2014 First International Conference on Systems Informatics, Modelling and Simulation A Short Term Tornado Prediction Model Using Satellite Imagery Viswanath Kambhampaty Vardhaman College of Engineering Hyderabad, India E-mail: [email protected] Rohith Gali Abstract— Environmental disaster is something that inflicts adverse effects on biodiversity, economy and human health. In order to reduce such adverse effects proper methodologies are to be brought into action to forecast them before in hand. This paper presents an alternative in predicting a tornado in advance by taking into account the results obtained from processing of satellite images contrary to the earlier methods where radar Images were analyzed. By using the satellite Images the noise levels can be reduced and higher accuracy of the results can be obtained compared to radar Images. Taking into consideration the observed values, the results obtained can be compared and a more accurate prediction can be made that would help us reduce huge losses in terms of both life and property. Till date none of the methods put forward were up to the mark and provided inaccurate predictions, and the drawbacks of which are eliminated in this paper providing with more accurate predictions. During the data mining process, obstacles like noise and attenuations deteriorate the quality of image hence denoising by k-mean clustering should be done after which coif wavelet transformations are applied on the image in order to obtain parameters such as square root balance sparsity norm threshold value which are used to find the wavelength ranges to formulate accurate predictions of tornadoes. The proposed technique capitulate an average accuracy of 92% in the prediction of a tornado. TABLE I. FUJITA SCALE F Scale F0. 40-72 mph F1. 73-112 mph F2. 113-157 mph F3. 158-207 mph F4. 208-260 mph F5. 261-318 mph Damage Light Moderate Considerable Severe Devastating Incredible EF Scale (U.S.) EF0. 65-85 mph EF1. 86-110 mph EF2. 111-135 mph EF3. 136-165 mph EF4. 166-200 mph EF5. Over 200 mph Each year, around 1000 tornadoes pass over the US, but only about 2% of them reach either F4 or F5 mark on the Fujita scale. Majority of the tornadoes are not sturdy enough, occurring in rural areas and doing little or no damage. Some tornadoes, however, are much disastrous, and can cut their way through major urban localities or through an entire small town. These usually strike without prior intimation and the devastation can be remembered for years. No accurate tornado prediction method has yet been invented in order to reduce the devastating effect in terms of both life and property. Previous approaches in the prediction of a tornado involved processing of radar based data in order to detect the occurrence of one. But this approach gave vague and inaccurate results since radar images consisted of high noise levels and attenuations which reduced the efficiency of the mechanism used over there. The methodology employed here involves clustering i.e, denoising and processing of satellite images and then finding the threshold and standard deviation values which are further used to locate the region of visible wavelengths in which tornado occurs. In future, using this mechanism, tornado can be predicted by matching the wavelengths of the then satellite images with the values calculated previously in this paper. The rest of the paper is organized as follows: Section II gives a glance to all the recent research carried on for the prediction of Tornado. Section III depicts the formation of tornado. Section IV describes experimentation methodology. Section V illustrates the experimental results and finally Section VI gives a conclusion. INTRODUCTION A tornado is an aggressively gyrating column of air that is in contact with both the surface of a cumulonimbus cloud and the earth or, in exceptional cases, the base of the cumulonimbus cloud. It is often referred to as a twister. The word tornado is derived from the Spanish word tronada, which means thunderstorm, which was taken from the Latin term tonare, meaning to thunder. The combination of tronada and tornar gave rise to the word tornado which means to turn. Generally, a tornado occurs when wind shear, which can produce a horizontal vorticity, integrates with thunderstorms. The horizontal vorticity is turned vertical by a thunderstorm and sets the storm rotating, transforming it into a supercell. A mesocyclone is the strong rotating updraft of storm. If the storm strengthens rapidly