View access options below. You previously purchased this article through ReadCube. Institutional Login. Log in to Wiley Online Library. Purchase Instant Access. View Preview. Learn more Check out. Abstract Remote sensing observations increasingly are used to obtain the detailed information needed about land surface state required for hydrological analyses. Related Information.
Close Figure Viewer. Browse All Figures Return to Figure. Previous Figure Next Figure. Email or Customer ID. Forgot password? Fig 4 shows elevation profiles for six of the most common devices used in OSFT activities. The Garmin Edge and Forerunner devices all rely on barometric altimeter sensors to determine elevation while the iPhone and Android applications do not. The Garmin Edge , , and Garmin Forerunner X devices all use barometric altimeter sensors to determine elevation [ 16 ], while the iPhone and Android devices rely on alternative methods [ 17 ].
Though newer iPhone and Android devices do contain barometric sensors, the Strava application does not request access to this information. In general, devices relying on barometric altimeter readings tend to be precise and similar in shape to the LiDAR-based elevation data.
- 50 Ways to Love Your LiDAR: How LiDAR is Used in Commercial Drone Applications - DRONELIFE!
- Victorias Demon Lover (Victorias Adventures in Time and Space Book 1)!
- The Prayer That God Answers: Experience the Power and Fullness of the Lords Prayer.
- Remote Sensing Center?
The overall accuracy of each segment effort is quite low however, with elevation fixes differing by up to 60 meters at the same point along a road. There are a number of factors at work here, primarily the small sample size of 30 devices for the Forerunner Xs over Segment A versus the for the Edge The non-barometric altimeter-enabled devices such as the iPhone and Android applications rely on a combination of location technologies, the primary method being global navigation satellite systems GNSS such as GPS-based elevation.
Notably, the mean and median elevation values, as well as the raw elevation values, shown in the plots for the iPhone and Android applications differ significantly from the interpolated NED profile shown in green. This implies that in this case either an alternate database is being used or that the elevation data from the device was not snapped to any database and instead GNSS elevation was reported unaltered.
Either way, the non-barometric altimeter devices differ substantially from the LiDAR elevation data in Santa Barbara and should not be relied on for accurate elevation values. These differences are discussed further in Section 3. Remember that Segment C contains an overpass so is not used in the accuracy training models. Segment D was included as it represents a different region.
To further refine the elevation model we split the segment efforts by device and year in order to determine if certain devices or software updates to devices produce higher accuracy mean and median elevation values. Any data prior to was excluded from our analysis as after splitting by device, there was not enough data to produce any meaningful results.
The results of this analysis found that there was little difference between years a proxy for device software updates and that segment efforts from should be included in the refined elevation model. In examining the accuracy of segment efforts from each of the barometric altimeter-enabled devices we found that Garmin Edge , and devices produced the most consistently accurate elevation profiles across the three training segments. This was likely due to these being the most popular devices used on the OSFT application and therefore had the least amount of variance across segments.
Other devices such as the Garmin Forerunner XT produced very accurate results for Segment B but very inaccurate results for the other two training segments. Again, the amount of data produced by each of these devices likely had the strongest impact on accuracy. It should also be noted that number of activity efforts does have an impact on the median and therefore the overall accuracy. Across all training segments, we found that randomly reducing the number of activity efforts to below began to have a significant negative impact on the overall accuracy and introduce high variance within each segment.
Combining segment efforts from the three top performing devices, namely the Garmin Edge , and , we report the accuracy for all training segments as well as Segment E , a segment that was not used in the training data. Table 4 lists the three measures of accuracy for these four road segments. The elevation profiles calculated via the median of our user-contributed, in-situ elevation data ISED segment efforts are compared with those of the National Elevation Dataset provided by the U.
Geological Survey [ 19 ]. Again, note that this NED is the source of elevation data that the Strava platform claims to use when a device does not have a barometric altimeter sensor. In the case of Segment D , the RMSE of the NED is relatively low over the entire segment indicating high accuracy overall, but the maximum offset HD is over double that of the user-contributed median elevation. The number of activity efforts that contribute to the mean also have an impact on the overall accuracy.
