An ENSO-oriented mining algorithm for marine abnormal association patterns

Tuesday, 16 December 2014
Cunjin Xue and Qing DONG, RADI Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Key Laboratory of Digital Earth Science, Chinese Academy of Sciences, Beijing, China
Spatiotemporal variation of marine environmental parameters and their relationship with ENSO is a complicated system. The relationships among them are mutually responding and driving, and they have attracted much more attention in the context of global change. With great challenges of traditional spatiotemporal analysis to deal with their relationships, we propose a novel algorithm for exploring marine association patterns before / when / after ENSO occurrence using the remote sensing images at large scale. The main works are as follows.

Firstly, monthly anomalies are defined and used to represent marine abnormal changes, and their quantitative levels are calculated with a continuous interval to represent intensity of variations.

Secondly, this paper designs a recursive mining algorithm to find frequent items within the context of ENSO occurrence. The key implementations are as follows.

Step 1: Scan the database one time, for each item, i.e. marine parameter, and each variation type, i.e. -2, -1, 0, 1 and 2, calculate its probability, denoted as, and conditional probability when ENSO occurrence, denoted as, respectively, where, i is one of marine parameters, K is one of variation types of marine parameters, L is one of variation types of ENSO events. And find the frequent 1-items if and only if is not less than.

Step 2: Generate candidate 2-items according to the Apriori’s linking algorithm, and for each candidate 2-item scan the database, calculate its probability and conditional probability when ENSO occurrence, and generate frequent 2-items.

Step 3: Generate frequent (m+1)-items from m-items using a recursive algorithm with “Linking-Pruning-Generating”, where m is not less than 2. In Linking phase, Apriori’s linking algorithm is done to generate the candidate (m+1)-items, Pruning phase is to remove the (m+1)-items whose sub items are not frequent on the property of non- monotonicity, while Generation phase is to generate the frequent (m+1)-items by checking its conditional probability when ENSO occurrence is great than its probability in databases.

Finally, marine bio-optical and dynamical parameters over Pacific Ocean from remote sensing imagery with periods from January 1998 to December 2012 are selected to demonstrate the algorithm and discover their ENSO-related spatiotemporal association patterns.