In order to achieve further improvement in the forecast accuracy of station temperatures and enhance the forecast capability for extreme temperatures， this study establishes a 24-hour national station daily maximum （minimum） temperature forecast model for Jiangxi Province based on the LightGBM machine-learning algorithm and the MOS forecast framework by using the surface observation data of 91 national stations in Jiangxi Province and the upper-air and surface forecast data of the ECMWF model from 2017 to 2019.The results of the 2020 evaluation show that the LightGBM model daily maximum （minimum） temperature forecast is consistent with the observed trend， and the annual average forecast is better than that of three numerical models， ECMWF， CMA-SH9 and CMA-GFS， two machine learning products， RF and SVM， and subjective revision products.In terms of the spatial and temporal distribution of forecast errors， the model's daily maximum （minimum） temperature forecast errors in winter and spring are slightly larger than those in summer and autumn； the daily maximum temperature forecast errors show the spatial distribution characteristics of "larger in the south and smaller in the north， and larger in the periphery than in the centre"， while the opposite is true for the daily minimum temperatures.In terms of important weather processes， the LightGBM model has the best prediction effect among the seven products in the high temperature process； in the strong cold air process， the LightGBM model is still better than the three numerical model products and the other two machine-learning models， but the prediction effect of the daily minimum temperature is not as good as that of the subjective revision products.After a simple empirical correction for the low-temperature forecast error in the strong cold air process， the model low-temperature forecast effect is close to that of the subjective revision product.The model significance analysis shows that the recent surface observation features also contribute to the model construction， and the results can be used as a reference for model improvement and temperature forecast product development.At present， the LightGBM model temperature forecast products have been applied to meteorological operations in Jiangxi Province.

This study utilized the 2020 -2021 China Meteorological Administration （CMA） Land Data Assimilation System （CLDAS） hourly surface air temperature （T2m） product in combination with the T2m forecast data from the CMA Shanghai Rapid Update Cycle Numerical Forecast （CMA-SH3）.A deep learning semantic segmentation model called MT-Cunet was developed to achieve a 24-hour T2m grid forecast that is updated on an hourly basis.The forecast results for 2022 were then tested and evaluated.Results showed that： MT- Cunet has demonstrated the most effective revision during the 3~9 h time horizon in the study range.It shows a significant reduction of 42.4% and 40.9% in the mean MAE and mean RMSE， respectively.The revision effect during the 10~24 h time horizon is also noteworthy， with a reduction of 26.7% and 26.3% in the mean MAE and mean RMSE， respectively.When evaluating low-temperature （≤0 ℃） and high-temperature （≥35 ℃） events， MT-Cunet exhibits a positive bias in high-temperature forecasts while showing a negative bias in low-temperature forecasts， and the magnitude of error is much smaller compared to CMA-SH3.On the spatial scale， MT-Cunet can substantially reduce the T2m forecast error in complex terrain and decrease the MAE dispersion of CMA-SH3， resulting in a more stable distribution of forecast errors.By examining and assessing the regional warming and cold wave processes in February and March 2022， it has been found that MT-Cunet demonstrates superior capability in predicting the timing and magnitude of temperature increases and decreases.In both warming and cold wave processes， the MAE of MT-Cunet is 28.9% and 33.8% lower than that of CMA-SH3， respectively.This suggests that the MT-Cunet model exhibits improved forecasting skills in transitional weather processes.Therefore， by employing a fast-updating cycle numerical model， it is possible to rapidly increase the number of forecast samples.Additionally， by refining the objective method of the semantic segmentation deep learning model， this approach effectively addresses the issue of poor performance in deep learning training caused by the limited amount of data in conventional numerical models.Furthermore， it opens up new possibilities for maximizing the utilization of domestic model resources and promoting the wider application of domestic model post-processing.

