Search

Semi-supervised few-shot learning aims to improve the model generalization ability by means of both limited labeled data and widely-available unlabeled data. Previous works attempt to model the relations between the few-shot labeled data and extra unlabeled data, by performing a label propagation or pseudo-labeling process using an episodic training strategy. However, the feature distribution represented by the pseudo-labeled data itself is coarse-grained, meaning that there might be a large distribution gap between the pseudo-labeled data and the real query data. To this end, we propose a sample-centric feature generation (SFG) approach for semi-supervised few-shot image classification. Specifically, the few-shot labeled samples from different classes are initially trained to predict pseudo-labels for the potential unlabeled samples. Next, a semi-supervised meta-generator is utilized to produce derivative features centering around each pseudo-labeled sample, enriching the intra-class feature diversity. Meanwhile, the sample-centric generation constrains the generated features to be compact and close to the pseudo-labeled sample, ensuring the inter-class feature discriminability. Further, a reliability assessment (RA) metric is developed to weaken the influence of generated outliers on model learning. Extensive experiments validate the effectiveness of the proposed feature generation approach on challenging one- and few-shot image classification benchmarks. [ABSTRACT FROM AUTHOR]

In order to evaluate the potential impact of an aerosol lidar monitoring network in China on improving air quality prediction, an observing system simulation experiment (OSSE) was conducted using the Weather Research and Forecasting Chemistry (WRF-Chem) model and three-dimensional variational (3DVAR) data assimilation (DA) system. The impact of lidar numbers and locations on DA performance was evaluated using sensitivity analysis. Three ground-based lidar monitoring networks with 590, 310, and 99 lidars were constructed based on the distribution of existing surface air pollution monitoring stations in China to study the added value of lidar measurements. Four assimilation experiments were conducted using varying aerosol extinction coefficient data (AEXT) from three lidar monitoring networks and surface particulate matter (PM) concentration data over a four-week period. The results showed that compared with only assimilating surface PM 2.5 , the correlation coefficient (CORR) of PM 2.5 was significantly higher, the root mean square error (RMSE) and mean fraction error (MFE) were significantly lower, and the mean mass concentration was closer to the simulated observation after adding lidar profile measurements to the 3DVAR DA system. Additionally, the evaluation of DA with 590 stations indicated spatial differences in the improvement of the positive effect of assimilation. Overall, the improvement effects of assimilation in Central and East China were better than those in other regions. Finally, it was observed that the positive effect of DA on PM 2.5 prediction increased with an increase in the lidar number. However, with an increase in the number of lidars, the improvement of the positive effect of DA decreased gradually. The results indicate that it is necessary to balance the forecasting effect and economic cost while establishing a lidar observation network. • The potential impact of lidar observation network on 3D aerosol simulation in China is evaluated for the first time. • A more accurate 3D initial field can have a more positive impact on the prediction of surface PM2.5. • With an increase in the number of lidars, the improvement of the positive effect of DA decreased gradually. [ABSTRACT FROM AUTHOR]

Based on the three-dimensional variational assimilation (3DVAR) algorithm, the extinction coefficient profiles of the CALIPSO (Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations) satellite were assimilated for the first time on the basis of the aerosol variables from the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC) mechanism of the Weather Research and Forecasting-Chemistry (WRF-Chem) model. A light aerosol pollution episode (case1) and a heavy pollution event (case2) were selected to simulate and predict using the WRF-Chem model. Two sets of data were assimilated: CALIPSO derived aerosol extinction coefficient profiles and surface PM mass concentration data obtained from ground monitoring stations across China. Then, various aerosol data sources including ground-based lidar in Beijing area, Himawari-8 (H-8) satellite aerosol optical depth (AOD) and surface PM monitors were used for comprehensive evaluate the impact of CALIPSO data assimilation on the model performance. The results showed that simultaneous assimilation of the CALIPSO extinction coefficient profiles and surface PM mass concentration data can effectively improve the model aerosol initial conditions and forecasting accuracy of PM 2.5 , PM 10 and AOD. The positive impact on PM 2.5 and PM 10 forecasting by the assimilation can be maintained for more than 14 h for the case1 and 24 h for the case2, and the initial field of AOD in assimilation experiment has been greatly improved comparing the model control run (cn, without data assimilation). For the simulation of the case2, the assimilation of CALIPSO extinction coefficient profiles reduced the root mean square error (RMSE), mean fraction (MFE) and bias (BIAS) of the AOD in the initial field of the model by 0.08 (19.5%), 0.34 (55.7%) and 0.16 (48.5%) compared to the cn experiment, respectively. When the CALIPSO extinction coefficient profiles and surface PM mass concentration data were assimilated simultaneously, the reduction of the RMSE, MFE and BIAS of the AOD was 0.6 (14.6%), 0.32 (52.5%) and 0.12 (36.4%), respectively. Additionally, aerosol extinction coefficient are remarkably underestimated in the cn experiment, particularly during heavy aerosol pollution case. Assimilating surface PM data can improve the extinction coefficient distribution below 500 m. The simulated aerosol extinction coefficient distribution at 1 km to 2 km can be significantly improved in both cases by assimilating the CALIPSO extinction coefficient profiles, and the value of model simulated extinction coefficient is nearly doubled compared to the cn experiment. Finally, the evaluation with independent ground-based lidar measurements indicates that combined assimilation of surface PM mass concentration and CALIPSO extinction coefficient profile can effectively improve the aerosol distribution near the ground and in the upper layer, and thus generate a more accurate 3D aerosol analysis and forecast field. • The CALIPSO extinction coefficient is first assimilated for the MOSAIC aerosol scheme in the WRF-Chem model. • Assimilating CALIPSO profile data can effectively improve the three-dimensional structure of aerosol, especially for the aerosol vertical profile. [ABSTRACT FROM AUTHOR]

