Relsa: automatic analysis of spatial data sets using visual reasoning techniques with an application to weather data analysis

A nalyzing large d is tr ib u ted sp a tio tem p o ra l d a tase ts such as weather d a ta re­ quires th e c h a r t in g and a n im a tio n o f the d a t a using v isualiza tion techniques and the abs trac tion o f the visual in fo rm ation into qualita t ive descrip tions th a t can be more efficiently com m un ica ted and reasoned a b o u t . The qua l i ta t ive descrip tions are often expressed in the form of sp a tia l ly contiguous, tem porally evolving features, patterns, and the ir sp a t ia l relations. M achine ex trac tio n of these fea tu res and pa tte rns is a key step in a u to m a t in g the ana lysis of these datase ts . This thesis describes RelS.A.. a co m p u ta tio n a l framework for au to m a tic a l ly in terpre ting s p a t ia l da ta se ts in qua lita ­ tive term s. RelS.A comprises th ree m ain com ponents: a s t ru c tu re discovery algorithm , which increm enta lly ex trac ts s t ru c tu re d objects: a re la tion aggrega tion mechanism, which su p p o r ts the building o f spa tia l re la tions between agg rega te spatial objects from those o f the constituen t ob jects : a n d a co rre la tion-d irected feature in terpre ta­ tion m echanism , which is able to use higher-level constra in ts to refine the extraction of under-constra ined spa tia l ob jec ts and the ir spa tia l re la tions from noisy background. RelS.A a lg o ri th m s construct m ultilayer descrip tions of the d a ta , using both the ag­ gregated sp a tia l relations from co n s ti tu en t ob jects a t finer levels and the correlation constra in ts from higher levels to im prove efficiency and accuracy o f the in terpreta tion. This thesis dem onstra tes th e RelS.A techniques on an ap p l ica t io n in weather d a ta analysis: e x tra c t in g and labeling w eather features such as h ig h / lo w pressure cells. ii pressure troughs, th e rm a l packings, an d fronts from the raw w ea the r d a t a . T he bidirectional cons tra in t p ro p a g a t io n allows higher-level knowledge o f co rre la t io n to guide lower-level featu re d e te c t io n a n d proves to be essential for the difficult feature ex tra c tio n problem s w here low level feature d e tec to rs suffer from large a m o u n ts of noise an d d is trac ting fea tu res . T h e experim enta l resu lts on th e w ea the r d a t a have shown th a t the RelS.A. te ch n iq u es a re able to ex tra c t noisy and am b ig u o u s w eather features using higher-level c o n s tra in ts .