Low accuracy of extended forecast remains an important scientific problem in the current stage， and qualified extended forecast is of great significance for disaster prevention and mitigation.In this study， the machine learning method was used to forecast the summer precipitation during the extension period （5~30 days） in China， and explore the possible contribution of soil moisture to extended forecast of precipitation.Based on the results， machine learning methods remarkably outweigh traditional linear models in terms of forecast accuracy， with Catboost， Lightgbm and Adaboost being the optimal machine learning methods.According to further analysis， the abnormal evaporation and sensible heat anomaly caused by the surface soil moisture anomaly in the Yangtze River Basin can lead to the atmospheric circulation and vertical movement anomaly， which eventually affects summer precipitation.The set of three optimal machine learning methods was applied to calculate the contribution rate of each forecasting factor in the model.It was found that the local soil moisture dominated the extended precipitation in the Yangtze River Basin from the 5th day to the 10th day， while the local soil moisture played a dominant role on previous precipitation from the 10th day to the 15th day， and the extended precipitation in the Yangtze River Basin during the period of Day 20~30 was basically controlled by large-scale circulation.Besides， the influence of non-local soil moisture on extended precipitation was evaluated， the results of which showed that the surface soil moisture in Indo-China Peninsula mainly contributed to the extended precipitation in the Yangtze River Basin from the 15th day to the 30th day.By adding the surface soil moisture of Indo-China Peninsula to the extended precipitation model in Northeast China， it was found that surface the soil moisture failed to improve the extended forecast accuracy of precipitation in this area， which verified the availability of the machine learning model.This study provides a certain reference for forecasting precipitation in the extended period and exploring the contribution rate of forecasting factors.

Sand and dust is a typical weather disaster which outbreaks in arid and semi-arid areas globally.This natural phenomenon， which is the result of stormy winds， raises a lot of dust from desert surfaces and decreases visibility to less than 1 km.The dust aerosol generated from dust storm dominates the aerosol loading in the troposphere and has comprehensive impacts on the global environment， weather， climate and ecology.Monitoring sand and dust from space using satellite remote sensing has become one of the most important issues in this field.However， sand and dust is difficult to accurately characterize by using single-band and linear models.Feng Yun-4A （FY-4A） imagery provides a good data source for timely and accurate monitoring of sand and dust.The machine learning models are important tools in sand and dust monitoring and forecasting.In this paper， the Normalized Difference Dust Index （NDDI）， Random Forests （RF） and Convolutional Neural Networks （CNN） were employed to monitor sand and dust based on the Advanced Geostationary Radiation Imager （AGRI） of geostationary FY-4A meteorological satellite in the Tarim Basin.The results showed that， sand and dust can be identified by NDDI_AGRI thresholds calculated using AGRI data.The determination of the NDDI_AGRI thresholds were obtained through statistical analysis of pixels， but it is necessary to take different thresholds for different times AGRI data.There are some identification errors in the cross region of cloud and land， and some vegetation coverage and desert by the NDDI_AGRI thresholds.The values of Precision， Recall， and F1-score of testing samples were all 100%； and the accuracy of cross validation of training samples was 99.5% for the sand and dust model of RF.The Loss and Accuracy in the estimation obtained using the CNN algorithm were about 0.1% and 99.9%， respectively， versus the training samples and testing samples.Both RF and CNN models have the ability and robustness to be used in sand and dust monitoring.The efficiency of two models had been checked using new dust events.Results show that the CNN algorithm preforms better than RF algorithm in identifying the junction of dust and non-dust.The RF and CNN algorithm have identification errors in some parts of sand and dust monitoring process， such as the mixed area of dust and clouds， and the Gobi area.The research results of this paper provide an important basis application of machine learning combined with FY-4A meteorological satellite data to monitor sand and dust operational.

At present， the extrapolation forecast based on radar echoes is the mainstay of disaster weather 0~2 hours nowcasting.This paper proposes a convolutional gated recurrent unit neural network （ConvGRU） by using radar mosaics at 6 min intervals obtained from the radar images provided by 11 doppler radars in Guangdong Province from 2015 to 2018.Through the automatic learning of massive data， the inherent characteristics of the data and the contained physical laws can be discovered using the proposed network.A multi-loss function weighting and hierarchical weighting strategy are proposed.Based on the ConvGRU framework， a three-layer self-encoding model （Encoder-Decoder） is built for training to establish a radar echo prediction model which predicts radar echoes for 20 consecutive frames in the next 2 hours by 6 minutes.The results are compared with the operationally applied methods including tracking radar echoes by correlation （TREC）， optical flow， and particle filter using typical case analysis and long-term verification.All the subjective and objective evaluation results indicate that the proposed ConvGRU method shows better forecasting performance in severe convective weather systems in predicting radar echo position， intensity and shape than other methods.These results indicate that the deep learning method can better grasp the characteristics of the strong echo area， and predict the strong echo accurately to a certain extent by automatic learning of time-series radar echo data.For the long-term evaluation results， the ConvGRU method has higher critical success index （CSI） and probability of detection （POD） scores than those of the traditional TREC， optical flow and particle filtering methods， and has the lowest false alarm rate （FAR） scores among all methods， suggesting it could be widely used in operational applications.However， the deep learning-based method has the limitation of losing spatial detail information in radar echoes due to the up-sampling and down-sampling operators， and the prediction performance of stratiform cloud precipitation is relatively poor.