enough, a downdraft may distort around the bottom portion of the mesocyclone, tightening and escalating the rotation and bringing it in contact with the ground to form a tornado. The factors affecting the formation of a tornado include humidity, temperature, and unsteadiness in the wind flow, and the presence of a storm, all of which contribute to the 978-0-7695-5198-2/14 $31.00 © 2014 IEEE DOI 10.1109/SIMS.2014.29 Vardhaman College of Engineering Hyderabad, India E-mail: [email protected] formation of the thunderstorms needed to produce tornadoes. A final aspect is wind shear, which gives these storms the aptitude to produce tornadoes. The EF (Enhanced Fujita) Scale given in Table I is the criterion to measure tornadoes based on wind damage. Keywords - Tornado; Coif wavelet; Image processing; k-means clustering; Satellite imagery. I. Narasimha Prasad Vardhaman College of Engineering Hyderabad, India E-mail: [email protected] II. RECENT RESEARCH Research is the systematic observation of phenomena for the purpose of learning new facts or testing the 105 application of theories to known facts. The prime purposes of research are documentation, innovation, analysis, or the development and research of methods and systems for the progression of human intellect. Advent to research depends on philosophical theory of knowledge, which varies noticeably both within and between humanities and sciences. The main intention behind literature survey is to show what you have read and understood about the hold of other researchers who have studied the issue/problem that you are analyzing and include that in your thesis. It can be done by simple summarization, evaluating and complementing or various ways that show that one has done the research and which led to new and better conclusions. Thomas A. Jones et al. [1] determined what percentage of Doppler radar–detected vortices leads to tornadoes and what vortex features are most useful in differentiating vortices that are not and are coupled with a tornado. Caren Marzban et al. [2] analyzed data produced by the national severe storms laboratory’s tornado detection algorithm which was a result of several methods in order to conclude which storm-scale vortex features based on doppler radar is the best predictors of tornadoes. V Lakshmanan, Indra Adrianto et al. [3] used the least-squares method in order to estimate shear, morphological image processing to establish gradients, fuzzy logic to simulate compact measures of tornado possibility and a classification neural network to generate the final spatio-temporal probability field. C. A. Doswell [4] emphasizes on the super cell identification criteria and also the difference between tornado bearing cloud and non tornadic clouds. Marzban, Caren, Gregory J. Stumpf [5] used circulation database for training the neural network which was truthed in order to determine which circulation s were associated with reports of actual tornado events at the ground. According to them if a circulation is detected within 20 minutes prior to the ground report of severe weather or a tornado or 5 minutes after a report, that particular circulation was classified as a prediction of a tornado. Hongkai Wang et al. [6] proposed and evaluated a method of identification of tornadoes automatically in doppler radar images by detecting the hook echoes. To evaluate the hook echo detection algorithm, hook echoes were detected in several radar datasets by using the algorithm and compared with those proposed by the experts. In their experiment on radar images, they obtained an average probability of detection of tornado equal to 61.5%. Davies, J. M [7] used a database of sounding s derived from the rapid update cycle model to examine thermodynamic features of F1 and higher intensity tornado events associated with small storm relative helicity and/or high lifting condensation level heights. Dean, A.R., and R.S. Schneider [8] in their study examined the frequency of important severe weather conditions and the relationship of these conditions to the performance of convective watches issued by NOAA storm prediction center. They analysed parameters such as potential energy, wind shear, etc. This helped in improving forecast performance. Ostby, Frederick P [9] discusses and differentiates the methods that were used to predict tornadoes in the late 70's and 90's. Mitchell, E. De Wayne et al. [10] along with NSSL developed and tested a tornado detection algorithm that has been designed to recognize the locally