Accurate estimation of tropical cyclone (TC) intensity, as an important task for meteorological hazard monitoring and warning, is often regarded as a problem of intensity grade classification or intensity regression. Good classification accuracy of TC intensity is of great significance to improve the accuracy of TC intensity estimation. Based on the infrared brightness temperature data of the Northwest Pacific Basin, a TC intensity classification method based on feature fusion of infrared satellite images is proposed in this paper. The proposed method divides infrared satellite images into three categories according to intensity grades of TCs: TS+STS, STY, VSTY+ViolentTY. Firstly, the features of the middle layer of Xception network which is one of deep convolutionary neural network models are extracted. Then, after 1×1 convolution and pooling, they are fused with the features of the output of the full connection layer. Finally, the fused features are sent into the Softmax classifier to determine the category. In this paper, by fusing the information from different convolutional layers, the information complementarity of global and local features can be realized. The results show that the proposed model has the best performance after the fusion of the features output from the seventh residual module in middle flow and the output features of the full connection layer. The classification accuracy of TC intensity grades is improved to 80.99%, which is 1.5% higher than the original Xception network. The proposed method can be applied to the task of TC intensity estimation.
Automatic detection of tropical cyclone (TC) regions from satellite images can provide regions of interest for intelligent TC positioning and intensity determination, and improve the efficiency and accuracy of intelligent disaster weather forecasting. There are currently few studies on automatic detection of TCs from satellite images. In recent years, deep learning technology has developed rapidly in various fields. This paper improves the Faster-RCNN target detection model in deep learning and applies it to the TC detection. The TC detection model designed in this paper is based on the original Faster-RCNN network framework, and the feature extraction network is changed from the original VGG16 network to the ResNet50 network . On this basis, this paper designs a feature fusion network Single Output Feature Fusion Networks (SOFFN). The feature layer used for detection can combine the semantic information of the high-level feature map and the high-resolution feature information of the low-level feature map, fuse different feature layers. At the same time, a new attention mechanism, Channel Linear Weighted Networks (CLWNet), based on the Squeeze-and-Excitation Networks (SENet) channel attention mechanism improvement is added to the model designed in this paper to improve the detection performance. In this paper, China's FY-2D satellite images are used to verify the performance of the proposed model. Experimental results show that the proposed model has achieved good results in TC detection.
The prediction of rapid intensification (RI) is an important part of tropical cyclone (TC) intensity change prediction task, RI is defined as TC subsequent intensity change increase greater than a certain threshold. Accurately predicting RI can help improve the accuracy of TC intensity prediction, thereby reducing people's economic and property losses. In recent years, an increasing number of researchers have started to use satellite imagery for RI prediction. In this study, deep learning is utilized to combine infrared with microwave satellite images for RI prediction of TCs. The core idea can be formulated as follows: The residual image between feature of historical satellite image sequence and current satellite image is taken as the final feature matrix. Multilayer ConvLSTM is used to extract the features of historical satellite image sequence (time resolution is three hours), the residual image between the generated feature and the current satellite image is taken as the feature matrix. Finally the feature matrix is input into the classifier to predict RI. Experiment shows that in the prediction problem of RI for TCs (when the RI threshold is 35kt) our method on the test dataset has reached 0.552 at probability of detection (POD), and the false alarm ratio (FAR) reached 0.847, and heidke skill score (HSS) reached 0.186. Compared with the current methods only using satellite images to the prediction of RI, our method reduced FAR by 1.7%, improved POD by 15.4% and HSS by 13.4%.
Different characteristics of satellite images are reflected in different channels, so the monitoring and early warning of meteorological disasters based on satellite image data of a single channel may not achieve satisfactory results. The infrared channel satellite image reflects the ground and cloud top infrared radiation or the temperature distribution. The water vapor channel satellite image reflects the spatial distribution of water vapor in the upper atmosphere. The two channel satellite images reflect atmospheric characteristics from different wavebands. This paper proposes an infrared and water vapor channel satellite image fusion model (TCIE-ResNeXtGAN) based on dual discriminator generative adversarial network and ResNeXt to improve the accuracy of tropical cyclone (Tropical Cyclone, TC) intensity estimation. For this reason, we define the factors affecting TC intensity estimation, namely the brightness temperature gradient in satellite images, based on our previous work, and introduce it to the loss function in the proposed deep learning model to guide the training process of the model. In this way, the purpose of improving the estimation accuracy of TC intensity using fused satellite images is achieved. To demonstrate the effectiveness of the fusion results in this paper, we compare performance between the existing three models and our model. The experimental results show that the proposed image fusion method in this paper can preserve the infrared and water vapor dual-channel information to the greatest extent, while improving the estimation accuracy of TC intensity using fused satellite image.
