Processing Pix4D Imagery with GCPs

Introduction:

The purpose of this activity was to compare the accuracy between processed UAS data imagery from Pix4D that utilized Ground Control Points (GCPs) contrasted to the imagery that didn't from the week previous.  A GCP is, according to Pix4D, "a characteristic point whose coordinates are known.  GCPs are used to georeference a project and reduce the noise." Pix4D recommends users to incorporate GCPs, as the angle differences in a given image set which have GCPs enable the images to adjust properly to project the data accurately in a 3D model.   

Methods: 

To start out, the data imagery gathered from the DJI Phantom 4 drone were uploaded to Pix4D.  When brought in, the shutter method defined for imagery capture was designated to "rolling shutter".  After the images were all imported, users were then asked to navigate to "GCP MTP Manager" import 16 individual GCPs which were collected at the Litchfield mine.  Once these were brought in, they were displayed as blue crosses atop the red circles which represent image locations displayed below in figure 1.

Figure 1

After the GCPs were imported, the Basic GCP/MTP Editor was used to mark the exact center of each GCP location in order to ensure the highest degree of location accuracy and representation.  Figure 2 below shows an example of a GCP being previewed for marking. 

Figure 2

This process was repeated for the next 15 GCPs.  Following that, point cloud, mesh, index, DSM and Orthomosaic are all ran.  Upon process completion, a quality report is generated, shown in figure 3 below.

Figure 3

After the processing was completed, the pinpoint accuracy of GCP marking can be shown in figure 4 below through the point cloud imagery.

Figure 4

Results

Once processed, the geotiff file created in Pix4D can be brought into ArcMap.  Figure 5 below shows a hillshade model of the GCP UAS Imagery.

Figure 5

It was then compared to the previous imagery that didn't utilize GCPs to assess overall accuracy (shown in figure 6 below).  The biggest distinguishing factor between the two seemed to be with the difference of elevation.  Notice how the max value reads 247 meters for the GCP map, and the non-GCP reads 108 meters.  

  

Processing Pix4D Imagery

Introduction:
Pix4D is a professional drone mapping and application software which uses images to create professional Orthomosaics, point clouds, and 3D mapping models.  In this exercise, students processed aerial images shot from a DJI Phantom 4 drone at 80 meters elevation over the Litchfield Mine located in Eau Claire, WI (taken on 9/30/16).  Pix4D was utilized in order to create a DSM and Orthomosaic model of the mine site. 

Methods:
Overlap required for this imagery required less overlap for areas of higher elevation rather than flat areas, which required more overlap. The recommended overlap is a minimum of 75% for frontal overlap, and at least 60% side overlap. If the user is flying over sand, snow, or uniform fields, there will be limited visual content mostly due to large uniform areas. Since the Litchfield mine is mostly uniform, a high overlap was selected (Figure 1).

Figure 1: Flight path, images captured, and photos processed in Pix4D

Rapid check is an alternative processing method which is speedier but less accurate.  In order to process multiple flights, the pilot needs to maintain height.  GCP's aren't essential for Pix4D, but they are strongly recommended if one wants to measure elevation points with high precision and accuracy.  

The first step involves processing the data with Pix4DMapper Pro application.  Students went to Create a New Project to carry this out, making a new folder location.  This was followed by taking the provided data from Professor Hupy and importing it to Pix4D into the created folder.  Figure 1 below shows the resulting pop-up screen once the destination folder is selected.  The Phantom 4 Camera is also switched from a global shutter to a rolling camera. To properly display the DSM, 3D Maps was also chosen.

Figure 2 - Image Properties of Litchfield Mine Survey

       
After the camera properties were set and the images to be processed were selected, the initial process was run before the point cloud and mesh and DSM, orthomosaic, and index were performed. This ensured the data would run correctly without having to wait for the processing time of the other functions. A quality report was then generated to preview outputs, and visualize the image overlay and images used along the flightpath (Figure 3).

Figure 3 - Previews of the orthomosaic and DSM outputs generated in the quality report

After this process was completed, the point cloud and mesh process and DSM, orthomosaic, and index process were performed. The results were then imported into ArcMap to be displayed as maps.

Results:

Figure 4 shows the orthomosaic output of the 197 photos used along the UAS flight path. This is a high resolution overlay because of the flight elevation and high-level photo overlay determined by the uniform texture of the Litchfield mine.

Figure 4 - Orthomosaic imagery from UAS of the Litchfield mine

Figure 5 then shows the digital surface model using a stretched color scheme to symbolize elevation. The elevation recorded by the UAS sensor uses an ellipsoidal earth method to record EXIF information rather than mean sea level used by most other software symbolizing elevation. 

Figure 5 - DSM elevation overlay of the imagery mosaic

Conclusion:
Point cloud imagery allows 3D modelling cost-effectively and is gaining in popularity across many sectors because UAS is highly customizable to achieve a wide variety of goals. For this project, the UAS point cloud data was used for volumetric analysis to calculate asset inventory. In order for the modelling to be accurate, attention to collection methods such as flight elevation and overlay are essential. This should be planned according to the objective before data collection. The methods and platform specifications should be recorded as metadata to assure accurate results. Flight elevation, shutter formats, platform used, date recorded, and other factors all play into the model produced and should be considered in project design and accounted for during processing.

