Recently, state-of-the-art classification performance of natural images has been obtained by self-supervised learning (S2L) as it can generate latent features through learning between different views of the same images. However, the latent semantic information of similar images has hardly been exploited by these S2L-based methods. Consequently, to explore the potential of S2L between similar samples in hyperspectral image classification (HSIC), we propose the nearest neighboring self-supervised learning (N2SSL) method, by interacting between different augmentations of reliable nearest neighboring pairs (RN2Ps) of HSI samples in the framework of bootstrap your own latent (BYOL). Specifically, there are four main steps: pretraining of spectral spatial residual network (SSRN)-based BYOL, generation of nearest neighboring pairs (N2Ps), training of BYOL based on RN2P, final classification. Experimental results of three benchmark HSIs validated that S2L on similar samples can facilitate subsequent classification. Moreover, we found that BYOL trained on an un-related HSI can be fine-tuned for classification of other HSIs with less computational cost and higher accuracy than training from scratch. Beyond the methodology, we present a comprehensive review of HSI-related data augmentation (DA), which is meaningful to future research of S2L on HSIs.

Compared with general object detection with horizontal bounding boxes in natural images, oriented object detection in remote sensing images is an active and challenging research topic as objects are usually displayed in arbitrary orientations. To model the variant orientations of oriented objects, general CNN-based methods usually adopt more parameters or well-designed modules, which are often complex and inefficient. To address this issue, the detector requires two key components to deal with: (i) generating oriented proposals in a light-weight network to achieve effective representation of arbitrarily oriented objects; (ii) extracting the rotation-invariant feature map in both spatial and orientation dimensions. In this paper, we propose a novel, lightweight rotated region proposal network to produce arbitrary-oriented proposals by sliding two vertexes only on adjacent sides and adopt a simple yet effective representation to describe oriented objects. This may decrease the complexity of modeling orientation information. Meanwhile, we adopt the rotation-equivariant backbone to generate the feature map with explicit orientation channel information and utilize the spatial and orientation modules to obtain completely rotation-invariant features in both dimensions. Without tricks, extensive experiments performed on three challenging datasets DOTA-v1.0, DOTA-v1.5 and HRSC2016 demonstrate that our proposed method can reach state-of-the-art accuracy while reducing the model size by 40% in comparison with the previous best method.

The scale parameter (SP), to control the sizes of objects, is of great significance in multiscale segmentation which is a prerequisite and foundational step for object-based change detection (OBCD). However, the appropriate SP is not readily apparent and the majority of the existing OBCD algorithms obtain the SPs by empirical or subjective trial-and-error ways that may lead to dissatisfactory accuracy and be time-consuming. To address this issue, an automatic approach for optimal segmentation scale selection for OBCD is proposed in this letter. First, a changed fuzziness image for bitemporal images was generated. Second, multiscale segmentation was implemented in a series of candidate scales, and the merging relationships between adjacent scales were built. Then, mapping the segments to the previous fuzziness image, a statistical metric describing the homogeneity of objects based on the Kullback-Leibler divergence was defined, the increments of the metric between merged objects and their child objects were calculated and weighted to identify the optimal SPs. For performance evaluation, Dempter-Shafer (DS) evidence fusion was utilized in the scales selected by the proposed approach in comparison with other state of art or empirical ones. The experimental results employing GF-1, Google Earth, and aerial images demonstrated the superiority and effectiveness of the SPs identified by the proposed approach in the OBCD task.

The reconstruction of road networks from high-resolution satellite images is of significant importance across a range of disciplines, including traffic management, vehicle navigation and urban planning. However, existing models are computationally demanding and memory-intensive due to their high model complexity, rendering them impractical in many real-world applications. In this work, we present Cascaded Efficient Road Network (CE-RoadNet), a novel neural network architecture which emphasizes the elegance and simplicity of its design, while also retaining a noteworthy level of performance in road extraction tasks. First, a simple encoder–decoder architecture (Effi-RoadNet) is proposed, which leverages smoothed dilated convolutions combined with an attention-guided feature fusion module to aggregate features from multiple levels. Subsequently, an extended variant termed CE-RoadNet is designed in a cascaded architecture to enhance the feature representation ability of the model. Benefiting from the concise network design and the prominent representational ability of the stacking mechanism, our network can accomplish better trade-offs between accuracy and efficiency. Extensive experiments on public road datasets demonstrate that our approach achieves state-of-the-art results with lower complexity. All codes and models will be released soon to facilitate reproduction of our results.

Optical and SAR image registration is the primary procedure to exploit the complementary information from the two different image modal types. Although extensive research has been conducted to narrow down the vast radiometric and geometric gaps so as to extract homogeneous characters for feature point matching, few works have considered the registration issue for non-flat terrains, which will bring in more difficulties for not only sparse feature point matching but also outlier removal and geometric relationship estimation. This article addresses these issues with a novel and effective optical-SAR image registration framework. Firstly, sparse feature points are detected based on the phase congruency moment map of the textureless SAR image (SAR-PC-Moment), which helps to identify salient local regions. Then a template matching process using very large local image patches is conducted, which increases the matching accuracy by a significant margin. Secondly, a mutual verification-based initial outlier removal method is proposed, which takes advantage of the different mechanisms of sparse and dense matching and requires no geometric consistency assumption within the inliers. These two procedures will produce a putative correspondence feature point (CP) set with a low outlier ratio and high reliability. In the third step, the putative CPs are used to segment the large input image of non-flat terrain into dozens of locally flat areas using a recursive random sample consensus (RANSAC) method, with each locally flat area co-registered using an affine transformation. As for the mountainous areas with sharp elevation variations, anchor CPs are first identified, and then optical flow-based pixelwise dense matching is conducted. In the experimental section, ablation studies using four precisely co-registered optical-SAR image pairs of flat terrain quantitatively verify the effectiveness of the proposed SAR-PC-Moment-based feature point detector, big template matching strategy, and mutual verification-based outlier removal method. Registration results on four 1 m-resolution non-flat image pairs prove that the proposed framework is able to produce robust and quite accurate registration results.

The remote sensing scene classification (RSSC) task plays a pivotal role in Earth observation missions, yet its progress remains constrained by the scarcity of high-quality labeled imagery. This article introduces a self-supervised learning (SSL) paradigm to address this challenge. First, for pseudo-label construction, a large set of long-interval satellite revisit imagery is collected and processed with pixel-level registration. The SIFT inliers retained during registration serve as saliency priors to guide asymmetric masking across views. This produces positive pairs that preserve global scene consistency while introducing controlled object-level ambiguities. Second, we propose a progressive layer-wise contrastive learning framework (MTC-Net) that couples the pseudo-label with the network’s representational hierarchy, forming a curriculum from local texture robustness to global semantic invariance. A dual-attention module with spatial–channel branches is further embedded to recalibrate intermediate features. The learning paradigm encourages the model to perform cross-view contextual reasoning rather than relying on pixel-wise correspondences. Experiments on three widely used datasets demonstrate that MTC-Net achieves competitive classification accuracy under limited-label settings, while ablation and visualization studies validate the effectiveness of establishing scene-level invariance through multi-temporal contrastive alignment.