Improvement the tropical cyclone forecasting process with effective features selection

The selection of adequate features for studying and forecasting of each atmospheric quantity regarding the meteorological ‎phenomena (for example, a tropical cyclone) is one of the basic challenges in system recognition and modeling concepts.
This study uses various methods of feature extraction including sequential forward selection, sequential backward ‎selection, mutual information, principal components analysis (PCA) and principal factor analysis (PFA) to extract quantities ‎ considering the wind speed at 10 meters above surface for a tropical cyclone activity period. This work studies 45 ‎various quantities of a Guno tropical cyclone activity domain which occurred in early June 2007 over the Arabian ‎Sea, the Gulf of Fars, and the OmanSea regions which caused significant damage to the region due to storm surges. These ‎quantities were initially selected from 286 quantities of the atmosphere which were determined in the global network of 1‎ by ‎1 ‎degree under the standard conditions including standard pressure levels (1000 mb to 10 mb), surface land ‎temperature (SLT), and surface sea temperature (SST). Due to the fact that the activity of the ‎interesting phenomena (i.e. Guno tropical storm) concentrates at the low and medial troposphere, it is possible to exclude ‎quantities of the upper atmosphere.
As a rule, the employment of neural network and fuzzy logic methods requires that the data be tested and validated. As a result, the Yemyin tropical ‎cyclone data was considered as testing data (it happened exactly 8 days after the Guno storm in the west of India, and, after ‎traveling over India, then entered into the southeast of Iran and, subsequently, into the northwest of Pakistan). The Nargis tropical cyclone data which occurred in May 2008 at the Bay of Bangal was considered as validation data because there ‎is currently no valid and acceptable data regarding the history of the old Guno tropical storm which occurred in this region approximately 35 ‎years ago.‎
The features extracted from the three first methods were tested using an adaptive network-based fuzzy inference system (ANFIS) approach. ‎Of course, 44 features were considered as inputs and a speed of 10 meters of surface was considered as output. The Guno storm data was considered for ‎training the network but the Yemyin storm data was used for testing the network and the Nargis storm data was used for the purpose of network validation.
Due to the fact that the KMO parameter for all three storms was greater than 0.80, the PCA method was used with high ‎confidence for data mining. The results show that the first 18 new components of the storms Guno and Yemyin are greater ‎than the others and they include more than 95% of existing information regarding 45 variables; consequently, the remaining components can be eliminated. However, this is true for the Nargis storm with the first 20 new components.‎
The results show that the computation time of the sequential backward selection method is very high in comparison with the sequential ‎forward selection method. The results show that the corresponding accuracy for sequential backward selection is also high.
The results of an ANFIS test for a different epoch show that the features extracted from the forward selection method have ‎minimum errors within three epochs.‎
The study of coefficient components shows the basic part of the second component is related to wind speeds of 10 meters height ‎over the surface and wind speeds in the layer 30 mb above the surface. Therefore, these components can be utilized as output for neural networks and ‎fuzzy systems such as the ones used in the ANFIS method.‎