The median value for our in-situ elevation data ISED from user-contributed observations of each segment are compared to the interpolated National Elevation Data profile for the same segment. Up to this point the focus of this research has been on constructing accurate user-contributed elevation profiles through comparison to existing high-resolution LiDAR data. However, an important benefit of these user-contributed elevation profiles is that they can contribute elevation profiles to regions where LiDAR data is either not available, inaccurate, or not suitable for determining elevation of a road.
An example of the former is when part of a road segment has fully or partially closed canopy cover from vegetation suggesting that laser pulses are unable to breach the canopy and return true ground elevation values. An example of the latter is found when trying to construct an elevation profile for a road segment that passes below an overpass, Segment C for example.
At a distance of roughly meters we see errors in the Raw LiDAR elevation profile which can be attributed to dense canopy vegetation cover that results in shorter LiDAR pulse returns or scattering based on leaf angle. This was initially cleaned to provide an accuracy comparison for the in-situ, user-contributed elevation profile based on the method discussed in Section 3.
Since the ISED profiles are constructed from an aggregate of thousands of cycling activities and rely on barometric sensors, they are less prone to such canopy errors. Additionally, ISED profiles can be used to supplement standard elevation profiling approaches in cases where elevation can not be determined from an areal view-point. Segment C is a road segment which passes under a highway overpass.
You are here
As shown in Fig 5 , the LiDAR data black points for this segment correctly reports a number of sudden elevation changes shown between and m along the X-axis, the overpassing highway. The blue line represents the median elevation reported from a reduced set of barometric altimeter enabled cycling devices along this segment. This line depicts a profile of the segment traversing under the highway overpass, unencumbered by the highway overpass, one that is not possible to recreate from aerial LiDAR scans.
The ISED approach can be employed in countless other situations where elevation cannot be accurately determined from airborne sources or the resolution of these source is inadequate. Furthermore, user-contributed elevation profiles can be generated across any time span, given a reasonable amount of data. For example, an elevation profile can be constructed for a road segment before and after a tectonic event to identify any major changes in slope or elevation.
While this is possible with traditional elevation acquisition technology, repeated, high temporal resolution, data collection is often time consuming and cost prohibitive. ISED profiles offer an alternative and supplemental method to these traditional approaches. In this section we discuss how to make the created road network elevation data Web-available without having to change the OpenStreetMap data model.
Having demonstrated that high-resolution elevation data can be generated from in-situ observations from user-contributed OSFT data, we turn our focus to the process of publishing these data. The existing node and way structure of OSM uniquely identifies nodes along a street and the street segment, respectively nodes and ways are also used to represent other point and line features. We use the identifiers associated with these nodes and ways as objects on which to link our user-contributed elevation data. The latitude and longitude coordinate geometry representations of these nodes are compared against our ISED 1m resolution road segments and the elevation value for each OSM node is determined by taking the elevation value from the closest ISED segment node.
In most cases, OSM road segments are made up of far fewer nodes than the ISED 1m resolution road segments as their function is to trace the curvature and interconnectedness of roads, not to provide an even sampling of nodes. Furthermore, these OSM nodes are typically not evenly distributed across the segment. Over all five segments in our sample dataset, we calculate an average RMSE of 6. As reported by the EMD, the overall shape of the profiles remains similar, which is not surprising given that one is merely a reduced set of values from the other.
While the RMSE between profiles is already approx. Provided these findings, access to the higher-resolution dataset may be of interest to many domains and application areas. Rather than changing the OpenStreetMap data model by updating ways with nodes every 1m which would essentially break the OSM data model for many other purposes and dramatically increase the data size , we decided to generate a supplementary dataset from ISED containing the higher horizontal resolution elevation data. A way represents a street segment and contains a geometry attribute which in turn links to a positional sequence object which lists a sequence of node URIs Fig 6.
An example of these relationships are shown in Fig 7. Each node in our dataset consists of a required set of predicates as well as an optional sameAs predicate that links to a LinkedGeoData node. The required predicates are shown in Table 5. Related work in this area has focused primarily on generating new datasets from user-contributed and crowd sourced data or on the accuracy and precision of local and global elevation datasets.
