SpaceNet: A Remote Sensing Dataset and Challenge Series
Foundational mapping remains a challenge in many parts of the world, particularly during dynamic scenarios such as natural disasters when timely updates are critical. Modifying maps is currently a highly manual process requiring a large number of human labelers to either create features or rigorously validate automated outputs. We propose that the frequent revisits of earth imaging satellite constellations may accelerate existing efforts to quickly revise foundational maps when combined with advanced machine learning techniques. Accordingly, the SpaceNet partners (CosmiQ Works, Radiant Solutions, and NVIDIA), released a large corpus of labeled satellite imagery on Amazon Web Services (AWS) called SpaceNet. The SpaceNet partners also launched a series of public prize competitions to encourage improvement of remote sensing machine learning algorithms. The first two of these competitions focused on automated building footprint extraction, and the most recent challenge focused on road network extraction. In this paper we discuss the SpaceNet imagery, labels, evaluation metrics, and prize challenge results to date.
SpaceNet: A Remote Sensing Dataset and Challenge Series
Adam Van Etten In-Q-Tel CosmiQ Works email@example.com Dave Lindenbaum In-Q-Tel CosmiQ Works firstname.lastname@example.org Todd M. Bacastow Radiant Solutions todd.bacastow @radiantsolutions.com
noticebox[b]Preprint. Work in progress.\end@float
The commercialization of the geospatial industry has led to an explosive amount of data being collected to characterize our changing planet. One area of particular interest is the application of computer vision and deep learning techniques to the extraction of information from satellite imagery at scale. To this end, CosmiQ Works, Radiant Solutions, and NVIDIA partnered to release SpaceNet data as a Public Dataset on AWS sn_s3 ().
Today, map features such as roads, building footprints, and points of interest are primarily created through manual mapping techniques. We believe that advancing automated feature extraction techniques will serve as important downstream use cases of map data, such as humanitarian and disaster response. Furthermore, we believe that solving this challenge is an important stepping stone to unleashing the power of advanced computer vision algorithms applied to a variety of remote sensing data applications in both the public and private sectors.
The first two SpaceNet challenges focused on building footprint extraction from satellite imagery. The third challenge addressed another foundational geospatial intelligence problem, road network extraction. The ability to create a road network is an important map feature (particularly if one is able to use this road network for routing purposes), while building footprint extraction serves as a useful proxy for population density zhangpop (), popnature (), popplos1 (). The potential application of automated building footprint and road network extraction ranges from the determination of optimal aid station locations during an epidemic in under mapped locations, to the establishment of an effective logistical schema in a disaster stricken region. There is a demonstrated need for both automated building footprint extraction as well as automated road network extraction, as the Humanitarian OpenStreetMap Team’s Tasking Manager hotosm () currently has scores of open tasks for both roads and buildings.
To address these challenges (and inspired by the ImageNet imagenet () model), the SpaceNet team released a large corpus of labeled satellite imagery in conjunction with public challenges aimed to increase the utility of satellite imagery. All SpaceNet data is distributed under a Creative Commons Attribution-ShareAlike 4.0 International License and is hosted as a public dataset on AWS and can be downloaded at no cost to end-users.
2 Source Imagery and Labels
2.1 Existing Datasets
Existing publicly available labeled overhead or satellite imagery datasets tend to be relatively small, or labeled with lower fidelity than desired for foundational mapping. For example, the ISPRS semantic labeling benchmark isprs_sem () dataset contains high quality 2D semantic labels over two cities in Germany and covers a compact area of 4.8 km; imagery is obtained via an aerial platform and is 3 or 4 channel and 5-10 cm in resolution. The TorontoCity Dataset torontocity () contains high resolution 5-10 cm aerial 4-channel imagery, and km of coverage; building and roads are labeled at high fidelity (among other items), but the data has yet to be publicly released. The Massachusetts Roads Dataset MnihThesis () contains 3-channel imagery at 1 meter resolution, and km of coverage; the imagery and labels are publicly available, though labels are scraped from OpenStreetMap and not independently collected or validated. Another useful overhead dataset (though not so relevant to foundational mapping) is the COWC dataset cowc () of cars, with 15 cm aerial imagery collected over six different geographic regions; labels consist of a point at the centroid of each car. A recently released satellite imagery dataset is xVIEW xview_data (), which contains 1400 km of 30 cm satellite imagery; labels, however, consist of bounding boxes that are not ideal for foundational mapping of buildings or roads.
