Abstract
Deep learning techniques are becoming increasingly important to solve anumber of image processing tasks. Among common algorithms, Convolutional NeuralNetworks and Recurrent Neural Networks based systems achieve state of the artresults on satellite and aerial imagery in many applications. While theseapproaches are subject to scientific interest, there is currently nooperational and generic implementation available at user-level for the remotesensing community. In this paper, we presents a framework enabling the use ofdeep learning techniques with remote sensing images and geospatial data. Oursolution takes roots in two extensively used open-source libraries, the remotesensing image processing library Orfeo ToolBox, and the high performancenumerical computation library TensorFlow. It can apply deep nets withoutrestriction on images size and is computationally efficient, regardlesshardware configuration.