3D visual grounding is the task of accurately locating objects in a three-dimensional scene based on textual descriptions. Current approaches mainly depend on downsampling and extracting the point features for fusion with text features. However, the main challenges of these methods are the poor point feature resolution and limited local context during multi-modal fusion, causing visual-linguistic misalignment, particularly for small objects described in the text. The intuitive solution is to get additional object-related point features, gathering more contextual information to augment the representation capability of multimodal features, thereby promoting the representation capabilities of multimodal features. Based on this, we introduce a novel 3D visual grounding framework named Context-aware Feature Aggregation (CFA). The CFA framework includes two key modules: (1) Point Augmented Aggregation Module (PAM), designed to compensate for downsampling-induced information loss by augmenting sampled points with neighboring context for more discriminative features; and (2) Dual Contextual Grouping Attention Module (DCGAM), which iteratively refines features and geometry coordinates from PAM, capturing more global context. We assess the performance of our CFA framework on two point-based datasets: ScanRefer and Nr3D/Sr3D. The CFA framework exhibits efficiency in 3D visual grounding, surpassing the performance of previous methods by experimental results. [ABSTRACT FROM AUTHOR]

Based on the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC) aerosol scheme of the Weather Research and Forecasting model coupled with online Chemistry (WRF-Chem) and the three-dimensional variational (3DVAR) assimilation method, a 3DVAR data assimilation (DA) system for aerosol optical depth (AOD) and aerosol concentration observations was developed. A case study on assimilating the Himawari-8 satellite AOD and/or fine particulate matter (PM2.5) was conducted to investigate the improvement of DA on the analysis accuracy and forecast skills of the spatial distribution characteristics of aerosols, especially in the vertical dimension. The aerosol extinction coefficient (AEC) profile data from The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO), surface PM2.5 and Himawari-8 AOD measurements were used for verification. One control experiment (without DA) and two DA experiments including a PM2.5 DA experiment denoted by Da_PM and a combined PM2.5 and AOD DA experiment denoted by Da_AOD_PM were conducted. Both DA experiments had positive effects on the surface PM2.5 mass concentration forecast skills for more than 60 h. However, the Da_PM showed a slight improvement in the analysis accuracy of the AOD distribution compared with the control experiment, while the Da_AOD_PM showed a considerable improvement. The Da_AOD_PM had the best positive effect on the AOD forecast skills. The correlation coefficient (CORR), root mean square error (RMSE), and mean fraction error (MFE) of the 24 h AOD forecasts for the Da_AOD_PM were 0.73, 0.38, and 0.54, which are 0.09 (14.06%), 0.08 (17.39%), and 0.22 (28.95%) better than that of the control experiment, and 0.05 (7.35%), 0.06 (13.64%), and 0.19 (26.03%) better than that of the Da_PM, respectively. Moreover, improved performance for the Da_AOD_PM occurred when the AEC profile was used for verification, as when the AOD was used for verification. The Da_AOD_PM successfully simulated the first increasing and then decreasing trend of the aerosol extinction coefficients below 1 km, while neither the control nor the Da_PM did. This indicates that assimilating AOD can effectively improve the analyses and forecast accuracy of the aerosol structure in both the horizontal and vertical dimensions, thereby compensating for the limitations associated with assimilating traditional surface aerosol observations alone. [ABSTRACT FROM AUTHOR]