Optimization of Parallel K-means for Java Paddy Mapping Using Time-series Satelite Imagery

Abstract

Spatiotemporal analysis of MODIS Vegetation Index Imagery widely used for vegetation seasonal mapping both on forest and agricultural site. In order to provide a long-terms of vegetation characteristic maps, a wide time-series images analysis is needed which require high-performance computer and also consumes a lot of energy resources. Meanwhile, for agriculture monitoring purpose in Indonesia, that analysis has to be employed gradually and endlessly to provide the latest condition of paddy field vegetation information. This research is aimed to develop a method to produce the optimized solution in classifying vegetation of paddy fields that diverse both spatial and temporal characteristics. The time-series EVI data from MODIS have been filtered using wavelet transform to reduce noise that caused by cloud. Sequential K-means and Parallel K-means unsupervised classification method were used in both CPU and GPU to find the efficient and the robust result. The developed method has been tested and implemented using the sample case of paddy fields in Java Island. The best system which can accommodate of the extend-ability, affordability, redundancy, energy-saving, maintainability indicators are ARM-based processor (Raspberry Pi), with the highest speed up of 8 and the efficiency of 60%.