: In order to improve the accuracy of objective location of tropical cyclones, a method of tropical cyclones center location based on the deviation Angle variance combined with saliency detection and wavelet transform is proposed in this paper. First, the rough foreground center of the satellite cloud image sensing region is obtained by means of the saliency detection segmentation algorithm. From the center outwards, the appropriate rectangular area is selected as the subsequent detection area. Then the binary image obtained after saliency detection and the binary image segmented by wavelet transform are conducted phase and calculation, and only the binary image in the dense cloud region containing the location of sharp change of brightness temperature is obtained. Finally, a reference center is set in the detection area to calculate the deviation angle matrix of the kernel area, and the deviation angle variance (DAV) matrix of the detection area is obtained by traversing each point in the detection area. The minimum value of the variance matrix corresponds to the location of the typhoon center. The mean location deviation of the eyed TC was 28km compared with the best track data from China, Japan and the United States, and that of the non-eye TC was 47km. Compared with the localization method without significant detection and wavelet transform segmentation, the localization method in this paper can reduce the localization deviation. It is also a potential approach compared with the same type of TC localization research.
Summer precipitation estimation is one of the key and difficult tasks in short-term climate prediction because of the large amount of convective precipitation in summer which is characterized by uneven distribution, large intensity, short duration and rapid change with time. In order to improve the accuracy of summer precipitation estimation, an efficient method by multi-time scale Support Vector Machine (SVM) with quantum optics inspired optimization (QOIO) is proposed in this paper. And the performance of the proposed method is verified by radar reflectivity and precipitation data of automatic weather stations (AWSs) in Shanghai. Using radar reflectivity and precipitat ion in the most relevant time scale, a rainfall estimation model based on multi-time scale SVM is established for each AWS to estimate next 6-minute precipitation. Compared with the traditional single Z-R relationship, linear regression, K-nearest neighbor and ordinary SVM, the results show the higher Threat Score and lower root mean square error can be obtained by the proposed method in summer precipitation estimation.
In order to improve the accuracy of numerical weather prediction(NWP) temperature, a support vector machine (SVM) model based on LASSO feature analysis is proposed to revise the predicted temperature for the next 12 hours. In this paper, high-resolution mode prediction data that include 2m temperature and related meteorological factors forecasted by the European Center of Medium range Weather Forecast ( ECMWF) , and the temperature data of the automatic stations in East China and coastal areas provided by the Shanghai Meteorological Bureau are used to build the proposed model. , In this paper, The results show that the root mean square error, absolute error and accuracy are greatly improved by the proposed prediction model. The comprehensive performance of the proposed method is better than that of the traditional linear regression technology.
Since the labels of training samples are related to bags not instances, the multiple instance learning (MIL) is a special ambiguous learning paradigm. In this paper, we propose a novel bag space (BS) construction and extreme learning machine (ELM) combination method named BS_ELM for MIL, which can capture the bag structure and use the efficiency of ELM. Firstly, sparse subspace clustering is performed to obtain the cluster centers and a new bag space is constructed. Then ELM is used to classify bags in the new space. Experiments on data sets demonstrate the utility and efficiency of the proposed approach as compared to the other state-of-the-art MIL algorithms.
Nowcasting or very short-term forecasting convective storms is still a challenging problem due to the high nonlinearity and insufficient observation of convective weather. As the understanding of the physical mechanism of convective weather is also insufficient, the numerical weather model cannot predict convective storms well. Machine learning approaches provide a potential way to nowcast convective storms using various meteorological data. In this study, a deep belief network (DBN) is proposed to nowcast convective storms using the real-time re-analysis meteorological data. The nowcasting problem is formulated as a classification problem. The 3D meteorological variables are fed directly to the DBN with dimension of input layer 6*6*80. Three hidden layers are used in the DBN and the dimension of output layer is two. A box-moving method is presented to provide the input features containing the temporal and spatial information. The results show that the DNB can generate reasonable prediction results of the movement and growth of convective storms.
