DEEP LEARNING APPLIED TO REMOTE SENSING: AN APPROACH FOR THE DETECTION OF CATTLE DRINKING FOUNTAINS USING PLANET IMAGES

The recent increscent of orbital sensors is producing an unprecedented amount of global surface data in the history of mankind, which has a tremendous potential for application of deep learning (DL) approaches. In this direction, this work implemented a process of object detection (i.e. cattle drinking fountains) over Planet images, acquired in two distinct months (i.e. nov/2017 and mar/2018), and identified more than 24,000 drinking fountains in an area of 18,000 km2, located in the state of Goiás. This approach, based on the neural network architecture of U-Net and trained with data obtained by visual inspection, produced results that are spatially consistent and compatible with high resolution images. The results indicate a potential of application in other satellite data (e.g. Worldview, Sentinel) and in the identification of other objects (e.g. treetop, pivot, road).