Towards Generating the Habitat Maps for UAE using a Transfer Learning Approach

In UAE, the habitat land use land cover (LULC) map is currently available only for Abu Dhabi emirate. This work aims to develop a generic workflow for producing maps for the entire UAE and also the basis of updating the land cover classification maps for entire region of UAE very frequently. Experiments are presented using temporal images acquired from Worldview-2 for detailed habitat sub-type classification in a small region of Abu Dhabi and Sentinel-2 imagery for habitat type classification for extending the work towards the Northern Emirates. In the proposed approach, we first perform an automatic change detection using statistical methods such as principal component analysis. Then we make use of Random Forest algorithm applied on the Fully-Connected features obtained from AlexNet framework using the inputs from unchanged locations. The training labels are thus generated using the baseline classified maps of Abu Dhabi. Eventually, our aim is to develop a robust classification approach based on automatic change detection to update the baseline maps for the entire country in short time intervals. The ultimate objective is to conduct this experiment using high resolution WorldView2 satellite imagery but in this work, we apply our idea to Sentinel-2 imagery for higher level habitat type classification for the Northern Emirates. This paper presents the updates in the project which is expected to complete in 2020.