A multilayer perceptron and multiclass support vector machine based high accuracy technique for daily rainfall estimation from MSG SEVIRI data

The current paper introduces a new multilayer perceptron (MLP) and support vector machine (SVM) based approach to improve daily rainfall estimation from the Meteosat Second Generation (MSG) data. In this study, the precipitation is first detected and classified into convective and stratiform rain by two MLP models, and then four multi-class SVM algorithms were used for daily rainfall estimation. Relevant spectral and textural input features of the developed algorithms were derived from the spectral MSG SEVIRI radiometer channels. The models were trained using radar rainfall data set colected over north Algeria. Validation of the proposed daily rainfall estimation technique was performed by rain gauge network data set recorded over north Algeria. Thus, several statistical scores were calculated, such as correlation coefficient (r), root mean square error (RMSE), mean error (Bias), and mean absolute error (MAE). The findings given by: (r = 0.97, bias = 0.31 mm, RMSE = 2.20 mm and MAE = 1.07 mm), showed a quite satisfactory relationship between the estimation and the respective observed daily precipitation. Moreover, the comparison of the results with those of two advanced techniques based on random forests (RF) and weighted ‘k’ nearest neighbor (WkNN) showed higher accuracy obtained by the proposed model.     

Optimization of One versus All-SVM using AdaBoost algorithm for rainfall classification and estimation from multispectral MSG data

In this paper, we implement the AdaBoost algorithm to optimize the classifications results of precipitations intensities carried out by One versus All strategy using Support Vector Machine (OvA-SVM). The model developed which combines the AdaBoost algorithm with a multiclass SVM is applied to images from the MSG (Meteosat Second Generation) satellite. Other variants to build multiclass SVMs, such as the OvO-SVM (One versus One SVM), SBT-SVM (Slant Binary Tree SVM) and DDAG-SVM (Decision Directed Acyclic Graph) are also implemented on which we tested the AdaBoost algorithm. The study showed that the AdaBoost algorithm performed better in the case of the OvA-SVM variant compared to the other variants.In order to evaluate the elaborated model, some classification techniques, such as the ECST Enhanced Convective Stratiform Technique (ECST), the SART where the Support vector machine, Artificial neural network and Random forest classifiers are combined, the Convective/Stratiform Rain Area Delineation Technique (CS-RADT) and the Random Forest technique (RFT) are applied. The classification results obtained show that AdaBoost with OvA-SVM (AdaOvA-SVM) presents very interesting performances where the evaluation parameters POD, POFD, FAR, BIAS, CSI and PC indicate the values 95.2%, 12.4%, 14.7%, 0.9, 88.1% and 96.5% respectively. Indeed, the AdaOvA-SVM technique has surpassed the CS-RADT, ECST and RFT techniques. As for the comparison with the SART, we noted that OvA-SVM presents very close results. The same trend was also observed when estimating precipitation. At the end of this study, it is shown that the AdaBoost algorithm performs better on a weak classifier or on a strong classifier operating in an unfavorable environment.     

Storm-time atmospheric density modeling using neural networks and its application in orbit propagation

Upper atmospheric densities during geomagnetic storms are usually poorly estimated due to a lack of clear understanding of coupling mechanisms between the thermosphere and magnetosphere. Consequently, the orbit determination and propagation for low-Earth-orbit objects during geomagnetic storms have large uncertainties. Artificial neural networks are often used to identify nonlinear systems in the absence of rigorous theory. In the present study, an attempt has been made to model the storm-time atmospheric density using neural networks. Considering the debate over the representative of geomagnetic storm effect, i.e. the geomagnetic indices ap and Dst, three neural network models (NNM) are developed with ap, Dst and a combination of ap and Dst respectively. The density data used for training the NNMs are derived from the measurements of the satellites CHAMP and GRACE. The NNMs are evaluated by looking at: (a) the mean residuals and the standard deviations with respect to the density data that are not used in training process, and (b) the accuracy of reconstructing the orbits of selected objects during storms employing each model. This empirical modeling technique and the comparisons with the models NRLMSIS-00 and Jacchia-Bowman 2008 reveal (1) the capability of neural networks to model the relationship between solar and geomagnetic activities, and density variations; and (2) the merits and demerits of ap and Dst when it comes to characterizing density variations during storms.     