intense vortices connected with tornadoes using the WSR-88d base velocity data. Colquhoun, J. R. [11] proposed a decision tree approach to forecast tornadoes which uses only meteorological parameters that are essential requirements for this phenomenon to develop. Hamill, Thomas M and Andrew T. Church [12] proposed a model that specifies the conditional probability that a major tornado will occur given that a thunderstorm occurs and given that certain RUC-2 CAPE and helicity values are forecasted. Trapp, Robert J. et al [13] collected a large dataset that were obtained using WSR88D mesocyclone detection algorithm to estimate the percentage of tornadic mesocyclones. Vasiloff, Steven V. [14] gave an alternate way where federal aviation administration’s terminal doppler weather radar can be used instead of the weather surveillance radar-1988 doppler (WSR-88D) for tornado detection in order to yield better results. Donaldson Ralph J and Paul R. Desrochers [15] showed that an improvement in the reliability and timeliness of tornado forecasting can be achieved through quantitative measurement by doppler radar of selected mesocyclone features such as excess rotational kinetic energy. Roger Edwards et al. [16] brought into light two loopholes in the present forecasting mechanisms they are 1) occasionally inaccurate synoptic-scale guidance by operational numerical weather prediction models, especially with regard to several critical parameters 2) model and observational difficulties regarding convective initiation and evolution. A. E. Mercer et al. [17] studied a sample of 50 tornado outbreaks and 50 primarily non tornadic outbreaks were simulated using the WRF to determine if they are able to distinguish outbreak type using synoptic-scale input. III. FORMATION OF TORNADO Tornadoes are usually the intense outcome of a very large thunderstorm called a supercell. During the storm warm air and cold air amalgamate. The warm air rises and the cold air drops. The warm air in due course twists into a spiral and forms a funnel cloud. Firstly, just before the thunderstorm builds up, a change in wind course and a rise in wind speed, at an increasing altitude, create an unseen horizontal gyrating effect in the lower atmosphere. Secondly, rising air within the thunderstorm’s updraft tilts the gyrating air to vertical from horizontal. Next, within a vast majority of the storm, an area of gyration, 2-6 miles wide is contained. The robust, most vicious tornadoes form within this area of gyration. Subsequently, a lower cloud base in the center of the storm becomes a gyrating wall cloud. This area is mostly rain-free. Finally, within a few minutes, a tornado builds up and starts to inflict its annihilation. In the initial stages, as the mesocyclone advances towards the ground, a noticeable condensation funnel materializes and descends from the bottom of the storm. As the funnel 106 descends, the rear flank downdraft (RFD) also comes in contact with the ground. This creates a gust of air front that can afflict damage, a reasonable distance from the tornado. The funnel cloud transforms into a tornado within minutes of the RFD approaching the ground. If the initial supercell thunderstorm is powerful enough it can give birth to numerous tornadoes. A new mesocyclone and a new tornado can be formed from the dissipation of one mesocyclone as shown in Figure 1. Later on, the tornado has an adequate supply of warm, moist in-flowing air to boost it, so it grows to maturity. This can take up to an hour. This is the most critical phase of the tornado and can be more than 1.6 km wide. The RFD turns into an area of cool surface winds and begins to envelop around the tornado, isolating the inflow of warm air, successfully choking the tornado. As the RFD cuts off the tornadoes air supply, the vortex begins to deteriorate. This dissolving stage only lasts a few minutes then the tornado weakens. The tornado is still potent of inflicting damage. The storm is shrinking, but the speed of winds can increase. information retrieved may go haywire from original value. Hence, the image must be denoised. Every image consists of various textures such as asphalt, barren lands, clouds, concrete, forests, grass and water bodies out of which not every texture is needed. These textures are to be segregated in order to get the desired texture for further analysis. Clustering is a task of grouping data into several segments whose members have a common resemblance. These segregations are based on