AF‐UNet: Intelligent Recognition Model for Atmospheric Fronts Based on Adaptive Fuser

The critical success index (CSI) is an established metric used in meteorology to verify the accuracy of weather forecasts. It is defined as the ratio of hits to the sum of hits, false alarms, and misses. Translationally, CSI has gained popularity as a unitary outcome measure in various clinical situations where large numbers of true negatives may influence the interpretation of other, more traditional, outcome measures, such as specificity (Spec) and negative predictive value (NPV), or when unified interpretation of positive predictive value (PPV) and sensitivity (Sens) is needed. The derivation of CSI from measures including PPV has prompted questions as to whether and how CSI values may vary with disease prevalence (P), just as PPV estimates are dependent on P, and hence whether CSI values are generalizable between studies with differing prevalences. As no detailed study of the relation of CSI to prevalence has been undertaken hitherto, the dataset of a previously published test accuracy study of a cognitive screening instrument was interrogated to address this question. Three different methods were used to examine the change in CSI across a range of prevalences, using both the Bayes formula and equations directly relating CSI to Sens, PPV, P, and the test threshold (Q). These approaches showed that, as expected, CSI does vary with prevalence, but the dependence differs according to the method of calculation that is adopted. Bayesian rescaling of both Sens and PPV generates a concave curve, suggesting that CSI will be maximal at a particular prevalence, which may vary according to the particular dataset.

BackgroundMethods to undertake diagnostic accuracy studies of administrative epilepsy data are challenged by lack of a way to reliably rank case-ascertainment algorithms in order of their accuracy. This is because it is difficult to know how to prioritise positive predictive value (PPV) and sensitivity (Sens). Large numbers of true negative (TN) instances frequently found in epilepsy studies make it difficult to discriminate algorithm accuracy on the basis of negative predictive value (NPV) and specificity (Spec) as these become inflated (usually >90%). This study demonstrates the complementary value of using weather forecasting or machine learning metrics critical success index (CSI) or F measure, respectively, as unitary metrics combining PPV and sensitivity. We reanalyse data published in a diagnostic accuracy study of administrative epilepsy mortality data in Scotland. MethodCSI was calculated as 1/[(1/PPV) + (1/Sens) – 1]. F measure was calculated as 2.PPV.Sens/(PPV + Sens). CSI and F values range from 0 to 1, interpreted as 0 = inaccurate prediction and 1 = perfect accuracy. The published algorithms were reanalysed using these and their accuracy re-ranked according to CSI in order to allow comparison to the original rankings. ResultsCSI scores were conservative (range 0.02–0.826), always less than or equal to the lower of the corresponding PPV (range 39%–100%) and sensitivity (range 2%–93%). F values were less conservative (range 0.039–0.905), sometimes higher than either PPV or sensitivity, but were always higher than CSI. Low CSI and F values occurred when there was a large difference between PPV and sensitivity, e.g. CSI was 0.02 and F was 0.039 in an instance when PPV was 100% and sensitivity was 2%. Algorithms with both high PPV and sensitivity performed best in terms of CSI and F measure, e.g. CSI was 0.826 and F was 0.905 in an instance when PPV was 90% and sensitivity was 91%. ConclusionCSI or F measure can combine PPV and sensitivity values into a convenient single metric that is easier to interpret and rank in terms of diagnostic accuracy than trying to rank diagnostic accuracy according to the two measures themselves. CSI or F prioritise instances where both PPV and sensitivity are high over instances where there are large differences between PPV and sensitivity (even if one of these is very high), allowing diagnostic accuracy thresholds based on combined PPV and sensitivity to be determined. Therefore, CSI or F measures may be helpful complementary metrics to report alongside PPV and sensitivity in diagnostic accuracy studies of administrative epilepsy data.

