3+, PSPNet and DeepLabV3+.

Change detection techniques, which extract different regions of interest from bi-temporal remote sensing images, play a crucial role in various fields such as environmental protection, damage assessment, and urban planning. However, visual style interferences stemming from varying acquisition times, such as radiation, weather, and phenology changes, often lead to false detections. Existing methods struggle to robustly measure background similarity in the presence of such discrepancies and lack quantitative validation for assessing their effectiveness. To address these limitations, we propose Representation Consistency Change Detection (RCCD), a novel deep learning framework that enforces global style and local spatial consistency of features across encoding and decoding stages for robust cross-visual style change detection. RCCD leverages large-kernel convolutional supervision for local spatial context awareness and global content-aware style transfer for feature harmonization, effectively suppressing interference from background variations. Extensive evaluations on S2Looking and LEVIR-CD+ datasets demonstrate RCCD’s superior performance, achieving state-of-the-art F1-scores. Furthermore, on dedicated subsets with large visual style differences, RCCD exhibits more substantial improvements, highlighting its effectiveness in mitigating interference caused by visual style errors. The code has been open-sourced on GitHub.

High-resolution visible remote sensing imagery, as a fundamental contributor to Earth observation, has found extensive application in land use classification. However, the heterogeneous array of optical sensors, distinguished by their unique design architectures, exhibit disparate spectral responses and spatial distributions when observing ground objects. These discrepancies between multi-sensor data present a significant obstacle to the widespread application of intelligent methods. In this paper, we propose a method tailored to accommodate these disparities, with the aim of achieving a smooth transfer for the model across diverse sets of images captured by different sensors. Specifically, to address the discrepancies in spatial resolution, a novel positional encoding has been incorporated to capture the correlation between the spatial resolution details and the characteristics of ground objects. To tackle spectral disparities, random amplitude mixup augmentation is introduced to mitigate the impact of feature anisotropy resulting from discrepancies in low-level features between multi-sensor images. Additionally, we integrate convolutional neural networks and Transformers to enhance the model’s feature extraction capabilities, and employ a fine-tuning strategy with dynamic pseudo-labels to reduce the reliance on annotated data from the target domain. In the experimental section, the Gaofen-2 images (4 m) and the Sentinel-2 images (10 m) were selected as training and test datasets to simulate cross-sensor model transfer scenarios. Also, Google Earth images of Suzhou City, Jiangsu Province, were utilized for further validation. The results indicate that our approach effectively mitigates the degradation in model performance attributed to image source inconsistencies.

Image segmentation represents a fundamental step in analyzing very high spatial-resolution (VHR) remote sensing imagery. Its objective is to partition an image into segments that best match with geo-objects. However, the diverse appearances of geospatial objects often lead to inter-object homogeneity and intra-object heterogeneity. Existing segmentation methods often struggle to accurately segment geo-objects with varying shapes and scales. To address these challenges, we propose Deep-Merge, a novel method that integrates deep learning and region adjacency graphs (RAGs) to accurately segment complete geo-objects in large VHR images. DeepMerge begins with an initial over-segmentation of the image and then iteratively merges similar regions to achieve complete geo-object segmentation. A deep learning model is employed to learn the similarity between adjacent super-pixel pairs. This approach only requires labels indicating whether adjacent superpixels belong to the same geo-object eliminating the need for object-level annotations, enabling weakly-supervised segmentation. A cross-scale module is incorporated to capture multi-scale information, enhancing the representation of superpixels. In addition, the feature distances between neighboring super-pixels are deemed as scale parameters (thresholds) to control the merging procedure, thus yielding an interpretable, predictable, stable, and the optimal scale parameter 0.5. DeepMerge can achieve high segmentation accuracy in a weakly supervised manner, which are validated on large-scale remote sensing images of 0.55 m resolution covering an area of 5,660 km². The experimental results demonstrate that DeepMerge achieves the highest F value (0.9552) and the lowest total error (0.0827), accurately segmenting geo-objects varying sizes and outperforming all competing methods.

As an inherent characteristic of Synthetic Aperture Radar (SAR) systems, the presence of speckle noise reduces the signal-to-noise ratio (SNR) of SAR images, leading to blurred image details and limiting the accuracy of change detection in SAR images. For C-band SAR images, different satellite imaging and ground object reflections can produce varying levels of noise, necessitating a network design adaptable to different noise levels. In this study, we introduce a focal loss function to enhance the robustness of the network, allowing for high-precision change detection of SAR images with different noise levels through parameter adjustment. To address the issue of low SNR in SAR images, we construct a deep learning network focused on feature exploration and utilization to effectively distinguish between noise and change information. The model initially employs discrete wavelet decomposition to extract high-frequency bands for noise reduction. Subsequently, conventional convolutions are utilized to extract high-level semantic information from frequency-domain images, with the resulting feature maps serving as patch inputs to the ConvMixer module to acquire blended spatial and positional information. In the decoder phase, valuable encoder features from multi-scale attention gates are concatenated with corresponding upsampled features to select encoder features based on different resolutions adaptively. Through comparative experiments on real datasets, the superiority of the model is demonstrated.