The eddy covariance long-term measurements commonly include data gaps due to system failures， quality control and quality assurance.In this study， the marginal distribution sampling （MDS） algorithm and three machine learning algorithms （random forest RF， support vector machine SVM and artificial neural networks ANN） were applied to fill the gaps of sensible heat flux （H）， latent heat flux （LE） and net ecosystem exchange in 2016 over an alpine ecosystem.Results indicate that the performance of RF is better than SVM and ANN.During the nighttime， the periods of sunrise and sunset， and in the winter and spring， the performance of three machine learning algorithms is relatively weak， compared to other periods or seasons.On the monthly and annual scales， the filled NEE budget is significantly influenced by the choice of gap-filling method， compared to H and LE.

Based on the temporal dependence of forecast errors derived from numerical weather prediction, the back-propagation (BP) neural network is used to establish the prediction model for predicting non-systematic forecast error. The effectiveness of this model is tested with the analysis and 24-hours forecast data produced by T213 model from 2003 to 2007. The results show that the established BP neural network model has a good ability on predicting non-systematic error in the next 24 hours. For most of 332 test samples, the spatial distribution of the predicted non-systematic errors is consistent with the truth. The non-systematic error estimated by BP neural network model can further correct forecasts on the basis of the systematic error correction, and its correction effectiveness is better than that of the systematic error correction only. For 332 test samples, the effective rate of systematic error correction on forecasts is 61%, but the effective rate of nonsystematic error further correction can increase to 82%.

A forecast model for the ice thickness is developed using winter ice thickness data observed in Zhangye National Wetland Park watershed and surface air temperature and ground temperature data observed in Zhangye meteorological station from December 2011 to March 2012. This model is based on BP neural network which can approximate any nonlinear function. The forecasting skill of this model is validated by comparing the forecasted ice thicknessresults with the observed ones. The results show that: The frozen thickness of forecasting model is able to have comparatively ideal forecasting effect to frozen thickness, mobile frozen water area thickness forecasting history intends to jointly lead height to amount to 96.8%, the model tries reporting accurate rate for 85.7%, motionless frozen water area thickness forecast history intends to jointly lead amounting to 87.8%, the model tries reporting accurate rate for 80.0%. They have very good actual application value. It is function index accords with actual request. This forecasting model based on BP neural network is of good performance in forecasting the ice thickness.

It is discussed that the method of remote sensing retrieval of temperature, relative humidity and cloud liquid water profiles from 12\|channel ground\|based microwave radiometer by BP neural network, which is on the basis ofsounding data, training the neural network with atmospheric profiles in four seasons, testingand analizingthe accuracy of the network. Finally, the results of retrieval with southern suburb of Beijing of 12\|channel microwave radiometer data show that: With only a few training cases, the retrieval method BPNN referredin this article is more realistic than that of the microwave radiometer.

使用误差反向传播网络(BPN)和约当网络(JN)两种人工神经网络(ANN)以及支持向量机(SVM)对降雨量进行了１ h和3 h预报的研究， 并与交叉相关法(CCM)外推预报的结果进行了比较。针对安徽省2003年6~7月的降水过程， 比较了网络(文中指BPN、 JN和SVM)和CCM预报降雨量与实况降雨量的雨带分布、 强降雨区域和强度； 使用命中率(HR)、 虚警率(FAR)、 漏报率(NAP)、 临界成功指数(CSI)、 相关系数(CC)和均方根误差(RMSE)这6个指标并结合天气分析检验网络和CCM的预报效果。结果表明： 网络和CCM对雨带和强降雨区域的预报比较准确， 但是对强降雨中心位置和强度的预报与实况存在差异； 在使用HR、 FAR、 NAP和CSI检验预报效果时设定的阈值对预报结果的评价有影响； 预报的中小尺度结构与天气分析的结果一致； 网络与CCM以及不同的网络之间的预报结果存在着差异； 连续预报的结果表明, 与CCM相比, 网络对3 h预报的效果优于1 h的。