To the best of our knowledge, little work has been done at the intersection of user-contributed elevation data from online social fitness tracking. A number of recent projects have explored the sensors on mobile devices from a user-generated data perspective to generate a range of interesting datasets and services as well as research findings [ 21 — 24 ]. Our previous work [ 25 ] has shown that sensors accessible on most current smart-phones can be employed to differentiate place types and could be used in contribution to volunteered geographic services [ 26 ].
Specific to the barometric sensor, existing work has identified this sensor in determining altitude estimations for indoor navigation [ 29 , 30 ], medical applications [ 31 , 32 ], and human movement and transportation research [ 33 , 34 ]. As data sources and platforms, social fitness tracking and activity applications such as Strava have been the focus of quite a few previous publications [ 35 ]. Griffen et al. Online social fitness tracking applications have gone on to sell a lot of the fitness activity data contributed by their users for various purposes such as urban design [ 7 , 38 ] and transportation infrastructure planning [ 6 ].
Systems have been designed for the purpose visualizing, sharing and analyzing much of this transportation data [ 39 ]. From an open geodata perspective, existing work has merged openstreetmap data with Shuttle Radar Topography Mission SRTM based digital elevation models approx. Recent work by Wang et al.
While their approach performs will when compared to benchmarks, there is no mention of the influence of canopy cover or other such obstacles. Furthermore, the frequency with which data is contributed to Strava means that our road elevation data can be updated daily or in some cases hourly and does not require an update from a third-party data provider.
Generating high-resolution elevation profiles is very often costly and time consuming. For a number of applications in many parts of the world, the spatial and temporal resolution of existing elevation data is not sufficient. The recent rise of online social fitness tracking applications has allowed individuals to publish local elevation data by way of barometric altimeters and GPS sensors in their mobile and wearable devices. Although each individual in-situ observation varies substantially in terms of spatial and temporal resolution and accuracy, the extensive amount of data from a variety of devices invites the construction of an aggregate, up-to-date elevation dataset for road networks at a 1m spatial resolution.
In this work, we have shown that elevation profiles generated from user-contributed data can approximate the accuracy of high resolution elevation profiles generated from ground-truthed LiDAR data. In fact, on average, our elevation profiles have a RMSE of 3. Furthermore, we demonstrated that user-contributed elevation profiles can be used to supplement existing elevation data sources in situations where they fall short, e. Lastly, we introduce a method to enhance OpenStreetMap, an existing open geographic dataset, through the addition of elevation values along road segments.
50 Ways to Love Your LiDAR: How LiDAR is Used in Commercial Drone Applications
Using the tenets of Linked Data, we present an approach to publishing our high-resolution, user-contributed elevation data and linking them back to existing spatial data sources without having to change the OSM data model. Future work in this area will involve expanding the scope of data sources from Strava cycling data to other platforms, e. Efforts are currently underway to expand the regional scope of this work outside of the two study areas presented.
One of the limitations of this work is that many of the regions that are in need of high resolution elevation models are places where uploading fitness tracking data is less common. We aim to explore the range of the various platforms and propose potential solutions for this limitation in future work. From a temporal perspective, next steps will focus on using in-situ elevation data to monitor changes in elevation and slope over time. Last, a RESTful application programming interface is in development that will return the elevation value of the closest known point provided geographic coordinates on the surface of the earth.
Browse Subject Areas? Click through the PLOS taxonomy to find articles in your field. Abstract Gaining access to inexpensive, high-resolution, up-to-date, three-dimensional road network data is a top priority beyond research, as such data would fuel applications in industry, governments, and the broader public alike.
Funding: The authors received no specific funding for this work. The concrete research questions addressed in this work are as follows : The quality of in-situ observations relies on the devices used and their sensors. Are the differences in vertical accuracy of devices that rely on barometric altimeters and those that do not, reflected in the data contributed to online social fitness tracking applications?
