3-band Worldview-2 images are standard natural color images, which means they have three channels containing reflected light intensity in thin spectral bands around the red, green and blue light wavelengths (659, 546 and 478 nanometres (nm) respectively). Worldview-2 is sensitive to light in a wide range of wavelengths. Each image covers 200m 2 on the ground and has a pixel resolution of ~50cm. The first Area of Interest (AOI) released in the SpaceNet dataset contains two sets of over 7000 images by the DigitalGlobe Worldview-2 satellite over Rio de Janeiro, Brazil. We hope that this demonstration of automated building detection will inspire other novel applications of deep learning to the SpaceNet data. In this post we demonstrate how DIGITS can be used to train two different types of convolutional neural network for detecting buildings in the SpaceNet 3-band imagery. NVIDIA is proud to support SpaceNet by demonstrating an application of the SpaceNet data that is made possible using GPU-accelerated deep learning. State-of-the-art Artificial Intelligence tools like deep learning show promise for enabling automated extraction of this information with high accuracy. This information can be used in important applications like real-time mapping for humanitarian crisis response, infrastructure change detection for ensuring high accuracy in the maps used by self-driving cars or figuring out precisely where the world’s population lives. The SpaceNet release is unprecedented: it’s the first public dataset of multi-spectral satellite imagery at such high resolution (50 cm) with building annotations. This public dataset of high-resolution satellite imagery contains a wealth of geospatial information relevant to many downstream use cases such as infrastructure mapping, land usage classification and human geography estimation. So SpaceNet 7 predictions are actually superior to SpaceNets 4 and 6 when comparing comparable building pixel areas: a ~8 ⨉ 8 pixel square in SpaceNet 4 yields a recall of ~0.1, whereas in SpaceNet 7 the recall is ~0.55.įigure 5 plots pixel sizes directly, demonstrating the far superior pixel-wise performance of SpaceNet 7 predictions in the small-area regime (~5⨉ greater for 100 pix² objects), though SpaceNet 4 predictions have a far higher score ceiling.DigitalGlobe, CosmiQ Works and NVIDIA recently announced the launch of the SpaceNet online satellite imagery repository. Of course the pixel areas are different by a factor of 64 (4m / 0.5)², so a 120 m² SpaceNet 4 building is a ~20 ⨉ 20 pixel square, whereas an 1000 m² SpaceNet 7 building occupies only a ~8 ⨉ 8 pixel square. The building area histograms look similar in Figure 4 for SpaceNets 4 and 7, yet the performance curves are very different SpaceNet 4 performance asymptotes at ~120 m², whereas SpaceNet 7 asymptotes at ~1000 m² with much lower recall. Right: Winning SpaceNet 7 predictions from 4m optical data. Middle: Winning SpaceNet 6 (originally published here) from 0.5m synthetic aperture radar data. Left: Winning SpaceNet 4 (originally published here) predictions from 0.5 optical data, here we focus on the blue (nadir) line. Comparison of building prediction recall (blue) for SpaceNets 4, 6, 7, overlaid on building histograms (red), with (IoU ≥ 0.5). This metric was illustrated in one of our SpaceNet 4 analysis blogs, see Figure 1.įigure 4. Performonce vs IOUįor all five of the SpaceNet challenges focused on buildings (SpaceNets 3 and 5 explored road networks), we used an intersection over union (IoU) metric as the basis for SpaceNet scoring. A follow-up post will dive deeper into the temporal change and tracking lessons from this challenge. We compare results to past SpaceNet challenges and note that despite the challenges of identifying small buildings in moderate resolution (4m) imagery, the pixels of SpaceNet 7 seem to overachieve when compared to SpaceNets past. In this post we dive into some of the building-level metrics for the SpaceNet 7 Multi-temporal Urban Development Challenge. SpaceNet is run in collaboration by co-founder and managing partner CosmiQ Works, co-founder and co-chair Maxar Technologies, and our partners including Amazon Web Services (AWS), Capella Space, Topcoder, IEEE GRSS, the National Geospatial-Intelligence Agency and Planet. Preface: SpaceNet LLC is a nonprofit organization dedicated to accelerating open source, artificial intelligence applied research for geospatial applications, specifically foundational mapping (i.e., building footprint & road network detection).
0 Comments
Leave a Reply. |