Accurate building extraction from remotely sensed data is difficult to perform automatically because of the complex environments and the complex shapes, colours and textures of buildings. Supervised deep-learning-based methods offer a possible solution to solve this problem. However, these methods generally require many high-quality, manually labelled samples to obtain satisfactory test results, and their production is time and labour intensive. For multimodal data with sufficient information, extracting buildings accurately in as unsupervised a manner as possible. Combining remote sensing images and LiDAR point clouds for unsupervised building extraction is not a new idea, but existing methods often experience two problems: (1) the accuracy of vegetation detection is often not high, which leads to limited building extraction accuracy, and (2) they lack a proper mechanism to further refine the building masks. We propose two methods to address these problems, combining aerial images and aerial LiDAR point clouds. First, we improve two recently developed vegetation detection methods to generate accurate initial building masks. We then refine the building masks based on the image feature consistency constraint, which can replace inaccurate LiDAR-derived boundaries with accurate image-based boundaries, remove the remaining vegetation points and recover some missing building points. Our methods do not require manual parameter tuning or manual data labelling, but still exhibit a competitive performance compared to 29 methods: our methods exhibit accuracies higher than or comparable to 19 state-of-the-art methods (including 8 deep-learning-based methods and 11 unsupervised methods, and 9 of them combine remote sensing images and 3D data), and outperform the top 10 methods (4 of them combine remote sensing images and LiDAR data) evaluated using all three test areas of the Vaihingen dataset on the official website of the ISPRS Test Project on Urban Classification and 3D Building Reconstruction in average area quality. These comparative results verify that our unsupervised methods combining multisource data are very effective.

Few-shot semantic segmentation has gained significant attention due to its ability to segment novel objects using only a limited number of labeled samples, thereby addressing the problem of overfitting caused by a lack of training data. Although this technique is widely studied in the field of computer vision, there are few methods for remote sensing images. Prevalent few-shot semantic segmentation methods can achieve remarkable results for natural images, but they are difficult to apply to remote sensing image processing because existing methods rarely take into consideration the large scale and resolution differences in remote sensing images. Consequently, it is hard for them to obtain correct semantic guidance from few annotated remote sensing images. To tackle these problems, this article proposes the pyramid correlation fusion network (PCFNet) to promote the ability to mine helpful information by calculating multi-scale pixel-wise semantic correspondence. Particularly, the dual distance correlation (DDC) module is designed to simultaneously compute the cosine similarity and Euclidean distance between query features and support features, producing adequate guidance information to determine the category of each pixel. Moreover, to improve segmentation accuracy for small objects, the scale-aware cross-entropy loss (SACELoss) is introduced to dynamically assign loss weights according to the actual sizes of objects. This enables smaller objects to be assigned larger weight values and thus receive more attention during training. Comprehensive experiments on both the iSAID-5 ⁱ and DLRSD-5 ⁱ datasets demonstrate that our method outperforms state-of-the-art few-shot semantic segmentation methods. Our code is available at https://github.com/TinyAway/PCFNet.

Graph convolutional networks (GCN) have emerged as a powerful alternative tool for analyzing hyperspectral images (HSIs). Despite their impressive performance, current works strive to make GCN more sophisticated through either elaborate architecture or fancy training tricks, making them prohibitive for HSI data in practice. In this paper, we present a Graph Convolutional RVFL Network (GCRVFL), a simple but efficient GCN for hyperspectral image classification. Specifically, we generalize the classic RVFL network into the graph domain by using graph convolution operations. This not only enables RVFL to handle graph-structured data, but also avoids iterative parameter adjustment by employing an efficient closed-form solution. Unlike previous works that perform HSI classification under a transductive framework, we regard HSI classification as a graph-level classification task, which makes GCRVFL scalable to large-scale HSI data. Extensive experiments on three benchmark data sets demonstrate that the proposed GCRVFL is able to achieve competitive results with fewer trainable parameters and adjustable hyperparameters and higher computational efficiency. In particular, we show that our approach is comparable to many existing approaches, including deep CNN models (e.g., ResNet and DenseNet) and popular GCN models (e.g., SGC and APPNP).

As one of the representative models in the field of image generation, generative adversarial networks (GANs) face a significant challenge: how to make the best trade-off between the quality of generated images and training stability. The U-Net based GAN (U-Net GAN), a recently developed approach, can generate high-quality synthetic images by using a U-Net architecture for the discriminator. However, this model may suffer from severe mode collapse. In this study, a stable U-Net GAN (SUGAN) is proposed to mainly solve this problem. First, a gradient normalization module is introduced to the discriminator of U-Net GAN. This module effectively reduces gradient magnitudes, thereby greatly alleviating the problems of gradient instability and overfitting. As a result, the training stability of the GAN model is improved. Additionally, in order to solve the problem of blurred edges of the generated images, a modified residual network is used in the generator. This modification enhances its ability to capture image details, leading to higher-definition generated images. Extensive experiments conducted on several datasets show that the proposed SUGAN significantly improves over the Inception Score (IS) and Fréchet Inception Distance (FID) metrics compared with several state-of-the-art and classic GANs. The training process of our SUGAN is stable, and the quality and diversity of the generated samples are higher. This clearly demonstrates the effectiveness of our approach for image generation tasks. The source code and trained model of our SUGAN have been publicly released.

Few-shot semantic segmentation aims to recognize novel classes with only very few labelled data. This challenging task requires mining of the correlation between the query image and the support images. Previous works have typically regarded it as a pixel-wise classification problem. Therefore, various models have been designed to explore the correlation of pixels between the query image and the support images. However, they focus only on pixel-wise correspondence and ignore the overall correlation of objects. In this paper, we introduce a mask-based classification method for addressing this problem. The mask aggregation network, which is a simple mask classification model, is proposed to simultaneously generate a fixed number of masks and their probabilities of being targets. Then, the final segmentation result is obtained by aggregating all the masks according to their locations. Experiments on both the PASCAL- $$5^i$$5 i and COCO- $$20^i$$20 i datasets show that our method performs comparably to the state-of-the-art pixel-based methods. This competitive performance demonstrates the potential of mask classification as an alternative baseline method for few-shot semantic segmentation.