wavelength as resemblance factor. The pixel values of various textures are used for this segregation which is shown in Table II. Figure 1. Formation of Tornado IV. EXPERMINTATION METHODOLOGY The primary purpose of this processing of satellite imagery is to identify the wavelength for prediction of tornado. The entire course of action employed is shown in the Figure 2. The pre-requirements for this methodology are satellite imagery of tornadoes with noise or without noise. When the satellite image containing noises such as illumination variations, occlusions, and scale variations, deformation of objects and so on, is analyzed, the Figure 2. Real Time Detection of Tornado 107 TABLE II. PIXEL VALUES OF VARIOUS TEXTURES Textures Water body Forest Grass Asphalt Barren lands Concrete Clouds Pixel values 0-20 21-33 34-81 82-140 141-199 200-224 225-255 Clusters are obtained based on the pixel values. The satellite image is thus segmented into various clusters and the region of interest which is cloud texture is extracted whose pixel value lies in the range of 225 to 255. K-means clustering is the mechanism deployed in this paper. It segregates the image such that the textures in a cluster have approximately same pixel values when compared with other clusters formed. In order to anticipate a fore-coming tornado a higher range k-means algorithm is adopted to cluster the image into six segments using Euclidean distance metric to avoid local minima. The clustering mechanism is applied for the image shown in Figure 3 and the resultant clusters produced are shown in Figure 4. MATLAB was used for the processing of images with image processing tool box, which provide image segmentation algorithms, tools and a complete data analysis, visualization and algorithm development platform. Wavelet transformation is applied on the cloud cluster obtained above. The coif wavelet transform is chosen because the Coiflets assure minimum signal disruption and provide a suitable technique. Coiflets are discrete wavelets designed by Ingrid Daubechies, to have scaling functions with vanishing moments. The Coifman wavelet (Coiflet) has enhanced steadiness and better regularity than the Daubechies wavelet, because it has more vanishing moments for the scaling function. The Coiflet has been established to be a popular and efficient basis in the signal and image processing. Further, the coif wavelet transformation converts the RGB image into gray scale image and eventually de noises the image. Figure 4. Images segmented as separate objects. (a) Cluster 1 (b) Cluster 2 (c) Cluster 3 (d) Cluster 4 (e) Cluster 5 (f) Cluster 6. To foretell the occurrence of a tornado an estimation of certain range of values from the clustered image should be made. Since the cloud image taken falls into visible infrared range, the wavelength will be in the visible spectrum range that would lie between 400 nm - 700 nm. The value of square root balance-sparsity norm threshold is calculated for those denoised, clustered cloud images. From the Table IV, it can be observed that, whenever a tornado occurs in the denoised, clustered image in frequency domain square root balance-sparsity norm threshold value ranges between 13 and 14. Soft threshold was chosen to find the value for square root balance-sparsity norm, because this procedure does not cause non-continuants at ck = ± . The notations used in this paper are given in Table III. TABLE III. SYMBOLS AND NOTATIONS USED FOR WAVELENGTH FACTOR ESTIMATION Symbol c s k cki 2 σ n ck k d d Figure 3. Original image 108 Notation Soft threshold value Wavelength of ith Pixel Square root balance-sparsity norm threshold Variance Standard deviation Number of pixels Mean wavelength Fixed threshold value Correction factor Wavelength factor TABLE IV. VALUES FOR VARIANCE, SQUARE ROOT BALANCE-SPARSITY NORM THRESHOLD, WAVELENGTH FACTOR AND THE WAVELENGTH Image No. Square root balance sparsity norm threshold ( ) Standard deviation (σ) Wavelength factor 13.29 13.99 13.80 13.94 13.85 13.93 13.63 13.89 13.05 13.99 13.02 13.87 13.03 13.07 13.97 2.84 2.75 2.68 2.89 2.70 2.78 2.54 2.74 2.75 2.68 2.76 2.74 2.72 2.67 2.60 43.29 44.66 45.82 42.54 45.55 44.23 48.29 44.87 44.69 45.91 44.53 44.85 45.23 46.08 47.25 Image 1 Image 2 Image 3 Image 4 Image 5 Image 6 Image 7 Image 8 Image 9 Image 10 Image 11 Image 12 Image 13 Image 14 Image 15 The values of variance, square root balance-sparsity norm threshold, wavelength factor and the wavelengths are calculated using Equations 1, 2, 3 & 4 are shown in Table 4. (1) For the determination of threshold values, equation (2) is used. 2 2 log n n (2) The equation (3) is used to calculate the variance. 