2.2 Challenge 1 - Rio De Janeiro Building Footprints
The first SpaceNet challenge aimed to extract building footprints from the DigitalGlobe WorldView-2 satellite imagery at 50 cm resolution. Imagery consists of 8-band multispectral imagery at 1m resolution, as well as pan-sharpened red-green-blue (RGB) imagery at 50 cm resolution. Released imagery was created from a mosaic of multiple images, and covers 2544 square kilometers. The challenge imagery was split into 200 meter tiles using the SpaceNet utilities python package sn_utils (), with 60% of the data released for training, 20% for testing, and 20% reserved for validation.
For Challenge 1 in the Rio de Janeiro Area of Interest (AOI) over 300,000 building footprints were labeled. A GIS team at DigitalGlobe (now Radiant Solutions, a Maxar Technologies company) provided a polygon footprint for each building, which were initially extracted through semi-automated means and subsequently improved by hand. Any partially visible rooftops were approximated to represent the shape of the building. Adjoining buildings were marked individually as unique structures (i.e. each street address was marked as a unique building), see Figure 1.
|1||Rio||2,544||382,534||50||WorldView-2||2011 – 2014|
|2, 3||Las Vegas||216||151,367||30||WorldView-3||2015-10-22|
2.3 Challenge 2 - Las Vegas, Paris, Shanghai, Khartoum Building Footprints
The second SpaceNet challenge aimed to extract building footprints from the DigitalGlobe WorldView-3 satellite in a continuous image strip. The source imagery is distributed as a Level 2A standard product that has been radiometrically and sensor corrected, and normalized to topographic relief using a coarse digital elevation model (DEM). It contains the original panchromatic band, the 1.24m resolution 8-band multi-spectral 11-bit geotiff, and a 30 cm resolution Pan-Sharpened 3-band and 8-band 16-bit geotiff. Table 1 provides collection details for each SpaceNet Area of Interest (AOI).
The dataset includes four areas: Las Vegas, Paris, Shanghai, and Khartoum. The labeled dataset consists of 24,586 scenes of size containing 302,701 building footprints across all areas, and are both urban and suburban in nature. The dataset was split 60%/20%/20% for train/test/validation. Each area is covered by a single image strip, which ensures that sun, satellite, and atmospheric conditions are consistent across the entire scene. Each building footprint was extracted by hand, and the complete labeling guidelines for this dataset can be found in Appendix A.
2.4 Challenge 3 - Las Vegas, Paris, Shanghai, Khartoum Road Extraction
Challenge 3 used the imagery from Challenge 2, only tiled into 400m chips. As with the previous challenges the chipped dataset was randomly split 60%/20%/20% for train/test/validation.
CosmiQ Works worked with Radiant Solutions to label the road networks for the entire area of the existing SpaceNet Round 2 Building Dataset (see Table 1). Because the competition was designed to enable the creation of routable road networks, a labeling schema based on OpenStreetMap (OSM) guidelines was established to ensure ground truth that would be usable by open source routing tools. In addition to the digitizing of road- centerlines, four other attributes were recorded: 1) road type, 2) surface type, 3) bridge and 4) lane number. The complete label guidelines can be found in Appendix B. Table 2 provides a summary of road labels by road type and area of interest.
|Road Type||AOI 2 Vegas||AOI 3 Paris||AOI 4 Shanghai||AOI 5 Khartoum||Total|
|Motorway||115.0 km||9.0 km||102.1 km||13.4 km||239.5 km|
|Primary||364.7 km||14.3 km||191.5 km||98.4 km||668.9 km|
|Secondary||416.9 km||58.1 km||500.7 km||65.8 km||1041.5 km|
|Tertiary||2.9 km||10.5 km||33.6 km||67.9 km||115.0 km|
|Residential||1645.6 km||232.4 km||938.3 km||484.5 km||3301.3 km|
|Unclassified||1138. km||95.4 km||1750.8 km||164.5 km||3148.7 km|
|Total||3685.0 km||425.3 km||3536.9 km||1029.5 km||8676.6 km|
2.5 Additional Data
In addition to the data detailed above, SpaceNet also hosts data from previous competitions such as the IARPA Functional Map of the World Competition fmow2018 () and the Urban 3D Challenge Dataset Urban3D2017 (). These datasets diversify the SpaceNet data corpus, with the goal encouraging new avenues of machine learning research.