Modelling of air temperature using remote sensing and artificial neural network in Turkey

The aim of this research was to forecast monthly mean air temperature based on remote sensing and artificial neural network (ANN) data by using twenty cities over Turkey. ANN contained an input layer, hidden layer and an output layer. While city, month, altitude, latitude, longitude, monthly mean land surface temperatures were chosen as inputs, and monthly mean air temperature was chosen as output for network. Levenberg–Marquardt (LM) learning algorithms and tansig, logsig and linear transfer functions were used in the network. The data of Turkish State Meteorological Service (TSMS) and Technological Research Council of Turkey–Bilten for the period from 1995 to 2004 were chosen as training when the data of 2005 year were being used as test. Result of research was evaluated according to statistical rules. The best linear correlation coefficient (R), and root mean squared error (RMSE) between the estimated and measured values for monthly mean air temperature with ANN and remote sensing method were found to be 0.991–1.254 K, respectively.     

State and disturbance estimation for test masses of drag-free satellites based on self-recurrent wavelet neural network

In recent years, the drag-free satellites have been widely used for some fundamental physical experiments, such as checking short-range effects of general relativity, geopotential determination and the exploration of static ocean current. And the space-borne detector of gravitational waves is one of the important applications for drag-free satellites in the future. In this study, the estimation of relative motion state and disturbance for test masses of drag-free satellite after release are researched. Firstly, the relative motion model between the test masses is established based on corresponding reference frames. Secondly, a self-recurrent wavelet neural network estimator is designed to estimate the pure gravitational relative motion state and disturbance for the test masses, and the sliding mode controller is used to transfer the relative motion state to the ideal state. Finally, the proposed estimation methods are verified by the simulation results.     

The impact of different background errors in the assimilation of satellite radiances and in-situ observational data using WRFDA for three rainfall events over Iran

Satellite radiances and in-situ observations are assimilated through Weather Research and Forecasting Data Assimilation (WRFDA) system into Advanced Research WRF (ARW) model over Iran and its neighboring area. Domain specific background error based on x and y components of wind speed (UV) control variables is calculated for WRFDA system and some sensitivity experiments are carried out to compare the impact of global background error and the domain specific background errors, both on the precipitation and 2-m temperature forecasts over Iran. Three precipitation events that occurred over the country during January, September and October 2014 are simulated in three different experiments and the results for precipitation and 2-m temperature are verified against the verifying surface observations. Results show that using domain specific background error improves 2-m temperature and 24-h accumulated precipitation forecasts consistently, while global background error may even degrade the forecasts compared to the experiments without data assimilation. The improvement in 2-m temperature is more evident during the first forecast hours and decreases significantly as the forecast length increases.     

改进U-Net网络及在遥感影像道路提取中的应用

高分辨率遥感图像分割在军事、民用等领域具有良好的应用前景,但由于复杂的背景条件以及干扰物的遮挡,导致现有算法无法较好地从遥感影像中提取道路细节信息。研究基于改进U Net网络模型,提出了MDAU-Net（multi dimension attention U-Net）网络结构模型,通过对U-Net网络结构加深至七层结构来提升精细分割道路的能力;并提出了一种多维注意力模块MD-MECA（multi dimension modified efficient channel attention）,将其添加至编码部分的特征传递步骤中,以达到对编码部分的特征传递进行优化的目的;其中利用DropBlock与Batch Normalization解决网络训练过程中出现的过拟合。试验结果表明:改进后算法可以有效提升道路的提取效果,在测试集上的准确率达到了97.04%。