Previous work has shown there to be significant differences in accuracy depending on the sensor availability of cycling computers. In this work we show the degree to which these differences in accuracy have permeated into the social fitness tracking application Strava. Given the accuracy of certain types of sensors, what elevation accuracy can be expected from user-contributed cycling data? Through the removal of systematically erroneous data from devices lacking barometric altimeters, we show that it is possible to generate elevation profiles for road segments, accurate to within meters of ground-truth LiDAR data.
Furthermore, we demonstrate that these data can be used to supplement existing approaches to producing high-resolution elevation profiles. Can elevation data contributed by users of an online social fitness tracking application, be used to augment existing open geographic data platforms, e.
Finally, the elevation data constructed from OSFT users is of a high spatial resolution with elevation values every one meter along road segments. The inclusion of these data in OpenStreetMap is not directly possible as one would have to change the underlying OSM data model. Can high-resolution, user-contributed elevation dataset be constructed and linked back to OpenStreetMap data using Linked Data technologies [ 8 ]?
Thereby we make elevation data available and query-able without requiring any modifications to OpenStreetMap. Of these nine, four were chosen and evaluated for trail classification on the Swanton Pacific Ranch scene. Maximum likelihood classification as Parallelepiped classification uses a decision rule to classify data. The decision boundaries form an n-dimensional parallelepiped classification in the image data space defined based on the standard deviation threshold from the mean of each class ITT Visual Information Solutions, If the pixels in the scene fall within the threshold it belongs to that class, if it falls in multiple classes it is assigned to the last class matched, and if it does not fall within a threshold it is unclassified.
The minimum distance classifier uses mean vectors for each training set, and calculates the Euclidian distance from each unknown pixel to the mean vector for each class in the training set. All unknown pixels are then classified to the closest vector from the training set classes. The Mahalanobis distance classification uses statistics for each class, similar to maximum likelihood classification, but assumes class covariances are the same so it is a faster method.
Convolution and morphology filters were investigated, both in combination and individually. ENVI also provides sieve and clump techniques that can be applied to classified image products. These were evaluated as well. The output results from each of the filters were evaluated and will be discussed in Chapter V.
The convolution filters produce images in which the brightness value of a given pixel is a function of some weighted average of the brightness of the surrounding pixels. The extent of the surrounding pixels considered by the convolution function can vary in size, and is known as a kernel. Median and Laplacian convolution filters were used in this research. Median filters smooth an image, removing regions of noise from an image smaller than the size of a user specified kernel. The Laplacian filter is a second derivative edge enhancement filter that is not dependant on edge direction.
Morphological operations in ENVI are dilation, erosion, opening, and closing. Dilate fills holes smaller than the user selected kernel in images, where erode removes small islands of pixels that are smaller than the kernel. Opening erodes the image Opening is used to smooth contours, and remove islands and peaks in an image. Closing dilates an image followed by erosion, and is used to fuse narrow breaks and fill small holes.
These pixels were ground truthed and then compared to classification images and filtered images to determine the performance of each method. This evaluation was conducted in two tiers. The first tier evaluated the four different classification techniques used: maximum likelihood, minimum distance, parallelepiped, and Mahalanobis distance. After evaluation, maximum likelihood was chosen as the best technique, and post classification filtering was conducted on maximum likelihood classifications only. The second tier of evaluations compared the performance of different filters and combinations of filters to each other.
The equation is then: 2 0. The points were selected from a spatial subset of Swanton Pacific Ranch that is forested and has many roads and trails, as seen in Figure The sampling followed the color conventions from the classes identified in the training set, with green points corresponding to trail classified points and red, blue, and yellow corresponding to a not trail classification. Of the points generated, after ground truth, 61 points were classified as trail and were classified as not trail. The full list of points is in the Appendix, but the point breakdown summary was as follows: Table 3.
Performance was The thematic products from each of the processes in this research result in a classification into one of only two classes. With a discrete two class classifier and an instance known as ground truth , there are four possible outcomes Fawcett, If the instance is positive trail and it is classified as trail, it is counted as a true positive; if it is negative classified not trail , it is counted as a false negative.