Lab 10: Visualizing and Refining Terrain Survey

Introduction:

This lab is based off of a previous exercise performed earlier in the semester which involved measuring a 115 x 115cm sandbox to produce a hill, ridge, plain, depression and valley.  After the terrain was molded, students created a grid system utilizing pins and strings in efforts to normalize the data.  576 sample points of the elevation model were recorded in order for the data to be normalized.  This was essential to successfully project the data.  Data normalization, according to Esri, can be defined as "the process of organizing, and cleaning data to increase efficiency for data use and sharing."  The data points gathered from the sandbox were entered into a table in x,y,z-(elevation) format in order for it to be used in ArcMap.  These points were converted into a grid system in ArcMap in order to display the elevation of individual data points gathered within the sandbox. The data sampling method chosen for these particular points was a systematic sampling technique.  This method was found to be the most accurate and effective with gathering individual points on the grid. It has consistent intervals which are recorded at specific sampling points. The group which performed this sampling method created equal intersections with the string and utilized equal intervals of five centimeters along the x and y axis of the sandbox.  The interpolation procedure in this lab helps in visualizing this data by displaying 3D models of the sandbox.  These series of maps helped in representing the entire surface of the sandbox simply based off of the plots of each recorded point.  Figure 1 below displays a section of the x, y, z data used for the data points.

Figure 1: Segment of normalized data on Excel 

Methods:
Once the data was normalized and a geodatabase was created, the x,y,z data points were imported into ArcMap by navigating to File-->Add XY Data.  This was exported as a feature class within the geodatabase.  Since the points were established in relation to a specific reference point at (0,0), a cadastral coordinate system was utilized without projecting the data.  A grid was then established and ran through a series of different interpolation methods in order to determine advantages/disadvantages of each, as well as how realistic of a representation of the sandbox terrain each method produced.  These methods (defined by ESRI) included Inverse Distance Weighted (IDW), Kringing, Natural Neighbor, Spline and Triangular Irregular Network (TIN).

The IDW tool uses a method of interpolation which calculates cell value by taking the average values of data points in the vicinity of each processing cell.  The closer a point is to the center of a cell being estimated, the more influence it has in the averaging process.   

The IDW tool seemed to provide a basic picture of what the original sandbox looked like, but is lacking specific definition.  This is likely due to the fact that there are only 100 total input units available.

Natural Neighbor interpolation finds the closest subset of input samples to a query point and applies weights to them based on proportionate areas to interpolate a value.  Since it only uses neighboring points, it is better suited for compact datasets and terrain that has higher elevation variability.  

The Natural Neighbor seemed to created an elevation that was a little distorted with elevation variance, though it served well in revealing the peaks and valleys.

Kriging is an advanced geostatistical procedure that generates an estimated surface from a scattered set of points with z-values.  More so than other interpolation methods, a thorough investigation of the spatial behavior of the phenomenon represented by the z-values should be done before you select the best estimation method for generating the output surface.  

Besides a couple spikes in the surface model, the Kriging created a smooth and accurate representation of the original sandbox. 

Spline interpolation method estimates values using a mathematical function that minimizes overall surface curvature, resulting in a smooth surface that passes exactly through the input points.

The spline tool created the smoothest surface out of all the models.  During processing, the Regularized option was selected in order to create the most accurate replica of the recorded terrain.       
TIN is a vector data structure that partitions geographic space into contiguous, non-overlapping triangles.  The vertices of each triangle are sample data points with x-,y-, and z-values.  These sample points are connected by lines to form Delaunay triagles.  TINs are used to store and display surface models.                   

The tin model strongly illustrated the slopes in the terrain, but poorly reflected what the sandbox looked like in that it was jagged.

Once all 5 interpolation methods were completed, the resulting output rasters were imported into ArcScene in order to produce a floating 3D view of elevation change in the sandbox.  When the raster is initially brought in, it's projected as a flat surface.  It can be modified to 3D by selecting "floating under a custom surface" under layer properties. The 3D surface was then exported as a JPEG and brought into ArcMap to be used as a visual aid for the maps produced in the results section.  A scale bar was established by navigating to data frame properties and selecting "centimeters".

Results/Discussion
Figure 2 displays the surface of the original sandbox.

Figure 2

Figures 3-7 below show the resulting maps of each Interpolation method.

IDW
Figure 3 below shows a map utilizing the IDW interpolation method.  This map had a fair representation of surface elevation, but did not have a very smooth surface.  This was uncharacteristic of the actual terrain of the sandbox, which wasn't nearly as bumpy.

Figure 3

Natural Neighbor
Figure 4 below shows a map utilizing the Natural Neighbor interpolation method.  The peaks of each of the "hills" appear jagged, which is unrepresentative of the actual sandbox that was measured.

Figure 4

Kriging
Figure 5 below shows a map utilizing the Kriging interpolation method.  Elevation changes are not as strongly pronounced, but the overall surface is considerably smoother than Natural Neighbor and IDW, having less pronounced variability from point to point.

Figure 5

Spline
Figure 6 below shows a map utilizing the Spline interpolation method.  It is obvious that the surface of this model is considerably smoother than any of the other previous 3D representations.  This is likely due to the fact that Spline utilizes a mathematical function which minimizes overall surface curvature.

Figure 6

TIN
Figure 7 below shows a map utilizing the TIN interpolation method.  It is very geometric and pointy by nature due to the triangles generated, unlike the actual surface of the sand.  Despite that distortion, it still represents the sandbox elevation well.

Figure 7

For this particular survey, the Spline interpolation method appeared to be the best survey technique for producing the most accurate representation of the sandbox.  The mathematical function utilized to minimize surface curvature proved to be very effective.

Conclusion
This survey is related to other field surveys in that it collects elevation data over many points.  What makes it unique is the fact that such recordings were made only inches apart.  It is not always realistic to perform a highly detailed grid based survey, nor it it always necessary.  Difficult terrain or private property may be factors that interfere with this.  Interpolation be used for much more than just elevation models.  Factors like temperature, wind speed and windchill could also be collected for a interpolation dataset, being that they're multiple factors that are all inter-related.