Precipitation nowcasting, which entails high-resolution forecasting of precipitation events within 1–2 hours, is significant to daily life and professional activities. Nevertheless, accurate short-term precipitation forecasting remains a considerable challenge at present. Traditional numerical weather prediction, which relies on intricate physical equations to simulate the Earth's atmospheric state, necessitates substantial computational resources and frequently yields lower accuracy for small-scale forecasts, thereby failing to meet the demands of precipitation prediction in complex regions. Most deep learning methodologies concentrate exclusively on the spatiotemporal prediction of a singular precipitation variable, thereby neglecting the dynamic spatiotemporal relationships between precipitation and other meteorological data within the meteorological system. Moreover, due to the rapid pace of climate change, long-term time series data is often inadequate for accurately addressing precipitation forecasting for extreme weather events, since past meteorological time series data may not accurately reflect the current atmospheric conditions. There is an urgent need to rely on short-term time series for prediction tasks. However, most current methods that rely on short-term time series for prediction perform poorly in forecasting moderate to heavy precipitation events. Inspired by spatiotemporal information transformation schemes, we introduce a spatiotemporal information(STI) transformation equation from chaotic dynamics into the field of computer vision and develop a neural network model framework based on spatiotemporal information transformation. This framework maps high-dimensional spatial information to the temporal information of future precipitation information, thereby facilitating the integration of dynamic spatiotemporal relationships between various meteorological data and precipitation, and enabling the mutual transformation of spatiotemporal information for enhanced forecasting accuracy. Furthermore, we propose an adaptive gradient loss function designed to improve the model's sensitivity to learning moderate-intensity precipitation. This research utilizes the US SEVIR dataset for training and testing, which encompasses data such as satellite visible light, infrared temperature, humidity, and cloud precipitation while employing multiple meteorological data for precipitation forecasting over the subsequent hour. We selected the Structural Similarity Index, Peak Signal-to-Noise Ratio, False Alarm Rate, Critical Success Index, and Heidke Skill Score as both quantitative and qualitative evaluation metrics. Experimental results demonstrate that the STI framework reduces the model's error in moderate to heavy precipitation events, making the model more sensitive to severe rainfall events. Furthermore, when the STI framework is integrated into other deep learning models and retrained, it further enhances their precipitation prediction accuracy. This finding indicates that the STI framework effectively captures the dynamic spatiotemporal relationships between various meteorological and precipitation data.

