An Ensemble Approach to Transfer Learning for Classification of Habitat Mapping

Abstract

The Environment Agency- Abu Dhabi developed extensive habitat, land cover, land use maps in 2015 using a very high resolution satellite imagery acquired between 2011 and 2013. This map can be used as a baseline map to allow efficient monitoring. In this work, we aim to establish a framework for short term updates to the maps to quickly enable efficient planning. With the availability of multi-spectral images, various spectral bands apart from visible (Red, Green and Blue) bands can be used in habitat mapping. This paper presents the work of land cover classification in the region of Abu Dhabi, UAE using a Worldview-2 satellite image. The proposed approach makes use of Random Forest algorithm, applied on the Fully-Connected features obtained from AlexNet framework using a 20% training samples on a 3-band input. Then, ensemble of outputs of Random Forest over different 3-bands combination is used to make the final prediction. The results are validated against the ground truth obtained from Environment Agency, Abu Dhabi. Eventually, our aim is to develop a robust classification approach and then adapt automatic change detection approaches to temporally update the baseline maps.