3 The SpaceNet Buildings Metric
The aim of the first two SpaceNet challenges was to extract building footprints from high-resolution satellite imagery, which is an object detection problem. Accordingly, we adopt a similar approach to ImageNet imagenet () and adopt a metric based upon the scale invariant intersection over union (IoU) metric, also known as the Jaccard index, which provides a measure of the overlap between two objects by dividing the area of the intersection by the area of the union.
As with Equation 5 of ImageNet, we select an IoU threshold of to denote a true positive detection. One additional feature that SpaceNet adopts from the ImageNet competition is the notion that each labeled region can have at most one true positive associated with that labeled region. This feature is implemented by a sequential search for a true positive sorted by decreasing IoU values. If a true positive is found, then the pair (the label and the proposed region) are removed from the sequence and the search continues.
With an object detection algorithm, the performance of an algorithm should depend on how many objects the algorithm detects (true positives) how many objects it fails to detect (false negatives), and how many non-objects it detects (false positives). Accordingly, for each scene in the SpaceNet data corpus, we compute the total number of true positives (proposals with IoU 0.5), false positives (proposals with IoU 0.5), and false negatives (ground truth buildings without a valid proposal). In a given city, we sum the true and false positives and negatives to compute total precision and recall scores; the city-wide F1 score is given by the harmonic mean of precision and recall:
The total SpaceNet Buildings Metric is tabulated via the arithmetic mean of the F1 scores of each city.
4 APLS: The SpaceNet Roads Metric
The third SpaceNet competition aimed to extract road networks from satellite imagery. Historically, pixel-based metrics have been used to assess the quality of road proposals, though such metrics are suboptimal for a number of reasons that we discuss below. To this end we developed a novel metric (Average Path Length Similarity: APLS) apls () to measure similarity between ground truth and proposal road graphs.
4.1 Pixel-Based Metrics
Road localization from overhead imagery has historically been treated as an image segmentation problem (i.e. identifying which pixels in an image belong to which class). In segmentation problems, each prediction pixel is either a true positive, false positive, or false negative. Three commonly used evaluation metrics are discussed below.
- Pixel-Based IoU
For object detection problems a threshold IoU value is often defined, above which the prediction is assumed to be a true positive. For pixel-based approaches, the number of true positive pixels forms the intersecting area, and the sum of true positive, false positive, and false negative pixels forms the union. IoU is subsequently calculated from Equation 1.
- Pixel-Based F1 Score
Each prediction pixel is either a true positive, false positive, true negative, or false negative. These can be combined into an F1 score via Equation 2.
- Relaxed F1
This is computed by first calculating the relaxed precision and relaxed recall. Relaxed precision is the fraction of predicted road pixels that are within a predetermined number of road pixels of a true road pixel relaxedf1 (). Relaxed recall is the fraction of true road pixels within Q pixels of a predicted road pixel. The F1 score is subsequently tabulated via Equation 2.
The three metrics listed above all fail to adequately incentivize the creation of connected road networks. For example, for the IoU and F1 metrics a slight error in the road width is heavily penalized, though a brief break in an inferred road (from a shadow or overhanging tree) is lightly penalized, as illustrated in Figure 3.
4.2 APLS Metric
Our focus on road networks naturally leads us to consider graph theory as a means to evaluate proposals. While many proposals exist for graph similarity matching, these tend to focus on the logical topology (connections between nodes) of the graph, whereas we care about both logical topology as well as the physical topology of the roads. Since we are primarily interested in routing, we propose a graph theoretic metric based upon Dijkstra’s shortest path algorithm Dijkstra1959 (). In essence, we sum the differences in optimal paths between ground truth and proposal graphs.