If the instance is negative not trail and it is classified as negative it is counted as a true negative; if it is classified as positive it is counted as a false positive. Given these four categories a 2 x 2 confusion matrix can be constructed and forms the basis of many commonly used metrics as seen in Figure Generally, one point in ROC space is better than another if it plots in the upper left of the graph, that is, it has a high true positive rate and low false positive rate. Classifiers on the left side of the graph with low false positive Alternatively classifiers on the upper right side of the graph classify nearly all positives correctly, but have high false positive rates.
Higher true positive rates and lower false positives rates are better for each discrete classifier. Each of the four classification techniques rely on the statistics of the training set to make a determination as to which class a pixel belongs, as discussed in Chapter IV. The training set mean, band list with description, and standard deviations are provided in Tables 4 through 7. Table 4. Training set mean for topographic modeling bands 3 through 10 by class Topographic Modeling band list Band Name Description 1 Slope Convention of 0 degrees for a horizontal plane 2 Aspect The direction azimuth that a surface faces, in degrees clockwise from North 0 deg 3 Shaded Relief Renders terrain in 3D by use of shadows that would be cast by the sun from the NW 4 Profile Convexity Rate of change of the slope intersecting with the plane of the z-axis and aspect direction 5 Plan Convexity Rate of change of the aspect intersecting with the x,y plane 6 Longitude Convexity Measures of the surface curvature orthogonally in the down slope direction 7 Cross Section Convexity Measures of the surface curvature orthogonally in the across slope direction 8 Minimum Curvature Minimum overall surface curvature 9 Maximum Curvature Maximum overall surface curvature 10 RMS Error Indication of how well the quadratic surface fits the digital elevation model 11 Slope Percentage The percentage or degree change in elevation over distance It ranked three out of four for true positive rate with Mahalanobis distance confusion matrix with metrics and thematic map showing trails in green The maximum likelihood classifier had the best true positive rate at Maximum likelihood confusion matrix with metrics and thematic map Minimum distance confusion matrix with metrics and thematic map The parallelepiped classifier had the lowest true positive rate with Parallelepiped confusion matrix with metrics and thematic map The Mahalanobis distance classifier was ranked second over the minimum distance and parallelepiped classifiers due to its lower false positive rate.
For the remainder of the research, the maximum likelihood classifier results were used to evaluate different filtering methods. Table 8. Graph for classification techniques 0. Generally, three different approaches were applied to the maximum likelihood classified image.
The first approach was to apply a median filter to remove noise in the image followed opening, closing, erosion, and dilation operators of varying structural size and comparing the results. The second approach was to use a post classification tool in ENVI named sieve and clump. The function looks at local neighborhood for each pixel and determines if the pixel is grouped with pixels of the same class, if not it re-labels the pixel unclassified. The clump function groups together similarly classified areas after they have been sieved. The third approach used a decision tree to create a final product.
Using the decision tree the first node separates trail classified points from non trail classified points.
The second node of the decision tree takes the trail classified points, known as node one survivors, and applies a Laplacian edge detector. The output from this approach represents the intersection of the survivors from both nodes of the decision tree in a binary mask. The naming conventions used attempt to capture the image processing method used as well as the size of the structuring element.
The sieve and clump products follow a similar naming convention, and all decision tree products start with the survivormask and follow the naming conventions for any processing that was applied to that mask. A summary of the image processing techniques used and their respective metrics are listed in Table 9. The most accurate processes were the products from the decision tree. This process works the best since it moves from the more liberal classification process and step-by-step removes unwanted artifacts based on trail characteristics.
The maximum likelihood classification identifies a high percentage of trails in the scene with an The Laplacian convolution enhances edges of long linear features trails without regard to direction, and the decision tree survivors are those pixels that are both classified as trail and have edges.
Laser-Carrying Airplanes Uncover Massive, Sprawling Maya Cities | WIRED
This product is smoothed with a median filter to remove noise in the scene while preserving edges, and finally the closing fuses any narrow breaks and fills small holes in the trail segments. The thematic map from this process is shown in Figure Binary image showing trail classified points in black from image processing Figure Rule image for same spatial subset as Figure 34 prior to image processing SUMMARY The overall result of this experiment represented by the analysis in Chapter V is very encouraging that the process of statistical classification followed by image processing can correctly identify roads and trails under canopy using LiDAR.
