Super-resolution (SR) reconstruction of remote sensing images aims to improve image resolution while ensuring accurate spatial texture information. In most multi-scale SR methods, the feature fusion at each layer contains only the multi-scale features of the current layer. However, this approach does not optimally use these multi-scale features over different layers, leading to their gradual disappearance during the process of transmission. To address this problem, we propose a Multi-Scale Residual Dense Network (MRDN) for SR. The feature fusion of each layer in MRDN contains multi-scale features from all preceding layers, rather than only fusing the features of the current layer. Specifically, MRDN concatenates the output of each layer and passes it to the subsequent multi-scale layers to facilitate feature fusion. MRDN maximizes the utilization of hierarchical features from the original low-resolution images, enabling adaptive learning of more effective features. In addition, efficient MRDN does not necessitate a substantial increase in network depth and complexity to achieve high performance. Experimental results indicate that MRDN outperforms the state-of-the-art methods on three remote sensing datasets. To demonstrate the generalizability of MRDN, we extend its application to three relevant tasks: natural image SR, real-world image SR, and small object recognition. MRDN achieves competitive results on these tasks, confirming its generalizability.

Accurate building extraction from very high-resolution (VHR) remote sensing images plays an important role in urban dynamic monitoring, planning, and management. However, it is still a challenging task to achieve building extraction with high accuracy and integrity due to diverse building appearances and more complex ground background in VHR remote sensing images. Recently, unity networking (UNet) has been proven to be capable of feature extraction and semantic segmentation of remote sensing images. However, UNet cannot achieve sufficient multiscale and multilevel features with larger receptive fields. To address these problems, an improved network based on UNet structure (Refine-UNet) is proposed for extracting buildings from the VHR images. The proposed Refine-UNet mainly consists of an encoder module, a decoder module, and a refine skip connection scheme. The refine skip connection scheme is composed of an atrous spatial convolutional pyramid pooling (ASPP) module and several improved depthwise separable convolution (IDSC) modules. Experimental results on the Jilin-1 VHR datasets with a spatial resolution of 0.75 m demonstrate that compared with UNet, pyramid scene parsing network (PSPNet), DeepLabV3+, and a deep convolutional encoder-decoder architecture for image segmentation (SegNet), the proposed Refine-UNet can obtain more accurate building extraction results and achieve the best precision of 95.1% and intersection over union (IoU) of 87.0%, indicating the great practical potential.

Farmland segmentation scenario from remote sensing images plays an important role in crop growth monitoring, precision agriculture and intelligent agriculture. To achieve high precision segmentation of farmland boundary, a Multi-task Deformable UNet combined Enhanced network (MDE-UNet) is proposed for farmland boundary segmentation. The network consists of two parts: a Multi-task Deformable UNet (MD-UNet) segmentation module with Deformable UNet (D-UNet) as the basic network and an enhancement module with a lightweight UNet improved by residual attention. In the MD-UNet segmentation module, three branches are used for precise segmentation of deterministic, fuzzy, and raw boundary, respectively. In the enhancement module, an improved lightweight UNet is designed, which can enhance the feature extraction ability of the MD-UNet segmentation module and further improve the segmentation accuracy. The accuracy and mIoU in the GF-2 farmland segmentation test dataset can reach 96.41% and 91.29% using the proposed model, respectively. The MDE-UNet method outperforms other representative deep learning methods such as DeepLab v3+, FCN-8s, SegFormer, and UTNet, and has potential for practical applications of farmland boundary segmentation.

Compared with most ship detection methods for synthetic aperture radar (SAR) images, ship contour extraction can provide more of the detailed shape and edge information of an observed ship and play a significant role in sea surface monitoring and marine transportation. In this study, a joint ship contour extraction method (faster region convolutional neural network (R-CNN), fast nonlocal mean (FNLM) filter and Chan–Vese model (FFCV) method) was proposed to obtain detailed ship information from SAR images, including ship detection in complex scenes and contour extraction in target slices. First, Faster R-CNN was employed to slice ships from large-scene SAR images. Then, FNLM filtering was applied to denoise and enhance the structural information of the target slices. Finally, an optimized Chan–Vese model was proposed in this article, which can not only accurately extract the contour of the observed ship but also reduce the computation time of the model. The SAR ship detection dataset (SSDD) was selected and finely relabeled to evaluate the contour extraction performance. An evaluation index$R_{N}$, including quantitative value and offset direction, was developed to evaluate the extraction accuracy of the target contour from the SAR images. Compared with the Mask R-CNN network, the average contour extraction accuracy index$R_{N}$of the proposed FFCV method reached −0.002 on all the images in the SSDD dataset, and its results were closer to the real ship contours while maintaining the applicability to complex scenes.

