The active recognition of interesting targets has been a vital issue for remote sensing. In this paper, a novel multi-source fusion method for ship target detection and recognition is proposed. By introducing synthetic aperture radar (SAR) sensor images, the proposed method solves the problem of precision degradation in optical remote sensing image target detection and recognition caused by the limit of illumination and weather conditions. The proposed method obtains port slice images containing ship targets by fusing optical data with SAR data. On this basis, spectral residual saliency and region growth method are used to detect ship targets in optical image, while SAR data are introduced to improve the accuracy of ship detection based on joint shape analysis and multi-feature classification. Finally, feature point matching, contour extraction and brightness saliency are used to detect the ship parts, and the ship target types are identified according to the voting results of part information. The proposed ship detection method obtained 91.43% recognition accuracy. The results showed that this paper provides an effective and efficient ship target detection and recognition method based on multi-source remote sensing images fusion.

Most of the existing saliency detection methods are affected by the complex background and weak contrast in remote sensing images (RSIs), which easily leads to confusion between salient object and background. To solve this problem, we present a general space–frequency joint saliency detection method based on spatial contrast analysis (SCA) and adaptive spectrum analysis (ASA). The proposed SCA constructs a spatial distribution function to give higher saliency to bright elements with similar appearance and compact distribution, which can avoid the interference of similar elements in the background. In addition, for the weak contrast-level salient object, an adaptive spectral energy function is proposed in ASA by combining regional dispersion and frequency spectrum characteristic, which can make full use of frequency spectrum and obtains the complete information of salient object. Finally, the SCA and ASA information are fused by a joint optimization module, which is proposed to integrate the saliency of different contrast levels objects; thus, the final saliency maps are obtained. Evaluated on the extended optical remote sensing saliency detection (EORSSD) dataset and our dataset, the proposed method is superior to other five classical methods in RSI saliency detection.

Fine-grained object detection in remote sensing images is highly challenging due to class imbalance and high inter-class indistinguishability. The strategies employed by most existing methods to resolve these two challenges are relatively rudimentary, resulting in suboptimal model performance. To address these issues, we propose a fine-grained object-oriented detection method based on relevance pooling guidance and class balance feature enhancement. Firstly, we propose a global attention mechanism that dynamically retains spatial features pertinent to objects through relevance pooling during down-sampling, thereby enabling the model to acquire more discriminative features. Next, a class balance correction module is proposed to alleviate the class imbalance problem. This module employs feature translation and a learnable reinforcement coefficient to highlight the boundaries of tail class features while maintaining their distinctiveness. Furthermore, we present an enhanced contrastive learning strategy. By dynamically adjusting the contribution of inter-class samples and intra-class similarity measures, this strategy not only constrains inter-class feature distances but also facilitates tighter intra-class clustering, making it more suitable for imbalanced datasets. Evaluation on the FAIR1M and MAR20 datasets demonstrates that our method is superior compared to other methods in object detection and achieves 46.44 and 85.05% mean average precision, respectively.

Existing thin cloud removal methods primarily rely on generative paradigms or discriminative paradigms. Generative paradigms often suffer from training instability, while discriminative paradigms exhibit insufficient feature representation, and their loss strategies lack physical consistency, resulting in suboptimal performance. To address these issues, we propose a dynamic bidirectional feature enhancement network for thin cloud removal in optical remote sensing images. Firstly, we design a multidimensional attention module comprising skip-recursive dilated convolution-based spatial attention module and frequency-domain-driven channel attention module. These modules efficiently capture long-range contextual dependencies and suppress cloud noise interference. Next, an adaptive local feature enhancement block is constructed using cross-fusion and adjacent feature propagation between dynamic convolutions, aimed at enhancing the ability of model to recover details. Then, we present a dynamic enhancement-based bidirectional information flow module to model the dynamic interaction between multi-task features, guiding detail recovery and feedback for optimized cloud removal features. Finally, we design a physics-aware joint loss function incorporating atmospheric light consistency constraints to ensure the physical authenticity of cloud-free images. Evaluation on the RICE and WHUS2-CR datasets demonstrates that the proposed method is superior to compared methods in thin cloud removal and can improve the performance of baseline.

Semi-supervised object detection (SSOD) in remote sensing is challenged by the accumulation of pseudo-label errors in complex scenes with dense objects and high intra-class variability. While teacher–student frameworks enable learning from unlabeled data, erroneous pseudo-labels such as false positives and missed detections can be reinforced over time, which degrades model performance. To address this issue, we propose the Error-Mitigation Teacher (EMT), a unified framework designed to suppress error propagation during SSOD training. EMT consists of three lightweight modules. First, the Adaptive Pseudo-Label Filtering (APLF) module removes noisy pseudo boxes via a second-stage RCNN and adjusts class-specific thresholds through dynamic confidence filtering. Second, the Confidence-Based Loss Reweighting (CBLR) module reweights training loss by evaluating the teacher model’s ability to reconstruct its own pseudo-labels, using the resulting loss as an indicator of label reliability. Third, the Enhanced Supervised Learning (ESL) module improves class-level balance by adjusting supervised loss weights according to pseudo-label statistics. EMT demonstrates consistent performance gains over representative state-of-the-art SSOD methods on DOTA, DIOR, and SSDD datasets. Notably, EMT achieves a 2.9% absolute mAP₅₀ improvement on DIOR using only 10% of labeled data, without incurring additional inference cost. These results highlight EMT’s effectiveness in improving SSOD for remote sensing.

The majority of existing object detection models struggle to fully exploit the intimate relationship between scene context and objects, and the feature fusion and proposal generation strategies tend to be relatively basic, resulting in poor model performance. To address these issues, we propose an object detection model based on adaptive multilevel fusion refinement network. Firstly, we propose an adaptive gated fusion network that dynamically assesses correlations between objects and scene information, generating a gated feature map to guide feature fusion and extract discriminative joint object-context features. Next, a proposal refinement model is proposed. By utilizing a learnable correlation-weighted coefficient, this model effectively merges low-level features with joint features, thereby mitigating spatial information deficits. We also propose an adaptive multidimensional offset strategy, which minimizes the impact of regression deviations on proposal quality by combining information offsets and spatial offsets. To optimize all subtasks, a novel multi-task loss function is proposed. Evaluated on the DOTA and HRSC2016 datasets shows that our method is superior to compared methods in object detection and harvests 77.26% and 90.68% mean average precision, respectively.