Generalized category discovery (GCD) aims to group unlabeled samples from known and unknown classes when only part of the labeled data in the known classes is given. It allows the model to adapt to dynamic environments by discovering novel categories. However, when we applied the GCD approach to the decentralized open world, we still encountered the following challenges: (1) none of labeled data easily obtained in the open world, (2) heterogeneous label spaces across different environments, (3)representation degradation caused by fine-tuning models with limited data in specific environments. To address the above challenges, we introduce a new and practical task, namely Cloud-edge GCD (CE-GCD). Different from semi-supervised GCD, CE-GCD assumes that we only have a base model trained on common public categories, and aims to perform personalized unsupervised novel category discovery in multiple environments with heterogeneous label spaces. Data from different environments or clients cannot be shared, only model parameters can be transferred. To tackle this problem, we propose a novel GCD framework based on energy-guided known class discrimination and multi-level contrastive learning. In each client, we first use the classifier of the base model to distinguish between known and unknown classes, and then perform unsupervised learning on the unknown classes. Each client transfers category information through prototypes to assist learning. Extensive experiments on multiple datasets demonstrate the effectiveness of our approach.

The accurate prediction of sea surface temperature (SST) is highly important for climate change research and the management of marine ecosystems. Traditional numerical models rely on complex physical processes and precise initial conditions, resulting in high computational costs and limited generalizability. Although deep learning methods can improve the physical consistency and interpretability of ocean processes by incorporating physical constraints, their performance may degrade under anomalous or extreme SST conditions, where rigid constraints limit the model’s adaptability to complex variations. In contrast, incorporating motion-aware mechanisms enables models to flexibly capture dynamic patterns from data, thus increasing their responsiveness to nonstationary processes. Therefore, we propose a novel SST prediction model that integrates optical flow guidance with spatiotemporal awareness, aiming to enhance the modeling of motion and spatiotemporal features in SST evolution. The proposed model consists of three key modulesa spatiotemporal information extraction module (SIEM), a motion trend extraction module (MTEM), and a spatiotemporal feature fusion module (SFFM). First, the SIEM, composed of multiple layers of SwinLSTM, captures spatial and temporal dependencies in the SST time series. The MTEM then estimates the optical flow to extract motion information from the SST data. Finally, the motion information is used to dynamically adjust the spatiotemporal features, which are then fused in the SFFM. To evaluate the performance of the proposed model, we conducted SST prediction experiments at both daily (1–10 days) and weekly (1–10 weeks) scales over the South China Sea (SCS) and the Pacific Ocean, and compared the results with several existing models. The experimental results demonstrate that our model outperforms existing methods across multiple evaluation metrics, with superior prediction accuracy and robustness.

Deploying deep learning-based remote sensing image interpretation models in orbit can alleviate data downlink pressure and enhance processing efficiency. To adapt to the arithmetic conditions of on-orbit platforms, deep learning models must undergo compilation before deployment, where the search and measurement of operator scheduling are critical aspects. The existing compilation methods related to the Tensor Virtual Machine (TVM) compiler suffer from inefficiencies caused by suboptimal starting point selection and extensive search spaces during the search process. Additionally, the clustering algorithm employed during the measurement of candidate scheduling relies on fixed dimensions and lacks flexible application management, resulting in slow convergence. This paper proposes corresponding optimization methods built upon the TVM framework to address these two challenges. In the search aspect, techniques such as starting point planning, salient reduction, and model sharing considerably reduce search time. In the measurement aspect, the clustering process is optimized through adaptive scheduling management and dimension adaptation, enhancing measurement efficiency. Experimental results indicate that, compared to current state-of-the-art deep learning compilers, the proposed method increases compilation speed by an average of 2.6 times while maintaining consistent inference speed.

Research about SAR target recognition has received a lot of attention in recent years. The relevant research results encountered some obstacles in their practical application. This is partly due to the long-tailed distribution of data in real-life scenarios. Specifically, a majority of data samples are concentrated in a few categories. The skewed distribution can cause learning bias toward the majority class. Although there have been some studies on long-tailed recognition for natural images, we found that these methods still need improvement when applied to SAR scenes. On the one hand, these studies use balanced datasets for performance testing, which is incompatible with imbalanced SAR target recognition. On the other hand, the predictions of these methods are unreliable for identifying high-value tail class SAR targets that are sensitive to faults. To address these issues, We propose a SAR long-tailed target recognition method based on evidence learning. The evidence learning head can output uncertainty as a measure of predictive confidence. On this basis, we demonstrate the use of uncertainty to further optimize the long-tailed SAR recognition performance and achieve optimal performance when the predicted distribution is unknown.

SAR ship detection is an important technology supporting water traffic monitoring and marine safety maintenance. In recent years, many methods based on deep neural networks have been used to improve the performance of SAR ship detection. These methods mainly focus on two issues: one is the false alarm of ship detection in complex inshore environments, and the other is the effective extraction and utilization of SAR ship features. The topic discussed in this paper is one of the culprits that has caused the aforementioned two problems, but has long been overlooked. Specifically, it pertains to the issue of intra-class imbalance in SAR ship detection. There are imbalances in the size distribution, azimuth distribution, and background distribution under the real data collection environment. However, since SAR ship detection is a single-class detection task, the aforementioned imbalances lack reliable descriptors during training. This paper proposes using evidence learning to obtain the epistemic uncertainty as a descriptor of biased learning on samples. Contrastive learning is used to further utilize the uncertainty label of samples to correct biased learning under intra-class imbalance. The proposed method is proven to be effective on multiple network models. AP50 reaches 94.8% on the HRSID dataset, 98.4% on SSDD dataset and 80.9% on the LS-SSDD dataset, both achieving SOTA performance.

In recent years, it has been a research hotspot to apply big data-driven deep learning methods to Synthetic Aperture Radar (SAR) target recognition with limited data. However, the problem caused by the long-tailed characteristics of SAR data has long been ignored. Specifically, a majority of data samples are concentrated in a few categories, leading to a skewed distribution of data. This skewed distribution can cause learning bias towards the majority class, which can subsequently degrade the recognition performance of the minority class. This issue is further exacerbated in limited sample conditions for SAR target recognition. After conducting research on target recognition for long-tailed natural images, this study has found that the existing methods used in this field cannot be easily applied to SAR target recognition. The primary reason is that SAR image data exhibit simultaneous and complex inter-class and intra-class long-tailed distributions. In response to this issue, we proposes the use of a multi-branch expert network and dual-environment sampling to address the long-tail problems in both inter-class and intra-class scenarios. The proposed method outperforms popular long-tailed target recognition methods on the long-tailed versions of the MSTAR and FUSAR datasets.

Simultaneous multitask processing is a common requirement in synthetic aperture radar (SAR) automatic target recognition (ATR), e.g., not only the category of the target but also the aspect angle of the target need to be identified at the same time. Moreover, the target recognition network is always expected to have the capability of incremental learning, i.e., acquire the processing capabilities for new tasks while maintaining the processing capabilities for old tasks. In this paper, an incremental multitask learning method based on structured pruning is proposed. The structured pruning, originally proposed for network compression, is used to learn with dominant neuron and release parameter space of convolutional neural network for new tasks. Through iterative pruning and training of new tasks, multitask target recognition is realized in a single convolutional neural network and could simultaneously output recognition results of multiple tasks. The experiments on the MSTAR dataset show that our method can simultaneously recognize the category and aspect angle of target, while does not decrease the corresponding accuracy compared to single-task processing.