An objective technique is presented for estimating tropical cyclone (TC) innercore two-dimensional (2-D) surface wind field structure using infrared satellite imagery and machine learning. For a TC with eye, the eye contour is first segmented by a geodesic active contour model, based on which the eye circumference is obtained as the TC eye size. A mathematical model is then established between the eye size and the radius of maximum wind obtained from the past official TC report to derive the 2-D surface wind field within the TC eye. Meanwhile, the composite information about the latitude of TC center, surface maximum wind speed, TC age, and critical wind radii of 34- and 50-kt winds can be combined to build another mathematical model for deriving the innercore wind structure. After that, least squares support vector machine (LSSVM), radial basis function neural network (RBFNN), and linear regression are introduced, respectively, in the two mathematical models, which are then tested with sensitivity experiments on real TC cases. Verification shows that the innercore 2-D surface wind field structure estimated by LSSVM is better than that of RBFNN and linear regression.
In order to extract the effective information in different modalities of the tumor region in brain Magnetic resonance imaging (MRI) images, we propose a brain MRI tumor image fusion method combined with Shearlet and wavelet transform. First, the source images are transformed into Shearlet domain and wavelet domain. Second, the low frequency component of Shearlet domain is fused by Laplace pyramid decomposition. Then the low-frequency fusion image is obtained through inverse Shearlet transform. Third, the high frequency subimages in wavelet domain are fused. Then the high-frequency fusion image is obtained through inverse wavelet transform. Finally, the low-frequency fusion image and high-frequency fusion image are summated to get the final fusion image. Through experiments conducted on 10 brain MRI tumor images, the result shown that the proposed fusion algorithm has the best fusion effect in the evaluation indexes of spatial frequency, edge strength and average gradient. The main spatial frequency of 10 images is 29.22, and the mean edge strength and average gradient is 103.77 and 10.42. Compared with different fusion methods, we find that the proposed method effectively fuses the information of multimodal brain MRI tumor images and improves the clarity of the tumor area well.
The data of current PM2.5 model forecasting greatly deviate from the measured concentration. In order to solve this problem, Support Vector Machine (SVM) and Particle Swarm Optimization (PSO) are combined to build a rolling forecasting model. The important parameters (C and γ) of SVM are optimized by PSO. The data (from February to July in 2015), consisting of measured PM2.5 concentration, PM2.5 model forecasting concentration and five main model forecasting meteorological factors, are provided by Shanghai Meteorological Bureau in Pudong New Area. The rolling model is used to forecast hourly PM2.5 concentration in 12 hours in advance and the nighttime average concentration (mean value from 9 pm to next day 8 am) during the upcoming day. The training data and the optimal parameters of SVM model are different in every forecasting, that is to say, different models (dynamic models) are built in every forecasting. SVM model is compared with Radical Basis Function Neural Network (RBFNN), Multi-variable Linear Regression (MLR) and WRF-CHEM. Experimental results show that the proposed model improves the forecasting accuracy of hourly PM2.5 concentration in 12 hours in advance and nighttime average concentration during the upcoming day. SVM model performs better than MLR, RBFNN and WRF-CHEM. SVM model greatly improves the forecasting accuracy of PM2.5 concentration one hour in advance, according with the result concluded from previous research. The rolling forecasting model can be applied to the field of PM2.5 concentration forecasting, and can offer help to meteorological administration in PM2.5 concentration monitoring and forecasting.