To measure the difference between ground truth and proposal graphs, the APLS metric sums the differences in optimal path lengths between nodes in the ground truth graph G and the proposal graph G’. In effect, this metric repeats the path difference calculation shown in Figure 4 for all paths in the graph. Missing paths in the graph are assigned the maximum proportional difference of 1.0. The APLS metric scales from 0 (poor) to 1 (perfect),
where number of unique paths, while length of path. The sum is taken over all possible source () and target () nodes in the ground truth graph. The node denotes the node in the proposal graph closest to the location of ground truth node . If path does not exist, the maximum contribution of 1.0 is used, thereby ensuring that missing routes are highly penalized.
Inherent to this metric is the notion of betweenness centrality, or the number of times a node appears along the shortest paths between other nodes. Missing nodes of high centrality will be penalized much more heavily by the APLS metric than missing nodes of low centrality. This feature is intentional, as heavily trafficked intersections are much more important than cul-de-sacs for routing purposes. Any proposed graph G’ with missing edges (e.g. if an overhanging tree is inferred to sever a road) will be heavily penalized by the APLS metric, so ensuring that roads are properly connected is crucial for a high score.
For small graphs the greedy approach of computing all possible paths in the network is entirely feasible (computing all possible paths for the 400m image chips of SpaceNet Challenge 3 takes less than 1 millisecond). For larger graphs, one must decide which nodes and paths are of paramount interest, lest the computational load become burdensome.
While we developed the APLS metric as a means to score computer vision competitions, unlike pixel-based metrics the APLS metric is agnostic to data type. The APLS metric applies equally well to road network proposals derived from optical imagery, radar, GPS tracks, LIDAR, and/or hand-crafted labels. Unfortunately, this metric is not differentiable.
4.2.1 Graph Augmentation
We select control nodes in the graph to denote points and routes of interest. In practice, control nodes are comprised of intersections, endpoints, and midpoints along edges. In order to capture the physical topology of roads, we inject midpoint nodes every 50 meters.
4.2.2 Node Snapping
Equation 3 is easy to apply if the proposal graph G’ has nodes coincident with the ground truth graph G. In practice, however, proposals will not align perfectly with ground truth graphs. Subsequently, we “snap” the control points of the ground truth graph onto the proposal graph. Proposal nodes must be within a given buffer distance of ground truth nodes. This buffer is 4 meters by default, which is equivalent to the labeling accuracy requirement for the road centerlines. Proposal nodes outside the buffer will not be snapped, and therefore highly penalized. This process creates a new proposal graph with nodes as coincident as possible to the nodes of the ground truth graph. Once the node snapping process is complete, routes between nodes in the ground truth graph (path) can be directly compared to routes in the proposal graph (path) using Equation 3. See Appendix C for more details.
4.2.3 Symmetric Comparisons
Once the ground truth to proposal node snapping procedure of Section 4.2.2 is completed, optimal path lengths can be compared between proposal and ground truth nodes. In order to penalize spurious road proposals, one must also perform the inverse operation by snapping proposal control nodes onto the ground truth graph. This is illustrated in Appendix C.
4.2.4 Total APLS Metric
APLS part 1 is computed by snapping ground truth control nodes onto the proposal graph and calculating Equation 3. APLS part 2 is computed via the reverse: snapping proposal nodes onto the ground truth graph and calculating Equation 3. The total ALS metric is the harmonic mean of APLS part 1 and APLS part 2, and scales from 0 (poor) to 1 (perfect).
5 SpaceNet Competition Results
In this section we cover highlights of the winning submissions, the full results can be found on the SpaceNet website sn_web (). All competitions were hosted on TopCoder.
5.1 SpaceNet Challenge 1 Results - Rio de Janeiro Building Footprints
The winning submission of the first competition utilized random forests to predict polygon candidates for buildings. The densely packed buildings of Rio de Janeiro proved difficult for all competitors, and the winning F1 score was 0.26, submitted by competitor wleite ().
