As a fundamental and challenging task in bridging language and vision domains, Image-Text Retrieval (ITR) aims at searching for the target instances that are semantically relevant to the given query from the other modality, and its key challenge is to measure the semantic similarity across different modalities. Although significant progress has been achieved, existing approaches typically suffer from two major limitations: (1) It hurts the accuracy of the representation by directly exploiting the bottom-up attention based region-level features where each region is equally treated. (2) It limits the scale of negative sample pairs by employing the mini-batch based end-to-end training mechanism. To address these limitations, we propose a Unified Semantic Enhancement Momentum Contrastive Learning (USER) method for ITR. Specifically, we delicately design two simple but effective Global representation based Semantic Enhancement (GSE) modules. One learns the global representation via the self-attention algorithm, noted as Self-Guided Enhancement (SGE) module. The other module benefits from the pre-trained CLIP module, which provides a novel scheme to exploit and transfer the knowledge from an off-the-shelf model, noted as CLIP-Guided Enhancement (CGE) module. Moreover, we incorporate the training mechanism of MoCo into ITR, in which two dynamic queues are employed to enrich and enlarge the scale of negative sample pairs. Meanwhile, a Unified Training Objective (UTO) is developed to learn from mini-batch based and dynamic queue based samples. Extensive experiments on the benchmark MSCOCO and Flickr30K datasets demonstrate the superiority of both retrieval accuracy and inference efficiency. For instance, compared with the existing best method NAAF, the metric R@1 of our USER on the MSCOCO 5K Testing set is improved by 5% and 2.4% on caption retrieval and image retrieval without any external knowledge or pre-trained model while enjoying over 60 times faster inference speed. Our source code will be released at https://github.com/zhangy0822/USER .

The precise location of the tropical cyclone (TC) center is critical for intensity estimation and trajectory prediction. Due to the variability of TC morphology and structure, there are still some challenges in locating its center automatically. The ability of the deep convolutional network to capture multilevel structural features of the images is exploited. Furthermore, a two-step scheme for locating the TC center is proposed, which contains the object detection for TCs with deep learning and the comprehensive decision for TC centers. In the object detection, considering the statistical scale distribution of TCs, the global and local features extracted by the network are combined to form the fusion feature maps through the upsampling and concatenation. The changes in the TC scale are accommodated by two different scale outputs. A high detection rate and a low false alarm rate are obtained with the object detection, which provides an initial position for the TC center. Within the scope of the TCs, the final position of the center is obtained through segmentation, edge detection, circle fitting, and comprehensive decision. The experimental results show that the average latitude and longitude error of the proposed method is about 0.237$^\circ$. For the TC in the initial phase or dissipation stage, the location results are usually superior to the results of the comparison algorithms.

Radar echo extrapolation is a basic but essential task in meteorological services. It could provide radar echo prediction results with high spatiotemporal resolution in a computationally efficient way, and effectively enhance the operational system's forecasting capability for meteorological hazards. Traditional methods perform extrapolation by estimating echo motions between contiguous radar data. This strategy is difficult to characterize complex nonlinear meteorological processes effectively, and it is difficult to benefit from large historical data. Recently, machine learning (ML) models have been used for radar echo extrapolation. These methods have effectively improved extrapolation quality in a data-driven way and from the statistical perspective. Although the ML-based methods show excellent performance, they usually produce blurry extrapolations. This leads to underestimating radar echo intensity and making echo lack small-scale details. Moreover, it makes models difficult to predict severe convective hazards. To solve this problem, a two-stage extrapolation model based on 3-D convolutional neural network and conditional generative adversarial network is proposed. These two models form the “pre-extrapolation” and “postprocessing” paradigm. The pre-extrapolation model is trained in the traditional way and performs rough extrapolation. The postprocessing model uses the pre-extrapolation result as input and is trained with the adversarial strategy. It could correct the echo intensity and increase the echo's details. In the experiment, our model could provide more precise radar echo extrapolations than other methods, especially for intense echoes and convective systems, in the data of North China from 2015 to 2016.

Remote sensing image captioning (RSIC) has great significance in image understanding, which describes the image content in natural language. Existing methods are mainly based on deep learning and rely on the encoder–decoder model to generate sentences. In the decoding process, recurrent neural network (RNN) and long short-term memory (LSTM) are normally applied to sequentially generate image captions. In this letter, the transformer encoder–decoder is combined with grid features to improve the RSIC performance. First, the pretrained convolutional neural network (CNN) is used to extract grid-based visual features, which are encoded as vectorial representations. Then, the transformer outputs semantic descriptions to bridge visual features and natural language. Besides, the self-critical sequence training (SCST) strategy is applied to further optimize the image captioning model and improve the quality of generated sentences. Extensive experiments are organized on three public datasets of RSCID, UCM-Captions, and Sydney-Captions. Experimental results demonstrate the effectiveness of SCST strategy and the proposed method achieves superior performance compared with the state-of-the-art image captioning approaches on the RSCID dataset.

In weather forecasting, numerical weather prediction (NWP) that is based on physical models requires proper post-processing before it can be applied to actual operations. Therefore, research on intelligent post-processing algorithms has always been an important topic in this field. This paper proposes a model output deep learning (MODL) method for post-processing, which can improve the forecast effect of numerical weather prediction. MODL is an end-to-end post-processing method based on deep convolutional neural network, which directly learns the mapping relationship between the forecast fields output by numerical model and the observation temperature field in order to obtain more accurate temperature forecasts. MODL modifies the existing deep convolution model according to the post-processing problem’s characteristics, thereby improving the performance of the weather forecast. This paper uses The International Grand Global Ensemble (TIGGE) dataset from European Centre for Medium-Range Weather Forecasts (ECMWF) and the observed air temperature of 2 m obtained from Tianjin meteorological station in order to test the post-processing performance of MODL. The MODL method applied to temperature in post-processing is compared with the ECMWF forecast, Model Output Statistics (MOS) methods, and Model Output Machine Learning (MOML) methods. The Root Mean Square Error (RMSE) of the temperature field predicted by MODL and the observed temperature field is smaller than the other models and the accuracy of the temperature difference of 2 °C (Acc) is higher, especially where the prediction time is in the first three days. The lightweight nature of MODL also makes it suitable for most operations.

Infrared small target detection (IRSTD) involves accurately locating the positions of small and dim targets in infrared images and serves as a fundamental component in modern sensing and measurement systems, such as maritime vessel search and airborne monitoring. In the spatial domain, background clutter and sensor-induced noise exhibit similar visual characteristics to true targets, leading to false alarms and missed detections. Existing detection methods primarily focus on extracting local spatial features, while overlooking the differences between targets and background in the frequency domain. To address this limitation, we propose a Frequency-Aware Attention Network (FAA-Net) for IRSTD. Specifically, we design a Frequency-Aware Attention Module (FAM) based on discrete cosine transform to capture informative high-frequency components and enhance local contrast. Furthermore, we construct a Transformer and Convolution Fusion Block (TCFB) to combine the global modeling ability of Transformers with the local feature extraction strength of convolutional layers. By applying FAMs and TCFBs to both shallow and deep layers, FAA-Net captures multi-scale semantic and structural features to generate robust target representations. From a measurement perspective, FAA-Net provides reliable pixel-level outputs that correspond to measurable physical quantities, such as target location and area. We conduct experiments on three publicly available datasets, including NUDT-SIRST, NUAA-SIRST, and IRSTD-1K. Experimental results demonstrate that our presented FAA-Net achieves superior performance compared to existing deep learning-based methods in IRSTD. The source code is available at https://github.com/Dr-Zhuang/FAA-Net-IRSTD.