An efficient algorithm based on continuous wavelet transform combining with pre-knowledge, which can be used to detect the defect of glass bottle mouth, is proposed. Firstly, under the condition of ball integral light source, a perfect glass bottle mouth image is obtained by Japanese Computar camera through the interface of IEEE-1394b. A single threshold method based on gray level histogram is used to obtain the binary image of the glass bottle mouth. In order to efficiently suppress noise, moving average filter is employed to smooth the histogram of original glass bottle mouth image. And then continuous wavelet transform is done to accurately determine the segmentation threshold. Mathematical morphology operations are used to get normal binary bottle mouth mask. A glass bottle to be detected is moving to the detection zone by conveyor belt. Both bottle mouth image and binary image are obtained by above method. The binary image is multiplied with normal bottle mask and a region of interest is got. Four parameters (number of connected regions, coordinate of centroid position, diameter of inner cycle, and area of annular region) can be computed based on the region of interest. Glass bottle mouth detection rules are designed by above four parameters so as to accurately detect and identify the defect conditions of glass bottle. Finally, the glass bottles of Coca-Cola Company are used to verify the proposed algorithm. The experimental results show that the proposed algorithm can accurately detect the defect conditions of the glass bottles and have 98% detecting accuracy.
An efficient algorithm for typhoon center location is proposed using fractal feature and gradient of infrared satellite cloud image. The centers are generally located in this region for a typhoon except the latter disappearing typhoon. The characteristics of dense cloud region are smoother texture and higher gray values than those of marginal clouds. So the window analysis method is used to select an appropriate cloud region. The window whose difference value between the sum of the gray-gradient co-occurrence matrix and fractal dimension is the biggest is chosen as the dense cloud region. The temperature gradient of the region, which is near typhoon center except typhoon eye, is small. Thus the gradient information is strengthened and is calculated by canny operator. Then we use a window to traverse the dense cloud region. If there is a closed curve, the region of curve is considered as the typhoon center region. Otherwise, the region in which there is the most texture intersection and the biggest density is considered as the typhoon center region. Finally, the geometric center of the center region is determined as the typhoon center location. The effectiveness is test by Chinese FY-2C stationary satellite cloud image. And the result is compared with the typhoon center location in the “tropical cyclone yearbook” which was compiled by Shanghai typhoon institute of China meteorological administration. Experimental results show that the high location accuracy can be obtained.
A new fast algorithm to segment man-made target from infrared image is given employing Bezier histogram and edge information. In order to reduce computation burden, an efficient approach to select region of interest (ROI) is proposed based on prior-information. Thus a piece of region, which contains a target to be segmented, is extracted from original image. The gray level histogram of ROI is smoothed by Bezier curve to restrain noise in the ROI. Thus Bezier histogram is obtained. Peaks of curvature curve in the Bezier histogram are detected to obtained segmentation thresholds. An optimal segmentation threshold is selected with a new criterion. The optimal threshold segments the ROI well. In order to obtain better segmentation result, firstly a new algorithm, which bases on discrete stationary wavelet transform and non-linear gain operator is proposed to enhance the detail of target in the ROI. Canny operator is used to extract the edge information of target in the enhanced ROI. Finally, excellent segmentation result is obtained by combining Bezier histogram threshold method with edge information of target. Experimental results show that a man-made target can be segmented effectively from complex background in infrared image by the new algorithm.
A least squares support vector machine (LS-SVM) based signal processing approach of reflective fiber optic displacement sensor is presented. The example for extending measuring range of the sensor using LS-SVM has been illustrated. From the experimental results, it can be clearly seen that not only the measuring range can be extended to the whole response characteristics of the fiber optics displacement sensor effectively, but also a desired linear relationship between the actual displacement and the LS-SVM predicted output can be obtained. This means the method proposed is very effective for the signal processing of the sensor.
An identification approach of nonlinear optical dynamic systems, based on adaptive kernel methods which are modified version of least squares support vector machine (LS-SVM), is presented in order to obtain the reference dynamic model for solving real time applications such as adaptive signal processing of the optical systems. The feasibility of this approach is demonstrated with the computer simulation through identifying a Bragg acoustic-optical bistable system. Unlike artificial neural networks, the adaptive kernel methods possess prominent advantages: over fitting is unlikely to occur by employing structural risk minimization criterion, the global optimal solution can be uniquely obtained owing to that its training is performed through the solution of a set of linear equations. Also, the adaptive kernel methods are still effective for the nonlinear optical systems with a variation of the system parameter. This method is robust with respect to noise, and it constitutes another powerful tool for the identification of nonlinear optical systems.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.