A NOVEL APPROACH TO RECOGNIZING PATTERNS IN REMOTE SENSING TIME-SERIES USING DEEP LEARNING

One of the most remarkable breakthroughs of Remote Sensing lies upon the devise of CubeSat standard. Such technology open up a myriad of possible applications that benefit from the higher spatio-temporal resolutions delivered by constellations of CubeSat compliant nanosatellites. Within this scenario, one has to investigate the new challenges and how to tackle them in order to harness this new kind of Remote Sensing Big Data. Among these challenges is the development of the means to extract useful information of pixels' observations throughout time in a fine-grained fashion. This work is a seminal study on using a special kind of deep learning approach, namely, deep Recurrent Neural Networks, for classifying long time-series of landcover's observations. The method was tested against the problem of identifying pastureland areas over high-res imagery from PlaneScope, a constellation of CubeSat nanosatellites. A discussion concerning limitations and capabilities of the proposed approach are also presented.