5.2 SpaceNet Challenge 2 Results - Las Vegas, Paris, Shanghai, Khartoum Building Footprints
The second buildings challenge yielded far better results, with a winning score of 0.69, submitted by xdxd (). This is likely due to a combination of three factors: increased resolution (30 cm vs 50 cm for SpaceNet Round 1), improved building labels (due to greater experience by the labeling team, higher resolution images, and lower density in the SpaceNet 2 cities), and/or improved competitor algorithms. The top competitor used a deep learning segmentation mask combined with thresholding and filtering to yield building polygons. Las Vegas proved the easiest city to classify, with Khartoum proving the most challenging, see Appendix D.
5.3 SpaceNet Challenge 3 Results - Las Vegas, Paris, Shanghai, Khartoum Road Networks
Given the novelty of the metric used and the demand for geospatial vector submissions entries to the first roads challenge were surprisingly good, with a winning score of 0.66 submitted by competitor albu albu (). The top performers all took an approach similar to the baseline algorithm outlined in Appendix E, but with greater attention paid to post-processing, see Figure 5.
Foundational mapping has myriad applications, from commercial (e.g. autonomous vehicles), to humanitarian (e.g. disaster response), to civil (e.g. traffic congestion mitigation). Satellite imagery may be able to aid foundational mapping, particularly in cases of natural disasters or other dynamic events where the revisit rate of satellites may be able to provide updates far faster than terrestrial or airborne methods, while covering a large geographic area.
The SpaceNet dataset provides a large corpus of high resolution multi-band imagery, with attendant validated building footprint and road network labels. The SpaceNet building footprint extraction challenges yielded algorithms with F1 scores of on multiple cities, and vast improvement in performance from Challenge 1 to Challenge 2.
In this paper, we demonstrated that the standard technique for scoring road detection algorithms (pixel-based F1 score) is suboptimal for routing purposes. Accordingly, we developed the graph-theoretic APLS metric based on shortest paths between nodes in ground truth and proposal networks. This metric rewards correct road centerline and intersection determination to a far greater extent than pixel-based metrics. The SpaceNet competition yielded multiple high quality submissions in graph structure format, and a winning score of 0.66. These outputs can be used directly for routing, and provide a step towards addressing the many problems that improved automated routing impacts.
We will continue adding data and labels to the SpaceNet data corpus, as well as hosting challenges to encourage the development of algorithms tailored to the unique aspects of satellite imagery. Our aspiration is that the large amount of labeled SpaceNet data will stir the imagination of an increasing cadre of computer vision experts, and lead to improvements in remote sensing analytics.
We thank Ryan Lewis, Lisa Porter, and Jake Shermeyer for meaningful contributions and helpful comments. We also thank Todd Stavish and Patrick Hagerty for assistance in launching SpaceNet.
- (1) AWS: Spacenet on aws. https://registry.opendata.aws/spacenet/ (2017)
- (2) Zhang, F., Du, B., Zhang, L.: A multi-task convolutional neural network for mega-city analysis using very high resolution satellite imagery and geospatial data. CoRR abs/1702.07985 (2017)
- (3) Tatem, A.J.: Worldpop, open data for spatial demography. Scientific Data 4 (01 2017) 170004 EP –
- (4) Stevens, F.R., Gaughan, A.E., Linard, C., Tatem, A.J.: Disaggregating census data for population mapping using random forests with remotely-sensed and ancillary data. PLOS ONE 10(2) (02 2015) 1–22
- (5) HOTOSM: Hot tasking manager. https://tasks.hotosm.org/ (2018)
- (6) Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3) (2015) 211–252
- (7) ISPRS: 2d semantic labeling contest. http://www2.isprs.org/commissions/comm3/wg4/semantic-labeling.html (2018)
- (8) Wang, S., Bai, M., Máttyus, G., Chu, H., Luo, W., Yang, B., Liang, J., Cheverie, J., Fidler, S., Urtasun, R.: Torontocity: Seeing the world with a million eyes. CoRR abs/1612.00423 (2016)
- (9) Mnih, V.: Machine Learning for Aerial Image Labeling. PhD thesis, University of Toronto (2013)
