Generative Street Addresses from Satellite Imagery

Abstract

Abstract:We describe our automatic generative algorithm to create street addresses from satellite images by learning and labeling roads, regions, and address cells. Currently, 75% of the world's roads lack adequate street addressing systems. Recent geocoding initiatives tend to convert pure latitude and longitude information into a memorable form for unknown areas. However, settlements are identified by streets, and such addressing schemes are not coherent with the road topology. Instead, we propose a generative address design that maps the globe in accordance with streets. Our algorithm starts with extracting roads from satellite imagery by utilizing deep learning. Then, it uniquely labels the regions, roads, and structures using some graph-and proximity-based algorithms. We also extend our addressing scheme to (i) cover inaccessible areas following similar design principles; (ii) be inclusive and flexible for changes on the ground; and (iii) lead as a pioneer for a unified street-based global geodatabase. We present our results on an example of a developed city and multiple undeveloped cities. We also compare productivity on the basis of current ad hoc and new complete addresses. We conclude by contrasting our generative addresses to current industrial and open solutions.

DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images

“…to extract realistic representations. Following the example-based generation idea, another approach is to use already existing data resources, such as aerial images [37,38], or geostationary satellite images [26,58]. Similar approaches extract road networks using neural networks for dynamic environments [53] from LiDAR data [59], using line integrals [32] and using image processing approaches [43,55].…”

“…Similar to the experiments of [26] and [37], we explored our baseline approach to follow some state-of-the-art deep learning models [19,24,28,44]. In contrast to those approaches, our dataset is more diverse, spanning three countries with significant changes in topology and climate; and significantly larger in area and size.…”

Detecting Hardly Visible Roads in Low-Resolution Satellite Time Series Data

“…Consequently, the segmentation task is to label the middle of the roads, allowing for identification and classification of roads at a finer level. We compare the performances of the following models: • U-Net: The baseline model, which resorts to a standard U-Net architecture that is commonly used in the context of road detection scenarios [14]. The model operates on data from a single timestamp of high quality (very few clouds in the test set; the tenth scene in the sequence of the twelve scenes, corresponding to June 2018).…”

Foreword to the Special Issue on Machine Learning for Geospatial Data Analysis

“…This Special Issue assembles six novel contributions in different areas of GeoData-driven machine learning. Topics span different disciplines of GIScience: generation of street address from satellite imagery [1], land-cover classification of polarimetric Synthetic Aperture Radar (PolSAR) images [2], extraction of buildings from maps to perform generalization [3], land-cover classification from satellite image time series [4], automatic selection of buildings based on cartographic constraints [5] and satellite image retrieval and recommendation [6].…”

“…In addition to tackling different applications, the papers employ a variety of machine learning tools to accomplish their goals. Deep learning is used in two works [1,4]: in the former, a convolutional neural network extracts roads from satellite images as an initial step to generate street addresses, while in the latter, a sequential convolutional recurrent neural network provides robust end-to-end land-cover and land-use mapping from satellite image time series. Decision trees in the form of single decision trees [3] or random forests [2] are used for building detection in the first and land-cover classification in the second contribution.…”