Convolutional Neural Networks (CNNs), such as U-Net, have shown competitive performance in the automatic extraction of buildings from Very High-Resolution (VHR) aerial images. However, due to the unstable multi-scale context aggregation, the insufficient combination of multi-level features and the lack of consideration of the semantic boundary, most existing CNNs produce incomplete segmentation for large-scale buildings and result in predictions with huge uncertainty at building boundaries. This paper presents a novel network with a special boundary-aware loss embedded, called the Boundary-Aware Refined Network (BARNet), to address the gap above. The unique properties of the proposed BARNet are the gated-attention refined fusion unit, the denser atrous spatial pyramid pooling module, and the boundary-aware loss. The performance of the BARNet is tested on two popular data sets that include various urban scenes and diverse patterns of buildings. Experimental results demonstrate that the proposed method outperforms several state-of-the-art approaches in both visual interpretation and quantitative evaluations.

As an inevitable phenomenon in most optical remote-sensing images, the effect of shadows is prominent in urban scenes. Shadow detection is critical for exploiting shadows and recovering the distorted information. Unfortunately, in general, automatic shadow detection methods for urban aerial images cannot achieve satisfactory performance due to the limitation of feature patterns and the lack of consideration of non-local contextual information. To address this challenging problem, the global-spatial-context-attention (GSCA) module was developed to self-adaptively aggregate all global contextual information over the spatial dimension for each pixel in this paper. The GSCA module was embedded into a modified U-shaped encoder–decoder network that was derived from the UNet network to output the final shadow predictions. The network was trained on a newly created shadow detection dataset, and the binary cross-entropy (BCE) loss function was modified to enhance the training procedure. The performance of the proposed method was evaluated on several typical urban aerial images. Experiment results suggested that the proposed method achieved a better trade-off between automaticity and accuracy. The F1-score, overall accuracy, balanced-error-rate, and intersection-over-union metrics of the proposed method were higher than those of other state-of-the-art shadow detection methods.

Remote sensing image change detection method plays a great role in land cover research, disaster assessment, medical diagnosis, video surveillance and other fields, so it has attracted wide attention. Based on a small sample data set from SZTAKI AirChange Benchmark Set, in order to solve the problem that the deep learning network needs a large number of samples, this work firstly uses non generative sample augmentation method and generative sample augmentation method based on DCGAN (Deep Convolutional Generative Adversarial Networks), and then constructs a remote sensing image change detection model based on an improved DeepLabv3+ network. This model can realize end-to-end training and prediction of remote sensing image change detection with sub-pixel convolution. Finally, Landsat 8, Google Earth, and OSCD (Onera Satellite Change Detection) dataset are used to verify the generalization performance of this network. The experimental results show that the improved network accuracy is 95.1% and the generalization performance is acceptable.

With the rapid development of unmanned aerial vehicle (UAV) technology, UAV remote sensing images are increasing sharply. However, due to the limitation of the perspective of UAV remote sensing, the UAV images obtained from different viewpoints of a same scene need to be stitched together for further applications. Therefore, an automatic registration method of UAV remote sensing images based on deep residual features is proposed in this work. It needs no additional training and does not depend on image features, such as points, lines and shapes, or on specific image contents. This registration framework is built as follows: Aimed at the problem that most of traditional registration methods only use low-level features for registration, we adopted deep residual neural network features extracted by an excellent deep neural network, ResNet-50. Then, a tensor product was employed to construct feature description vectors through exacted high-level abstract features. At last, the progressive consistency algorithm (PROSAC) was exploited to remove false matches and fit a geometric transform model so as to enhance registration accuracy. The experimental results for different typical scene images with different resolutions acquired by different UAV image sensors indicate that the improved algorithm can achieve higher registration accuracy than a state-of-the-art deep learning registration algorithm and other popular registration algorithms.

Image registration plays a vital role in the mosaic process of multiple UAV (Unmanned Aerial Vehicle) images acquired from different spatial positions of the same scene. Aimed at the problem that many fast registration methods cannot provide both high speed and accuracy simultaneously for UAV visible light images, this work proposes a novel registration framework based on a popular baseline registration algorithm, ORB—the Oriented FAST (Features from Accelerated Segment Test) and Rotated BRIEF (Binary Robust Independent Elemental Features) algorithm. First, the ORB algorithm is utilized to extract image feature points fast. On this basis, two bidirectional matching strategies are presented to match obtained feature points. Then, the PROSRC (Progressive Sample Consensus) algorithm is applied to remove false matches. Finally, the experiments are carried out on UAV image pairs about different scenes including urban, road, building, farmland, and forest. Compared with the original version and other state-of-the-art registration methods, the bi-matching ORB algorithm exhibits higher accuracy and faster speed without any training or prior knowledge. Meanwhile, its complexity is quite low for on-board realization.

Deep learning (DL) has been widely applied to enhance the accuracy and reconstruction of snowpack parameters due to its powerful nonlinear mapping ability. Here, we propose a broadband snow albedo (BA) estimation method that infuses the asymptotic radiative transfer (ART) model into the DL model and thus imposes constraints on the input variables of the DL model and compares it to the MOD10A1 BA product and the BA simulated using the ART model alone. The results show that coupling a physical mechanism-based snow radiative transfer model with a DL model can successfully improve the estimation accuracy of BA. Compared to the MOD10A1 BA product and the BA simulated using the ART model alone, the accuracy of the BA simulated by the convolutional neural network model under the constraints of the ART model is improved by 12% and 3%, respectively, and the accuracy of the BA simulated by the long short-term memory model under the constraints of the ART model is improved by 14% and 5%, respectively. The coupling paradigm that integrates the snow radiative transfer model into the DL model effectively balances the advantages of both models. This coupling paradigm creates a complementary association between the snow radiative transfer model and the DL model. This research effectively enhances the simulation and predictive capabilities of snow albedo and is expected to promote the in-depth application of remote sensing big data in combination with physical modeling in the field of cryosphere parameter estimation.