이 연구에서 우리는 코로나 홀(Coronal hole, CH)의 정보(위치, 면적)를 이용하여 CIR(Corotating Interaction Regions)과 지자기폭퐁(Geomagnetic Storm)에 대한 경험적인 예보를 수행하였다. 이것을 위해 1996년 1월 <TEX>$\sim$</TEX> 2003년 11월까지의 미국 국립 천문대-Kitt Peak 관측소의 He I <TEX>$1083{\AA}$</TEX> 영상으로부터 코로나 홀 자료를 얻고, Choi et al.(2009)로부터 확인된 CIR과 지자기폭풍 자료를 활용하였다. 지자기 폭풍을 일으키는 코로나 홀의 특성을 고려하여 코로나 홀의 중심이 <TEX>$N40^{\circ}$</TEX>와 <TEX>$S40^{\circ}$</TEX> 사이, <TEX>$E40^{\circ}$</TEX>와 <TEX>$W20^{\circ}$</TEX> 사이에 위치하고 태양 반구에 대한 면적 비율이 다음과 같은 세 가지 경우를 선택하였다: (1) case 1: 0.36% 이상, (2) case 2: 0.66% 이상, (3) case 3: <TEX>$1996{\sim}2000$</TEX>년 동안에는 0.36%, <TEX>$2001{\sim}2003$</TEX>년 동안에는 0.66% 이상. 우리는 각 경우에 대하여 예보의 성공 유무를 확인할 수 있는 예보 분할표(Contingency Table)를 만들고, 그들의 태양 주기 위상(Solar cycle phase)에 대한 의존성을 조사하였다 분할표로부터 우리는 PODy(the probability of detection yes), FAR(the false alarm ratio), Bias(the ratio of "yes" predictions to "yes" observations) 그리고 CSI(critical success index)와 같은 예보 평가 지수를 결정하였다. 이와 같은 예보에서 PODy와 CSI가 상대적으로 더 중요한 사실을 고려하여, 우리는 가장 좋은 후보가 case 3이라는 것을 발견하였다. 이 경우에 두 가지 예보에 대한 예보평가 지수는 아래와 같다: CH-CIR의 경우는 PODy=0.77, FAR=0.66, Bias=2.28, CSI=0.30이고, CH-storm의 경우는 PODy=0.81, FAR=0.84, Bias=5.00, CSI=0.16이다. 또한 태양 활동 극대기 이후 감쇄기간 동안의 지수들이 태양 극대기 이전의 값들 보다 훨씬 잘 예보되고 있음을 알 수 있다. 따라서 코로나 홀을 이용한 CIR의 예보는 충분한 가능성을 보여주고 있으나, 지자기 폭풍의 예보는 너무 많은 허위 예보로 인하여 다소 어려울 것으로 비상된다. In this study, we suggest an empirical forecast of CIR (Corotating Interaction Regions) and geomagnetic storm based on the information of coronal holes (CH). For this we used CH data obtained from He I <TEX>$10830{\AA}$</TEX> maps at National Solar Observatory-Kitt Peak from January 1996 to November 2003 and the CIR and storm data that Choi et al. (2009) identified. Considering the relationship among coronal holes, CIRs, and geomagnetic storms (Choi et al. 2009), we propose the criteria for geoeffective coronal holes; the center of CH is located between <TEX>$N40^{\circ}$</TEX> and <TEX>$S40^{\circ}$</TEX> and between <TEX>$E40^{\circ}$</TEX> and <TEX>$W20^{\circ}$</TEX>, and its area in percentage of solar hemispheric area is larger than the following areas: (1) case 1: 0.36%, (2) case 2: 0.66%, (3) case 3: 0.36% for 1996-2000, and 0.66% for 2001-2003. Then we present contingency tables between prediction and observation for three cases and their dependence on solar cycle phase. From the contingency tables, we determined several statistical parameters for forecast evaluation such as PODy (the probability of detection yes), FAR (the false alarm ratio), Bias (the ratio of "yes" predictions to "yes" observations) and CSI (critical success index). Considering the importance of PODy and CSI, we found that the best criterion is case 3; CH-CIR: PODy=0.77, FAR=0.66, Bias=2.28, CSI=0.30. CH-storm: PODy=0.81, FAR=0.84, Bias=5.00, CSI=0.16. It is also found that the parameters after the solar maximum are much better than those before the solar maximum. Our results show that the forecasting of CIR based on coronal hole information is meaningful but the forecast of goemagnetic storm is challenging.

To identify suitable merging methods to improve regional precipitation estimates using multiple sources of precipitation data, this study applied four different approaches (multiple linear regression (MLR), feedforward neural network (FNN), random forest (RF) and long short-term memory network (LSTM)) to merge four satellite precipitation products and one reanalysis data in the Jiangsu, Zhejiang and Shanghai of China. The pros and cons of the merging approaches are analyzed comprehensively, using correlation coefficient (CC), root mean square error (RMSE), relative bias (RB), probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI) as evaluation indexes. Our results show that: (1) All merging approaches can improve the accuracy of precipitation estimations, but only RF and LSTM can improve the daily precipitation event detection capacity. These approaches can significantly reduce errors in moderate precipitation scenarios, but do not effectively improve accuracy in light and heavy precipitation scenarios. (2) MLR was the least expensive computing cost method in our study and performed better than the other three methods when gauge density was low. However, MLR had the worst daily precipitation event detection capacity (CSI = 0.67). (3) FNN performed moderately in most experiments (CC = 0.87, RMSE = 4.65 mm/day, RB = 1.19 %, POD = 0.94, FAR = 0.29, CSI = 0.70). (4) The merged data generated by RF was the most accurate and had the best daily precipitation event detection capacity (CC = 0.87, RMSE = 4.61 mm/day, RB = − 0.33 %, POD = 0.97, FAR = 0.20, CSI = 0.78). RF performed best in moderate precipitation scenarios. However, it performed worse than other methods when gauge density was low. (5) LSTM was the most robust methods and performed best in light precipitation scenarios. The FAR of the LSTM-generated data was the smallest (0.15) among four fusion methods. However, LSTM had the most expensive computing cost and the worst accuracy of the merged data (CC = 0.86, RMSE = 4.68 mm/day, RB = − 9.36 %).