- (10) Mundhenk, T.N., Konjevod, G., Sakla, W.A., Boakye, K.: A large contextual dataset for classification, detection and counting of cars with deep learning. CoRR abs/1609.04453 (2016)
- (11) Lam, D., Kuzma, R., McGee, K., Dooley, S., Laielli, M., Klaric, M., Bulatov, Y., McCord, B.: xview: Objects in context in overhead imagery. CoRR abs/1802.07856 (2018)
- (12) CosmiQWorks: Spacenet challenge utilites. https://github.com/SpaceNetChallenge/utilities/ (2018)
- (13) Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR. (2018)
- (14) Goldberg, H., Brown, M., Wang, S.: A benchmark for building footprint classification using orthorectified rgb imagery and digital surface models from commercial satellites. In: Proceedings of IEEE Applied Imagery Pattern Recognition Workshop 2017. (2017)
- (15) CosmiQWorks: Apls metric. https://github.com/CosmiQ/apls (2017)
- (16) Mnih, V., Hinton, G.E.: Learning to detect roads in high-resolution aerial images. In: Proceedings of the 11th European Conference on Computer Vision: Part VI. ECCV’10, Berlin, Heidelberg, Springer-Verlag (2010) 210–223
- (17) Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische Mathematik 1(1) (Dec 1959) 269–271
- (18) Boeing, G.: Osmnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks. Computers, Environment and Urban Systems 65 (2017) 126 – 139
- (19) CosmiQWorks: Spacenet on aws. https://spacenetchallenge.github.io/ (2018)
- (20) wleite: Spacenet round 1 winner: wleite’s implementation. https://github.com/SpaceNetChallenge/BuildingDetectors/tree/master/wleite (2017)
- (21) xd_xd: Spacenet round 2 winner: Xd_xd’s implementation. https://github.com/SpaceNetChallenge/BuildingDetectors_Round2/tree/master/1-XD%****␣spacenet_2018_arxiv.bbl␣Line␣125␣****_XD (2017)
- (22) albu: Spacenet round 3 winner: albu’s implementation. https://github.com/SpaceNetChallenge/RoadDetector/tree/master/albu-solution (2018)
- (23) CosmiQWorks: Broad area satellite imagery semantic segmentation. https://github.com/CosmiQ/basiss (2018)
- (24) Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. CoRR abs/1612.01105 (2016)
- (25) Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. CoRR abs/1505.04597 (2015)
- (26) scikit image: Skeletonize. https://scikit-image.org/docs/dev/auto_examples/edges/plot_skeleton.html (2018)
- (27) yxdragon: Skeleton network. https://github.com/yxdragon/sknw (2018)
Appendix A Appendix A. SpaceNet Building Footprint Labeling Guide
Buildings will primarily be within urban/suburban areas.
Rooflines will first be extracted to best represent the shape of the footprint and for any buildings where the base of the building is partially visible due to off-nadir angle the polygon will be shifted to best fit the base corners.
Buildings will have squared corners (assuming they are true to the shape of the building).
Extraction of building corners will be within 5 pixels of the location on the imagery.
Building footprints will be closed polygons and will not overlap.
No attributes will be populated.
Adjoining buildings with different heights shall be captured at one visible roof for the whole block and shifted to that visible roofs baseline.
Appendix B Appendix B. The SpaceNet Roads Dataset Labeling Guidelines
The SpaceNet Roads Dataset labeling guidelines
Road vectors must be drawn as a center line within 2m (7 pixels) of observed road.
The centerline of a road is defined as the centerline of the roadway. If a road has an even number of lanes, the centerline shall be drawn on the line separating lanes. If the road has an odd number of lanes then the centerline should be drawn down the center of the middle lane.
Divided highways should have two centerlines, a centerline for each direction of traffic. See below for the definition of a divided highway.
Road vectors must be represented as a connected network to support routing. Roads that intersect each other should share points as an intersection like instructed through OSM. Roads that cross each other that are not connected such as while using an overpass should not share a point of connection.
Roads must not bisect building footprints.
Sections of a road that are a bridge or overpass must be labeled as a bridge via a Boolean flag.
Divided highways must be represented as two lines with traffic direction indicated when possible.
Surface type must be classified as: paved, unpaved, or unknown.
Road will be identified by type: (Motorway, Primary, Secondary, Tertiary, Residential, Unclassified, Cart Track).