The point densities of current LiDAR systems are capable of penetrating second growth forest canopy, and producing accurate digital terrain models DTMs. In this data set the slope values, convexities along different planes, and curvatures provided sufficient statistics to characterize unpaved road and trail segments within the study area.
Supervised classification techniques traditionally used in remote sensing on multispectral and hyperspectral images were applied successfully to identify roads and trails after training data was identified. Four classification techniques were used in the experiment to identify the best techniques for classifying roads and trails in forested areas. Among maximum likelihood, Mahalanobis distance, minimum distance, and parallelepiped classifiers maximum likelihood produced the highest true positive rate and was chosen as the sole classifier to conduct image processing.
However, the Mahalanobis distance classifier did result in the highest accuracy rating and may warrant more investigation in further work. Image processing techniques were applied to the result of the maximum likelihood classification and compared to each other.
- Revenge of the Barons Revenge.
- The Trusted Source for the Business of Drones..
- Navigation menu;
- 50 Ways to Love Your LiDAR: How LiDAR is Used in Commercial Drone Applications.
Three different approaches were used and evaluated against each other. The sieve and clump operators maintained the highest true positive rate. The results from the decision tree approach reduced the true positive rate, but also had the lowest false positive rate which led to it having the highest accuracy rate. In the evaluations of both the classification techniques and the image processing methods, the best techniques depend on the application. The Mahalanobis distance classifier and the decision tree processing techniques are more Either set of techniques results in a thematic map that successfully maps roads and trails under canopy.
Thematic map products from these processes could prove invaluable for both military and commercial applications. During intelligence preparation of the battlefield IPB , commanders and intelligence professionals can use this process to define the lines of communications LOC , which an enemy threat may use to move personnel and supplies through an area of operations.
In the field of LiDAR, there are many ongoing research and development programs in government and commercial sponsored studies. The accuracy of road and trail network mapping using LiDAR can be used for quantitative terrain analysis without the need for ground reconnaissance in areas unobservable to electro-optical imagery. LiDAR provides the ability to determine road networks on large scales in denied areas where ground survey is not available. This research demonstrates the value of LiDAR collected data in areas where traditional remote sensing techniques for intelligence preparation of the battlefield are insufficient.
Geophysical Research Letters, 32 21 , L21S Airborne laser scanning: basic relations and formulas. Processing of shuttle laser altimeter range and return pulse data in support of SLA International Archives of Photogrammetry and Remote Sensing, 32 3 , 64— Clode, S. Detection and vectorization of roads from lidar data.
Photogrammetric Engineering and Remote Sensing, 73 5 , — Crutchley, S. The light fantastic: Using airborne laser scanning in archeological survey. Swindon: English Heritage. David, N. Pathway detection and geometrical description from ALS data in forested mountaneous area. Paris, France. De Veaux, R. Intro stats 3rd ed. Diaz, J.
LiDAR: The Up-and-Coming Tool for Just About Everything
Identifying roads and trails hidden under canopy using Lidar. ArcMap Version Food and Agriculture Organization of the United Nations. Global forest resources assessment main report. Using a helicopter based laser altimetry system FLI-MAP to carry out effective dike maintenance and construction policy. Floods, from Defence to Management, — Airborne and spaceborne laser profilers and scanners. Topographic laser ranging and scanning. CRC Press. Laser scanning for the environmental sciences.
ENVI Help. Lefsky, M. Berlin: Springer. Richmond, R. Topographic laser ranging and scanning: Principles and processing. Sibson, R. Interpolating Multivariate Data pp. Szeliski, R. Computer vision: Algorithms and applications. New York: Springer. Vosselman, G. Airborne and terrestrial laser scanning. White, R. Remote Sensing, 2 4 , — Defense Technical Information Center Ft. Belvoir, Virginia 2. Richard C. Dan C.