The existence of quasi-stationary alongshore atmospheric fronts typically located 30–70 km off the east coast of Taiwan is demonstrated by analyzing synthetic aperture radar (SAR) images of the sea surface acquired by the European Remote Sensing Satellites ERS-1 and ERS-2. For the data interpretation, cloud images from the Japanese Geostationary Meteorological Satellite GMS-4 and the American Terra satellite, rain-rate maps from ground-based weather radars, sea surface wind data from the scatterometer on board the Quick Scatterometer (QuikSCAT) satellite, and meteorological data from weather maps and radiosonde ascents have also been used. It is shown that these atmospheric fronts are generated by the collisions of the two airflows from opposing directions: one is associated with a weak easterly synoptic-scale wind blowing against the high coastal mountain range at the east coast of Taiwan and the other with a local offshore wind. At the convergence zone where both airflows collide, air is forced to move upward, which often gives rise to the formation of coast-parallel cloud bands. There are two hypotheses about the origin of the offshore wind. The first one is that it is a thermally driven land breeze/katabatic wind, and the second one is that it is wind resulting from recirculated airflow from the synoptic-scale onshore wind. Air blocked by the mountain range at low Froude numbers is recirculated and flows at low levels back offshore. Arguments in favor of and against the two hypotheses are presented. It is argued that both the recirculation of airflow and land breeze/katabatic wind contribute to the formation of the offshore atmospheric front but that land breeze/katabatic wind is probably the main cause.

Air pollution, particularly fine particulate matter (PM2.5), poses significant public health and environmental risks. This study explores the effectiveness of spatiotemporal graph neural networks (ST-GNNs) in forecasting PM2.5 concentrations by integrating remote-sensing hyperspectral indices with traditional meteorological and pollutant data. The model was evaluated using data from Switzerland and the Gauteng province in South Africa, with datasets spanning from January 2016 to December 2021. Key performance metrics, including root mean squared error (RMSE), mean absolute error (MAE), probability of detection (POD), critical success index (CSI), and false alarm rate (FAR), were employed to assess model accuracy. For Switzerland, the integration of spectral indices improved RMSE from 1.4660 to 1.4591, MAE from 1.1147 to 1.1053, CSI from 0.8345 to 0.8387, POD from 0.8961 to 0.8972, and reduced FAR from 0.0760 to 0.0719. In Gauteng, RMSE decreased from 6.3486 to 6.2319, MAE from 4.4891 to 4.4066, CSI from 0.9555 to 0.9560, and POD from 0.9699 to 0.9732, while FAR slightly increased from 0.0154 to 0.0181. Error analysis revealed that while the initial one-day ahead forecast without spectral indices had a marginally lower error, the dataset with spectral indices outperformed from the two-day ahead mark onwards. The error for Swiss monitoring stations stabilized over longer prediction lengths, indicating the robustness of the spectral indices for extended forecasts. The study faced limitations, including the exclusion of the Planetary Boundary Layer (PBL) height and K-index, lack of terrain data for South Africa, and significant missing data in remote sensing indices. Despite these challenges, the results demonstrate that ST-GNNs, enhanced with hyperspectral data, provide a more accurate and reliable tool for PM2.5 forecasting. Future work will focus on expanding the dataset to include additional regions and further refining the model by incorporating additional environmental variables. This approach holds promise for improving air quality management and mitigating health risks associated with air pollution.