Number of lanes will be listed as number of lanes for each centerline as defined in rule 1. If road has two lanes in each direction, the number of lanes shall equal 4. If a road has 3 lanes in one direction and 2 directions in another the total number of lanes shall be 5.
Definition of Divided Highway
A divided highway is a road that has a median or barrier that physically prevents turns across traffic. A median can be:
A median is not:
Yellow hatched lines on pavement
Road Type Guidelines
All road types were defined using the Open Street Maps taxonomy for key=highway. The below descriptions were taken from the Open Street Maps tagging guidelines for highway and the East Africa Tagging Guidelines.
motorway - A restricted access major divided highway, normally with 2 or more running lanes plus emergency hard shoulder. Access onto a motorway comes exclusively through ramps (controlled access). Equivalent to the Freeway, Autobahn, etc.
primary - National roads connect the most important cities/towns in a country. In most countries, these roads will usually be tarmacked and show center markings (in Sudan, however, primary roads might also be unpaved).
secondary - Secondary roads are the second most important roads in a country’s transport system. They typically link medium-sized places. They may be paved but in in some countries they are not.
tertiary - Tertiary roads are busy through roads that link smaller towns and larger villages. More often than not, these roads will be unpaved. However, this tag should be used only on roads wide enough to allow two cars to pass safely.
residential - Roads which serve as an access to housing, without function of connecting settlements. Often lined with housing.
unclassified - The least important through roads in a country’s system - i.e. minor roads of a lower classification than tertiary, but which serve a purpose other than access to properties. Often link villages and hamlets. (The word “unclassified” is a historical artifact of the UK road system and does not mean that the classification is unknown; you can use highway=road for that.)
Cart track - This is a dirt path that shows vehicle traffic that is less defined than a residential.
Road Label GeoJSON Schema
“road_id”: int; Identifier Index
Not a bridge
Appendix C Appendix C. APLS Metric
APLS Graph Augmentation
A couple of complications arise when comparing road networks, both relating to the selection of which routes to compare. Accordingly, we select “control nodes” in the graph between which routes are of interest. In practice, control nodes are comprised of intersections, endpoints, and midpoints along edges. In order to capture the physical topology of roads, we inject nodes every 50 meters. Obviously, for very large graphs one would be far more parsimonious with the selection of control nodes, but for the 400m SpaceNet tiles we can afford to select a high density of control nodes. This process is shown in Figure 7.
Equation 3 is easy to apply if the proposal graph G’ has nodes coincident with the ground truth graph G. In practice, however, proposals will not align perfectly with ground truth graphs. Subsequently, we “snap” the control points of the ground truth graph onto the proposal graph. Proposal nodes must be within a given buffer distance of ground truth nodes. This buffer is 4 meters by default, which is equivalent to the labeling accuracy requirement. Proposal nodes outside the buffer will not be snapped, and therefore highly penalized. This process creates a new proposal graph with nodes as coincident as possible to the nodes of the ground truth graph. Once the node snapping process is complete, routes between nodes in the ground truth graph (path) can be directly compared to routes in the proposal graph (path using Equation 3.
Once the ground truth to proposal node snapping procedure of Figure 8 is completed, optimal path lengths can be compared between proposal and ground truth nodes. In order to penalize spurious road proposals, one must also perform the inverse operation by snapping proposal control nodes onto the ground truth graph. This is illustrated in Figure 9.
Appendix D Appendix D. Building Detection Results
Here we show a few snapshots of the winning algorithm for SpaceNet Challenge 2.
Appendix E Appendix E. Road Baseline Algorithm
The APLS metric is designed to reward correct node placement and connectivity, and so should prove a better metric than pixel-based F1 for automated route inference evaluation. In this section, we discuss the CosmiQ baseline procedures for exploring the application of APLS to satellite imagery basiss ().
While any number of approaches are valid (and encouraged) to tackle the SpaceNet challenge, the most obvious workflow begins with inferring a segmentation mask using convolutional neural networks (CNNs). We assume a road centerline halfwidth of 2m and create training masks using the raw imagery and SpaceNet geojson road labels (see Figure 11).