2 ck ck 2 i n 1 (3) Fixed threshold value is calculated by using equation (4). 255 d 2 2 log n k dn d n 1 V. Wavelength (λ) = * d Observed result 575.35 624.79 632.37 593.06 630.90 616.07 658.16 623.26 583.23 642.27 579.78 622.14 589.40 602.28 660.07 Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado Tornado d mainly affect the formation of tornado are acknowledged and studied. Images contain primary lower attributes such as texture, pixel values and intensity. In this study, pixel values are considered as the main factor that is to be extracted from the satellite images. Using pixel values, wavelength is calculated and this wavelength is used for the detection of tornado. The actual analysis was carried out on a sample of 579 images and out of those 579 images the results of 50 images have been tabulated in table 5. Out of 50 images tabulated in Table IV and Table V, 38 of them were tornadic and rest of the 12 were non-tornadic images. Of these 38 tornadic images 35 were experimentally proved to be true and the remaining 3 images contradicted the observed results, and out of the 12 non-tornadic images analyzed, 11 turned out to be in accordance with the observed results and 1 image negated with the observed result. This analysis yileded an accuracy of 92.1% for the tornadic images and an accuracy of 91.6% for the nontornadic images. The mean accuracy for both these results turn out to be 92%. Since the above results and research is more accurate than the previous methodology, it is more convincing and satisfactory. Consider image 18 which is of a tornado. The results found experimentally shows a threshold value of 13.15 and wavelength equal to 631.76 nm. Since it is in the range of wavelength and threshold values established earlier, it is experimentally proved to be a tornado. Taking into account image 40 which is a non-tornadic image, the experimental results showed a threshold value of 13.18 and wavelength equal to 617.06, which falls under the established range turned out to be contradicting the observed result. The soft threshold expression is shown in equation (1). sign c k c k , c k c s k 0, c k d d (4) EXPERIMENTAL RESULTS Tornado is one of the biggest and most brutal form of a natural disaster, forcasting which can prevent lot of damage. Many researches have put forward many theories which have turned to be inaccurate due to the lack of proper analysis and study of various parameters that involve in the formation of a tornado. Mining of data from satellite images is even more challenging and cumbersome task since it involves processing of not one but hundreds of satellite images in order to extract the required results. The satellite images are obtained from Indian meteorological department and are used to analyze in order to identify the tornado. The extraction of factors which 109 TABLE V. IDENTIFICATION OF TORNADO Image No. Image 16 Image 17 Image 18 Image 19 Image 20 Image 21 Image 22 Image 23 Image 24 Image 25 Image 26 Image 27 Image 28 Image 29 Image 30 Image 31 Image 32 Image 33 Image 34 Image 35 Image 36 Image 37 Image 38 Image 39 Image 40 Image 41 Image 42 Image 43 Image 44 Image 45 Image 46 Image 47 Image 48 Image 49 Image 50 VI. Threshold Wavelength Experimental Observed ( ) (λ) results result 13.01 616.13 Tornado Tornado 13.06 595.80 Tornado Tornado 13.15 631.76 Tornado Tornado 13.99 672.64 Tornado Tornado 13.85 651.90 Tornado Tornado 13.73 671.95 Tornado Tornado 13.93 659.95 Tornado Tornado 13.91 645.10 Tornado Tornado 14.11 642.27 Non Tornado Tornado 13.45 604.40 Tornado Tornado 13.62 619.05 Tornado Tornado 13.9 660.57 Tornado Tornado 13.65 637.12 Tornado Tornado 13.89 636.49 Tornado Tornado 14.2 587.27 Non Tornado Tornado 13.88 618.53 Tornado Tornado 13.17 648.68 Tornado Tornado 12.56 607.21 Non Tornado Tornado 13.11 646.76 Tornado Tornado 13.61 632.85 Tornado Tornado 13.76 611.85 Tornado Tornado 13.79 673.02 Tornado Tornado 13.96 628.92 Tornado Tornado 16.51 663.19 Non Tornado Non Tornado 13.18 617.06 Tornado Non Tornado 12.06 646.26 Non Tornado Non Tornado 12.2 668.92 Non Tornado Non Tornado 12.27 652.69 Non Tornado Non Tornado 12.36 608.78 Non Tornado Non Tornado 12.65 637.35 Non Tornado Non Tornado 12.16 620.54 Non Tornado Non Tornado 12.68 647.08 Non Tornado Non Tornado 11.97 613.13 Non Tornado Non Tornado 12.53 611.28 Non Tornado Non Tornado 12.42 602.81 