MOTIVATION This paper aims at presenting an automatic meteorological data mining approach based on analyzing and mining heterogeneous remote sensed image datasets, with which it is possible to forecast potential rainstorms in advance. Firstly, automatic MCS cloud detection and tracking methods are proposed to identify the geo-referenced cloud objects in satellite remote sensed images. Next, a data integration modeling mechanism is designed to extract meaningful properties of those detected

more »

... s, by integrating the heterogeneous image data and observed data into a unified view. Finally, based on the integrated global data schema, a two-phase data mining method employing machine learning techniques, the C4.5 decision tree algorithm and dependency network analysis, is proposed to analyze and forecast the meteorological activities of all clouds in order to discover the hidden correlations between the satellite observed data and the eastward evolvement trends of MCS clouds. Moreover, the meteorological environment factors that may cause the eastward evolvement trends of MCS clouds are also analyzed and conceptually modeled. An MCS Cloud Analysis System (MCAS) is also successfully designed and implemented by applying the above approaches, through which it is possible for meteorologists to forecast potential heavy rainstorms more easily. MCS IDENTIFICATION AND TRACKING Since the satellite image data are spatio-temporal, all MCS clouds need to be detected and tracked correctly and efficiently from complete image sequences. Subsequently, it is necessary to extract some representative attributes of the tracked MCS from the corresponding data collections in order to characterize them. To address the above problems, we propose a fast tracking and characterization method of multiple moving clouds from meteorological satellite images based on feature similarities. The method is based on the fact that in a relatively small time-span, the deformation of an MCS is progressive and detectable, which means that at two consecutive satellite images the same MCS will maintain a relatively similar moving speed, shape, area and texture. The complete flow of our method is described as follows. Firstly, image processing techniques are applied on the satellite remote sensed images to make MCS segmentation, by which all the MCS structures in the images are labeled. Then, a group of important features of each MCS are extracted as its representation, based on which feature similarities of different MCS in consecutive images are computed to track the same MCS along the time axis. Finally, in the characterization stage, the qualified MCSs are categorized into four types according to their evolvement trends indicated on the satellite images: (1) MCS moving East (E) out of the Tibetan Plateau; (2) MCS moving Northeast (NE) out of the Tibetan Plateau; (3) MCS moving Southeast (SE) out of the Tibetan Plateau; and (4) MCS staying in the Tibetan Plateau (STAY-IN). The categorization is implemented by computing the direction angle values from the movement trajectory of each MCS, and is crucial to meteorologists for predicting and evaluating the potential occurrences of heavy rainfalls. MCS DATA MINING Due to the heterogeneity of our data sources, appropriate attributes of each MCS, including TBB value, HLAFS values and feature values, should be firstly integrated to constitute a consistent input dataset serving for data mining purposes. We designed a data integrator to integrate the relevant data of TBB images and HLAFS variables into a unified dataset. Then, we implemented a spatial data mining approach directly from the integrated dataset to yield the meteorological knowledge.