Nowcasting is a critical technology for disaster prevention and mitigation, and the accuracy of radar echo extrapolation directly impacts forecasting performance. In most deep learning-based models, accurately predicting heavy precipitation remains a challenging task. Focusing on the region of China, this study proposes an improved model based on residual and attention mechanisms—RA-UNet—for precipitation nowcasting with a lead time of 3 h. The model introduces the residual neural network (ResNet) and the convolutional block attention module (CBAM) to integrate multi-scale features into the U-Net encoder–decoder architecture, enhancing its ability to capture the spatiotemporal evolution of precipitation systems. Meanwhile, depthwise separable convolutions are employed to replace conventional convolutions, significantly improving computational efficiency while preserving model performance. To evaluate the model’s performance, experiments were conducted using 6 min resolution radar echo data from China in 2024, with comparisons made against the optical flow (OF) method and the U-Net model. The experimental results show that RA-UNet demonstrates significant advantages in 3 h forecasting: its mean absolute error (MAE) is reduced by approximately 7%, the false alarm rate (FAR) decreases by about 20%, and it outperforms the comparison models in metrics such as the critical success index (CSI) and structural similarity index (SSIM). Notably, RA-UNet effectively mitigates intensity degradation in long-term forecasts, successfully predicting the trend of &gt;40 dBZ strong echo cores in two typical cases and significantly improving the premature dissipation problem of precipitation fields. This study provides a new approach to refined forecasting of complex precipitation systems, and future work will combine multi-source data fusion with physical constraint mechanisms to further enhance precipitation event prediction capabilities.

Weather radars are commonly used to track the development of convective storms due to their high resolution and accuracy. However, the coverage of existing weather radar is very limited, especially in mountainous and ocean areas. Geostationary meteorological satellites can provide near global coverage and near real-time observations, which can compensate for the lack of radar observations. In this paper, a deep learning method was used to estimate the radar composite reflectivity from observations of China's new-generation geostationary meteorological satellite FY-4A and topographic data. The derived radar reflectivity products from satellite observations can be used over regions without radar coverage. In general, the deep learning model can reproduce the overall position, shape, and intensity of the radar echoes. In addition, evaluation of the reconstruction radar observations indicates that a modified model based on the attention mechanism (Attention U-Net model) has better performance than the traditional U-Net model in terms of all statistics such as the probability of detection (POD), critical success index (CSI), and root-mean-square error (RMSE), and the modified model has stronger capability on reconstructing details and strong echoes.

Lightning is one of the frequent natural disasters, which seriously affects the secure and stable operation of the power system, especially the distribution network lines with weak reliability. In order to improve the power supply reliability of the distribution network, higher requirements are put forward for the accuracy of lightning warning. Therefore, this paper establishes a lightning warning model based on comprehensive multi-meteorological factor thresholds and analyzes the meteorological factor data such as atmospheric field strength, echo intensity, echo-top height, and vertical cumulative liquid water content under thunderstorm weather. The threshold value of each factor warning is obtained, and the corresponding threshold weight is calculated by the entropy weight method. According to the weight of each threshold, the comprehensive threshold index of lightning warning is obtained, and the lightning warning is based on this index. A total of 105 lightning data from May to June 2022 in Nanchang city were analyzed as samples. The thresholds of atmospheric field strength, echo intensity, echo-top height, and vertical cumulative liquid water content were 1.2 kV/m, 40 dBZ, 8 km, and 5.2 kg·m−2, respectively. The corresponding weights of each factor were 0.4188, 0.2056, 0.2105, and 0.165, respectively. This model was used to warn a thunderstorm event in July 2022 in Nanchang area. The success rate of the model warning was 0.91, the false alarm rate (FAR) was 0.11, and the critical success index (CSI) was 0.80. Compared with the single-factor threshold lightning warning model, the warning FAR is decreased by 6%, and CSI is increased by 14% while ensuring the high warning success rate.