Non Tornado Non Tornado CONCULSION Identification TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE FALSE TRUE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE REFERENCES This paper proposes an alternate and more accurate method of predicting a tornado by choosing Square root balance- sparsity norm threshold and spectral characteristics of cloud in the visible and infra-red spectrum to differentiate between tornadic and non-tornadic clouds from the satellite images. The methodology employed here surpasses previous approaches where radar images were processed. Therefore more accurate prediction of tornadoes can be made by exploiting k-mean clustering algorithm and coif wavelet transformations. The risk of getting deviating and inaccurate results due to presence of noise in radar images can be eliminated by utilization of satellite images for forecasting of tornado. Extra parameters such as cloud dimensions, velocity, stratospheric temperatures and surface temperature that effect the precipitation of the cloud can be taken into considereration which will be stressed on in our future work. [1] Jones, T., K. McGrath, and J. Snow, ”Association between NSSL Mesocyclone Detection Algorithm Detected Vortices and Tornadoes,” Weather Forecasting, Vol.19, 2004, pp. 872–890. [2] Caren Marzban, E. Dewayne Mitchell and Gregory J. Stumpf, “The notion of best predictors: an application to tornado prediction,” weather forecasting, Vol.14, 1999, pp. 1007-1016. [3] V Lakshmanan, Indra Adrianto , Travis Smith and Gregory Stumpf, “A Spatiotemporal Approach to Tornado Prediction,” IEEE International Joint Conference, Vol.3, 2005, pp. 1642 - 1647. [4] C. A. Doswell III and Donald W. Burgess, “Tornadoes and Tornadic storms: A Review of Conceptual Models,” The Tornado: Its Structure, Dynamics, Prediction, and Hazards, Geophysics Monogr., No. 79, American Geophysics Union, 1993, pp. 161–172. [5] Marzban, Caren, Gregory J. Stumpf, “A Neural Network for Tornado Prediction Based on Doppler Radar-Derived Attributes,” Journal of Applied Meteorology, Vol.35, 1996, pp. 617–626. [6] Hongkai Wang , Mercer R.E, Barron J.L. and Joe Paul, ”SkeletonBased Tornado Hook Echo Detection,” International Conference on Image Processing, Vol. 6, 2007, pp. 361-364. 110 [13] Trapp Robert J, Gregory J. Stumpf and Kevin L. Manross, “A Reassessment of the Percentage of Tornadic Mesocyclones” Weather Forecasting, 2005, Vol.20, pp. 680–687. [7] Davies, J.M, “Tornadoes in Environments with Small Helicity and/or High LCL Heights,” Weather Forecasting, Vol.21, 2006, pp. 579–594. [8] Dean A.R and R.S. Schneider, “Forecast Challenges at the NWS Storm Prediction Center Relating to the Frequency of Favorable Severe Storm Environments,” 24th Conference Severe Local Storms, Savannah, 2008, 9A.2. [14] Vasiloff Steven V, “Improving Tornado Warnings with the Federal Aviation Administration's Terminal Doppler Weather,” Radar, Bulletin of the American Meteorological Society, 2001, Vol.82, pp. 861–874. [15] Donaldson, Ralph J and Paul R. Desrochers, “Improvement of Tornado Warnings by Doppler Radar Measurement of Mesocyclone Rotational Kinetic Energy,” Weather Forecasting, 1990, Vol.5, pp. 247–258. [9] Ostby, Frederick P, “Improved Accuracy in Severe Storm Forecasting by the Severe Local Storms Unit During the Last 25 Years: Then Versus Now,” Weather Forecasting, 1999, Vol.14, pp. 526–543. [10] Mitchell, E. De Wayne, Steven V. Vasiloff, Gregory J. Stumpf, Arthur Witt, Michael D. Eilts, J. T. Johnson and Kevin W. Thomas, “The National Severe Storms Laboratory Tornado Detection Algorithm,” Weather Forecasting, 1998, Vol.13, pp. 352–366. [16] Edwards Roger, Stephen F. Corfidi, Richard L. Thompson, Jeffry S. Evans, Jeffrey P. Craven, Jonathan P. Racy, Daniel W. McCarthy and Michael D. Vescio, ”Storm Prediction Center Forecasting Issues Related to the 3 May 1999 Tornado Outbreak,” Weather Forecasting, 2002, Vol.17, pp. 544–558. [11] Colquhoun J. R, “A Decision Tree Method of Forecasting Thunderstorms, Severe Thunderstorms and Tornadoes,” Weather Forecasting, 1987, Vol.2, pp. 337–345. [17] A.E. Mercer, C.A. Doswell III, M.B. Richman, and L.M. Leslie, “Evaluation of WRF Forecasts of Tornadic and Nontornadic Outbreaks When Initialized With Synoptic-Scale Input,” Monthly Weather Review, 2009, Vol.137, pp. 1250–1271. [12] Hamill Thomas. M and Andrew T. Church, “Conditional Probabilities of Significant Tornadoes from RUC-2 Forecasts,” Weather Forecasting, 2000, Vol.15, pp. 461–475. 111
© Copyright 2024 Paperzz