Development of deep segmentation methods applied to satellite imagery

点击翻译按钮获取中文摘要

In collaboration with UrtheCast inc., the goal of this project is to explore the concrete usability of deepconvolutional neural networks in the field of satellite imagery. Since this project spans over a short period oftime (only 6 months) we will focus on image segmentation, a topic at the core of the research activities of pr.Jodoin and a glaring issue for UrtheCast. As such, we look forward to answer the following four questions:1) Which state-of-the-art network configuration works best and what is its accuracy compared to thenone-deep-learning solutions currently deployed by UrtheCast?2) How many manually annotated satellite images are required to properly train a convolutional neuralnetwork? Also, does weakly annotated data can be used to increase the accuracy of a network?3) How can very large satellite images (from 2,000 x 2,000 to 6,000 x 6,000 pixels) be processed (both attraining and at test time) on a single 12 Gb GPU?4) Can a model trained on a RGB dataset be transferred to a multispectral dataset as well as data from theUrtheCast satellites?Since deep segmentation models are new to the world of remote sensing, this project is fundamentallyimportant both for UrtheCast and the Canadian society as a whole. With its "UrtheDaily(TM)" project,UrtheCast is about to deploy a constellation of satellites which will acquire an unprecedented number ofimages that no one can manually analyse. This project will thus provide the company with the tools forprocessing those images. The same models could also be used to process data acquire by other Canadiansatellites such as RadarSat2 and SCISAT.