为了提高紫外线预报准确率， 应用BP（Back Propagation Learning Algorithm）神经网络模型和支持向量机（Support Vector Machines， 简称SVM）回归方法建立重庆市主城区紫外线辐射强度客观预报模型。统计相关分析结果显示， 不同季节影响紫外线辐射强度的主要因素并不相同。对所有相关分析因子用逐步回归方法， 按方差贡献大小筛选出预报因子， 以每日紫外线平均辐射量为预报对象， 分季节建立预报模型。比较用不同方法建立的预报模型发现， 两种非线性模型（BP模型和SVM模型）的拟合能力优于线性逐步回归模型， 但独立样本检验结果表明， 3种模型的预报准确率基本相当。将3种方法所建预报模型应用T213数值预报资料进行业务试报， 得到较好预报效果。

以1960-2007年共48年6月份西行进入南海海域的热带气旋样本为基础， 将热带气旋中心附近最大风速作为台风强度， 以气候持续预报因子作为模型输入， 采用模糊神经网络方法, 进行了热带气旋强度预报模型的预报建模研究。结果表明， 对175个独立预报样本模糊神经网络方法的南海热带气旋强度24 h的预报平均绝对误差为3 m·s-1。另外， 根据相同的热带气旋样本及预报因子， 还进一步将该预报方法与国内外普遍采用的气候持续法热带气旋强度预报方法进行对比分析， 结果表明， 气候持续预报方法的预报误差明显偏大， 独立样本强度预报平均绝对误差为4.54 m·s-1。

By means of the Doppler radar measurements and automatic precipitation station data collected in the Linyi region during four precipitation processes of 2005. The Back-Propagation Neural Network (BPNN) was used to estimate the rainfall. In order to contrast with neural network, the improved window probability matching method（WPMM）was employed to determine the relationship between radar echo intensity (Z) and precipitation intensity (R), and the Z-R relation was further used to estimated the rainfall. Based on analysis index study such as mean relative error, root mean square error, correlation coefficient, correlation curve slope, the results suggested that the hourly rainfall andaccumulation rainfallestimationof the precision from BPNN is higher than from Z-R relation, and the precision from calibration samples is higher than evaluation samples. Rainfall estimation of BPNN was in good consistence with those observation by rain-gauge andcan truly reflect the precipitation status over the ground surface. Rainfall estimation of Z-R relation would yield underestimation of different degree with the change of rainfall intensity.

According to the meteorological data of 32 stations in Tibet from November 2003 to October 2005,based on the interpretation of numerical forecasting products of ECWMF,T213 and so on,an artificial neural network model is constructed with dynamic learning rate BP algorithm to forecast daily maximum/minimum temperature from one day to seven days.The results show that the neural network model has preferable adaptive learning and non-linear mapping abilities and its forecast results can satisfy the precision requirement for real-time forecast,which has good referenced value in short and mid-range real-time operational forecast of extreme temperature.

To set up the short-term climate prediction model in this paper,both the neural network and connection weights of genetic algorithm are optimized,the best individual in evolution process is reserved. Thus it can overcome the defects of unstability of neutral network initial weight and falling easily into local solution.As the applied example of a short-time climate forecast model,April mean precipitation in Guangxi area is taken as the predict and,the antecedent montly mean 500 hPa potential field and sea surface temperature filed in some high correlation areas are taken as the pedictors.Predictive performance between the new model and linear regression model for same predictors is discussed based on the independent samples.Results show that the model is superior in prediction accuracy and stability compared with the traditional method.

Based on itself-period change of predictand characterized by mean generating function method,500 hPa monthly mean height field and the predictor of monthly mean SST,a new short-term forecast pattern is established by the artficial neural.June rainfall in north-,center-and south-parts of Guangxi,respectively,as predicative object is carried out predicative experiment.The results show that the new method has more better prediction ability and physical foundation than the regression prediction patterns of mean genrating function,500 hPa height field and SST.

Based on the deep research of neural networks and time series, the two methods are studied within a unified framework of system theory. Three optimum methods are proposed based on space, time, mode relativity using improved BP. The experiments reveal that the proposed methods are very effective in processing of data in HEIFE area.

In this paper, based on the Back-Propagation(B-P) algorithm of neural network, some recognition models of hail cloud with 3 and 4 parameters are formulated by using the data of radar echo and part radiosonde over Chengdu, Neijiang and Luzhou Results of analysis and verification indicate that the models of hail cloud recognition formulated by B-P algorithm possess not only good fitting ratio and forecasting accuracy, but also superiority comparing with other methods because of self-organization, self-learning and self-adaptability.