Introduction

How accurate is the elevation mapping service provided by Green Aero Tech?  While our primary hardware manufacturer senseFly (https://www.sensefly.com/applications/surveying.html) specifies “absolute X, Y, Z accuracy of down to 3 cm / 5 cm”, this has only been documented on mining survey environments.  So, to better appeal to our farming and agriculture customers, we set out to document and assess accuracy in a real-world farm situation.  The resulting case study shows and confirms just how powerful this UAV tool can be for agriculture producers.

First, we chose a field and environment that farmers would be familiar with.  The parcel of land was a quarter-section of bare farmland that was divided into two smaller fields, totaling about 120 acres of crop area.  Because of a light snow the night before, there was a thin covering on the field which provided an additional challenge, compared to an ideal condition of imaging bare dirt in direct sunlight.  The limited accumulation was not considered to be enough to impact the elevation values, so the operation moved forward.  By mid afternoon, the snow had disappeared entirely.

Ground Control Points (GCPs)

The goal for this case-study was to set a reasonable amount of Ground Control Points (or GCPs) at strategic points along the edge of the field (at least 3 corners) as well as a number of additional points on the interior of the field. In this manner, multiple reference points would be made available to “dial-in” the accuracy of the UAV imagery by giving a clearly-identifiable ground point (something that would contrast with the ground and be highly-visible from the air at a 90-100m altitude) with a known and fixed location (Longitude, Latitude, Elevation). On occasion, it would be possible to use large natural or fixed features of the land, such as culverts or boulders, if they are located in the subject area.

senseFly documentation recommends at least 3 GCPs for a subject area, to a typical maximum of 5 GCPs, however they do not specify an exact field area.  In discussion with other users experienced with this type of UAV data collection, we have heard recommendations of as many as 20 GCPs for a 160 acre quarter-section of land, depending on topography.  Future learning and analysis will be done to determine what the optimal number of points is for efficient mapping.

Using RTK survey-grade Leica equipment (base and rover), 10 GCPs were marked out on the field using a Jeep driving along the outside of the field and then additional points by driving through the field itself.  On the field, a large (approx. 75cm x 75cm) “X” was spray-painted on the ground at the precise point where the shot was taken.  This process took about 15 minutes for the field and the shot data would later be integrated to the Pix4D image processing software.  File format of raw shot data was output as XYZ metric points in NAD83, using UTM 15N for this particular region.

In traditional circumstances, an RTK-equipped vehicle (4×4 truck, SUV, or ATV) would be used to manually drive the field in a grid-pattern, however due to the early season the fields were much too wet to allow an accurate or efficient driving pattern.  In fact, even in the early morning with overnight freezing temperatures, the Jeep would still get mildly stuck in seemingly-good areas, making a full-field pattern virtually impossible.

 

Flight Specifications and Conditions

Weather: -4°C, approximately 90% cloud cover (overcast), 7.6m/s average wind speed from North direction

The UAV flight lasted approximately 28 minutes, starting at 11:15am and ending around 11:43am, covering a distance of 20.9km and averaging a speed of 46.7km/h.

Data result: 258 images captured at approximately 90m altitude resulting in 181.7 acres at 3.24cm/px average resolution.

 

Elevation Data Processing

After a quick in-field data processing to validate the captured images and ensure that GCPs were visible and that the snow-covered field provided sufficient contrast for the mapping software, the UAV and data were returned to the office for complete processing using our Core i7 4770K system, complete with 32GB RAM and 1TB SSD.  RTK shot data was now overlaid with the UAV data, and manually tied into several images showing the GCPs on the field, which took about 30 minutes for initialization and manual annotation.  The full processing took about 3 hours (automated) in Pix4D, and resulted in approximately 21GB of raw data along with various reports and formats.

To audit our results and determine what level of accuracy this platform is truly capable of, we purposely did not manually tie in 3 GCPs to the processing system, meaning that the software would ignore these 3 “hard” points.  After processing is completed, we are then able to come back and calculate the relative difference between what the software reports elevation to be (based on UAV imagery) and what the RTK equipment recorded.  By not “calibrating” in all 10 GCPs, we may potentially be sacrificing some absolute accuracy, however the trade-off of gaining provable audit “check points” should outweigh the minor loss.  As can be noted in the image, the chosen points ranged with both interior and edge GCPs.

 

GIS Analysis and Accuracy Assessment

Using our preferred GIS tool, QGIS v2.6.0 Desktop (http://www.qgis.org/en/site/), we loaded in the RTK shot data and labeled it overtop a Google Map Satellite background image of the area.  Then the RGB orthomosaic and generated Digital Surface Model (DSM) was loaded.

Using the QGIS “Identify” tool, it was a simple matter to scroll around the various layers to each individual GCP and obtain the DSM elevation value at the same location.  Basically, we zoomed in to a 1:200 zoom level and selected the GCP, and the Identify Results table displayed the GCP’s reported elevation in meters (field_4) as well as the underlying DSM attribute (Band 1) which was the elevation output from Pix4D in meters also.  The difference between these two values is considered to be the error margin and should be within 5cm’s to meet the senseFly accuracy claims.

 

Accuracy Results Table

As shown in the above table, the Elevation recorded by the RTK rover (in meters) when compared to the post-processed DSM Elevation value (also in meters) came in very close on the three “un-calibrated” audit GCPs (#02, #04, #08, highlighted in red), and the overall accuracy averaged less than 2cm difference between the two values on every point.  The equipment being used (Leica RTK base & rover with Differential GPS / L2 mode) technically only guarantees about a 2cm accuracy, so the UAV imagery when processed to DSM falls within even that margin of error.

 

 

 

Conclusions

In conclusion, the results manifest from this case study truly speak for themselves.  The senseFly eBee Ag in conjunction with accurately-collected RTK Ground Control Points consistently meets (and even exceeds) the manufacturer’s stated vertical accuracy benchmark of 5cm.  This is certainly also attributed to the immense processing power of Post-Flight Terra3D by Pix4D which easily performs the complex algorithm calculations and data handling in an efficient manner.

Even under less-than-ideal circumstances (overcast cloud, snow cover, indistinct ground features) the equipment and software performed admirably, and proved the technology under real-world circumstances that are likely to be encountered by agriculture producers looking to obtain elevation data about their land.  The limited field access required also greatly expanded the time window when these types of reports can be procured.  Reports of this nature can be gathered in early spring before seeding, reports can be analyzed and reviewed for permit requirements during the summer, and a drainage plan or similar operation can be started immediately after harvest in the fall.

Green Aero Tech intends to fully leverage this system of UAV imagery combined with RTK GCPs to make affordable elevation mapping services available to any level of farm operation or rural landowner.