Automatic building extraction from high-resolution remote sensing images has many practical applications, such as urban planning and supervision. However, fine details and various scales of building structures in high-resolution images bring new challenges to building extraction. An increasing number of neural network-based models have been proposed to handle these issues, while they are not efficient enough, and still suffer from the error ground truth labels. To this end, we propose an efficient end-to-end model, EU-Net, in this paper. We first design the dense spatial pyramid pooling (DSPP) to extract dense and multi-scale features simultaneously, which facilitate the extraction of buildings at all scales. Then, the focal loss is used in reverse to suppress the impact of the error labels in ground truth, making the training stage more stable. To assess the universality of the proposed model, we tested it on three public aerial remote sensing datasets: WHU aerial imagery dataset, Massachusetts buildings dataset, and Inria aerial image labeling dataset. Experimental results show that the proposed EU-Net is superior to the state-of-the-art models of all three datasets and increases the prediction efficiency by two to four times.

As two of the most widely used remote sensing images, optical and synthetic aperture radar (SAR) images show abundant and complementary information on the same target owing to their individual imaging mechanisms. Consequently, using optical and SAR images simultaneously can better describe the inherent features of the target, and thus, be beneficial for subsequent remote sensing applications. In this article, we propose a novel modular fully convolutional network model to improve the accuracy of land cover classification by fully exploiting the complementary features of the two sensors. We investigate where and how to fuse the two images in the joint classification network. A cross-gate module with a bidirectional information flow is proposed to achieve the best fusion performance. In addition, to validate the proposed model, we construct a multiclass land cover classification dataset. Exhaustive experiments show that the proposed joint classification network presents superior results than state-of-the-art classification models using single-sensor images.

Emergency flood monitoring and rescue need to first detect flood areas. This paper provides a fast and novel flood detection method and applies it to Gaofen-3 SAR images. The fully convolutional network (FCN), a variant of VGG16, is utilized for flood mapping in this paper. Considering the requirement of flood detection, we fine-tune the model to get higher accuracy results with shorter training time and fewer training samples. Compared with state-of-the-art methods, our proposed algorithm not only gives robust and accurate detection results but also significantly reduces the detection time.

Deep learning-based change detection methods have achieved remarkable success through the feature learning capability of deep convolutions. However, the network structures of existing methods are simply modified from the semantic segmentation models, ignoring the essential characteristics of change detection, thereby limiting their applications. In this work, we propose a category-awareness based difference-threshold alternative learning network (D-TNet) for remote sensing image change detection. Our motivation is to characterize the different change magnitudes for different land cover changes, and represent the semantic content differences of various objects. Thus, our D-TNet consists of a difference map learning path and a threshold map learning path, realizing self-adapting thresholds selection by assigning each pixel a unique threshold. The two paths are alternatively optimized to make the difference map more discriminative, as well as making the threshold map more adaptive. In addition, a category-awareness attention mechanism is introduced in D-TNet, which learns a pixel-to-category relationship to benefit in representing the heterogeneity of land covers. Finally, experimental results on three change detection datasets verify the effectiveness of our D-TNet in both visual and quantitative analysis. Code will be available at: https://www.researchgate.net/profile/Ling-Wan-4.

In recent years, change detection (CD) has achieved remarkable success through using deep learning. However, most existing methods rely on label teaching, and thus have many limitations when dealing with the complexity and diversity of remote sensing scenes. In this work, we propose a self-supervised learning pretraining framework for remote sensing image CD (SSLCD). Our motivation is to leverage the intrinsic structure of multitemporal data to learn general and robust representations, generating competitive pretrained models for the CD task. On the one hand, an intrinsic structure learning strategy is introduced, which enforces feature invariance and change consistency across temporal phases and augmented views, learning discriminatory representations related to changes while simultaneously mitigating noises associated with irrelevant changes. On the other hand, a self/cross-reconstruction mechanism is proposed, which extends masked image modeling to multitemporal images by predicting missing parts of the pre-phase image using the post-phase image, thereby enhancing the model’s capacity in modeling high-level contextual information. Finally, models pretrained using SSLCD are extensively evaluated on three CD datasets, and the results demonstrate that SSLCD outperforms existing remote sensing pretraining methods as well as the state-of-the-art CD methods.

Corners play an important role on image processing, while it is difficult to detect reliable and repeatable corners in SAR images due to the complex property of SAR sensors. In this paper, we propose a fast and novel corner detection method for SAR imagery. First, a local processing window is constructed for each point. We use the local mean of a 3 x 3 mask to represent a single point, which is weighted by a Gaussian template. Then the candidate point is compared with 16 surrounding points in the processing window. Considering the multiplicative property of speckle noise, the similarity measure between the center point and the surrounding points is calculated by the ratio of their local means. If there exist more than M continuous points are different from the center point, then the candidate point is labelled as a corner point. Finally, a selection strategy is implemented by ranking the corner score and employing the non-maxima suppression method. Extreme situations such as isolated bright points are also removed. Experimental results on both simulated and real-world SAR images show that the proposed detector has a high repeatability and a low localization error, compared with other state-of-the-art detectors.