Owing to the characteristics of cloud and snow, it is difficult to detect them when they coexist. Thus, we propose an end-to-end semantic segmentation network for cloud and snow detection based on hybrid feature (CSD-HFnet) to be applied to remote sensing images (RSIs) in which cloud and snow coexist. First, the local binary pattern (LBP), gray-level co-occurrence matrix (GLCM), and superpixel segmentation are combined as basic features to preserve the textural, spatial and shape information of the cloud and snow objects. We also propose deep learning feature extraction network to obtain the multi-scale deep learning features, which is utilized to better distinguish cloud from snow. The original spectral bands, the basic features, and the multi-scale deep learning features are input into the feature integration module simultaneously to form the primary multi-scale hybrid features (PMHF). Then, the PMHF are filtered, sorted and weighted by the feature filtering & sorting block and attention mechanism block to obtain the advanced multi-scale hybrid features (AMHF). Finally, the AMHF are input into the cloud and snow segmentation network, which consists of several memory-capable gate circuits for training and validation of cloud and snow detection. The results indicate that CSD-HFnet with AMHF can provide reliable detection under the condition of cloud and snow coexistence. CSD-HFnet can detect cloud and snow in multi-spectral RSIs of various spatial resolutions with an OA of 95.37%. Moreover, CSD-HFnet exhibits excellent cloud detection ability for red-green-blue (RGB) images with an OA of 95.88%, higher than other state-of-the-art cloud detection methods.

Considering the tradeoff between spatial resolution and temporal resolution, spatiotemporal fusion has become a promising technique to monitor snow cover dynamics with both high spatial and temporal resolutions. The representative spatiotemporal fusion methods, e.g. Spatial Temporal Data Fusion Approach (STDFA), usually exist obvious phenomenon of spectral distortion when the surface reflectance changes nonlinearly, which affects the quality of the spatiotemporal fusion image. To address this issue, an effective STDFA-Matching-Pix2pix-Generative Adversarial Network (SMPG) algorithm combining the unmixing-based method, deep learning method, pre-matching and post-matching module is proposed to reduce the spectral distortion of STDFA fusion image. The high-temporal-low-spatial (HTLS) resolution MOD09GA data and high-spatial-low-temporal resolution (HSLT) Landsat 8 data are selected in this study. SMPG algorithm is firstly employed to obtain daily high-spatial-high-temporal (HSHT) images, and then daily snow cover results with a spatial resolution of 30 m are obtained by calculating the normalized difference snow index (NDSI). SMPG algorithm is further compared with STDFA, Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), Flexible Spatiotemporal DAta Fusion (FSDAF), Swin SpatioTemporal Fusion Model (SwinSTFM), and Generative Adversarial Network-based SpatioTemporal Fusion Model (GAN-STFM). The experimental results indicate that the proposed algorithm yields better overall performance in daily spatiotemporal fusion image and snow cover result with a spatial resolution of 30 m. The mean correlation coefficient (CC) of SMPG can achieve 0.962, which is 0.06-0.36 higher than that of other spatiotemporal fusion methods. The error between the percentage of snow cover area obtained through SMPG and validation data is within 0.84%.

Accurate and timely crop mapping is essential for agricultural applications, and deep-learning methods have been applied on a range of remotely sensed data sources to classify crops. In this article, we develop a novel crop classification method based on spatiotemporal deep-learning fusion technology. However, for crop mapping, the selection and labeling of training samples is expensive and time consuming. Therefore, we propose a fully automated training-sample-selection method. First, we design the method according to image processing algorithms and the concept of a sliding window. Second, we develop the Geo-3D convolutional neural network (CNN) and Geo-Conv1D for crop classification using time-series Sentinel-2 imagery. Specifically, we integrate geographic information of crops into the structure of deep-learning networks. Finally, we apply an active learning strategy to integrate the classification advantages of Geo-3D CNN and Geo-Conv1D. Experiments conducted in Northeast China show that the proposed sampling method can reliably provide and label a large number of samples and achieve satisfactory results for different deep-learning networks. Based on the automatic selection and labeling of training samples, the crop classification method based on spatiotemporal deep-learning fusion technology can achieve the highest overall accuracy (OA) with approximately 92.50% as compared with Geo-Conv1D (91.89%) and Geo-3D CNN (91.27%) in the three study areas, indicating that the proposed method is effective and efficient in multi-temporal crop classification.