ABSTRACTIn this paper, we have proposed a method of sea fog prediction using model sea surface wind (SSW) data and hourly optical satellite data with 500 m spatial resolution. Three sea fog cases in the Yellow Sea were selected from Geostationary Ocean Color Imager (GOCI) images, and their shifts were determined from the fog cluster centroid displacements at successive hours. The 4 km resolution Weather Research and Forecasting (WRF) model SSW were used to predict sea fog shifts. The sea fog prediction results were verified using validation indices like probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI). The predictions started with high initial skill scores (POD = 0.76, FAR = 0.18, and CSI = 0.66), and showed values 0.59, 0.30, and 0.46 for POD, FAR, and CSI, respectively for predictions after two hours. After three hours their values still remained close to the medians though decreased, but afterwards decreased considerably for POD and CSI. Thus the method is found suitable for short-term prediction of sea fog. Additionally, trajectory sensitivities were compared between 20 and 4 km resolution WRF which, in general, resulted in less errors for 4 km WRF model SSW.

Large-scale flash flood warning in China using deep learning

Granularity of the grid (both horizontally and vertically) is a key consideration when conducting localised Numerical Weather Prediction (NWP) modelling. Generally speaking, an NWP model with a finer grid can explicitly resolve more processes and require less parameterisation. However, a finer grid also requires more computation and it is not always clear that a finer grid will produce a more accurate forecast. In this study, we explore the sensitivity of rainfall prediction over Singapore to grid resolution. We use the Weather and Research Forecasting model (WRF) to forecast rainfall over Singapore and explore the performance of horizontal resolutions ranging from 1 km to 12 km. We test the performance on a set of dates from across the years 2020–2021 against both ground observations and radar-derived rain rates. When compared to ground observations, we show that, overall, the higher resolution produces the highest Critical Success Index (CSI) for rain rates in excess of 0.5 mm/h. When compared against radar-derived rain rates, the 1 km domain produces superior CSI values for all rain rates. The daily-average hourly Fractional Skill Score (FSS) was then calculated for some dates and showed agreement with the CSI results with the exception of a north-east monsoon day where, for heavier rain rates, the 3 km domain has superior FSS. We also investigate a particularly heavy rain event on 10 January 2021 and show that the 3 km grid has highest CSI for rain rates of 4 mm/h and 16 mm/h (based on both ground-based and radar-derived rain rates), however, the 1 km has superior FSS for this event.

The Critical Success Index (CSI) and Gilbert Skill score (GS) are verification measures that are commonly used to check the accuracy of weather forecasting. In this article, we propose that they can also be used to simplify the joint interpretation of positive predictive value (PPV) and sensitivity estimates across diagnostic accuracy studies of epilepsy data. This is because CSI and GS each provide a single measure that takes the weather forecasting equivalent of PPV and sensitivity into account. We have re-analysed data from our recent systematic review of diagnostic accuracy studies of administrative epilepsy data using CSI and GS. We summarise the results and benefits of this approach.

A form of the critical success index (CSI) is used by the National Weather Service to indicate the value of warnings. This verification statistic assumes that the times when an event was neither expected nor observed are of no consequence. It can be shown that the CSI is not an unbiased indicator of forecast skill but is proportional to the frequency of the event being forecast. This innate bias is demonstrated theoretically and via example. An unbiased verification statistic appropriate for forecast of rare events is presented and applied to severe convective weather warnings. Comparisons of this score to the CSI show the extent of the penalty the CSI extracts from forecasters who work in areas that are not climatically prone to given events.