In November 2016, Australian UAV (AUAV) was tasked to undertake a 1200 hectare aerial mapping project in the Mid-Western Region of NSW, using our Unmanned Aerial Vehicles (UAVs, more commonly known as ‘drones’).

Although AUAV regularly undertakes large-scale aerial mapping of up to several thousand hectares, this project presented a unique challenge. It required both high-resolution colour and multi-spectral mapping (to record and analyse vegetation health), and for both to be delivered to survey-grade accuracy.

Project Purpose

AUAV was engaged by the Aquatic Ecology and Restoration Research Group at the University of New England to undertake the aerial mapping as one component of a comprehensive environmental study on soil hydrology.   

Challenge 1: Combined RGB and NDVI Imagery

RGB (Red, Green, Blue) is what we think of as normal colour imagery, like that taken by standard digital cameras. The results look similar to what you might be familiar with from Google Maps or other satellite mapping sources, but with drones we capture much higher resolution data, resolving detail down to the centimetre level.

NDVI (Normalised Difference Vegetation Index) imagery uses a multi-spectral camera to capture specific imaging bands in the near-infrared range. This data is then run through a special processing algorithm to create the NDVI false-colour imagery, being the standard for documenting and assessing crop and vegetation health.

One of the survey  areas showing a 3D Digital Elevation Model (DEM), RGB colour map, and NDVI vegetation map

Although the Sequoia multi-spectral camera can also simultaneously capture RGB imagery, it is not of sufficient quality to meet the requirements of high-resolution mapping or to generate an accurate 3D elevation model. This meant that we needed to capture the entire area with both a high-resolution RGB sensor and the NDVI sensor, effectively doubling the size of the 1200ha area to a 2400ha area, given it needed to be flown twice.

Challenge 2: Survey Grade vs Mapping Grade Accuracy

Ground Control Point (GCP) target

Most of our large-scale projects require only ‘mapping grade’ accuracy, i.e. around 1-3m, which is what a standard high-end GPS will give you. Typically, ‘survey grade’ (i.e. 5cm accuracy) is only needed for smaller sites up to a few hundred hectares, for example bulk earthworks or precision engineering projects.

Mapping accuracy is fine for most environmental monitoring work, where the outputs are used only for human interpretation and analysis, such as invasive weed identification.

Survey accuracy is needed, however, when 3D volume or flood modelling calculations are to be made from the elevation data, or if multiple surveys need to be precisely aligned over time for comparison.

In order to achieve survey-grade accuracy, control targets need to be placed on the ground throughout the area, such that they are visible in the aerial photography. These positions are then surveyed using very sophisticated equipment to around 2cm global accuracy. This greatly increases the workload involved, as instead of just flying over the area, we also need to move throughout the site to place and survey these marks in a grid pattern, with a marker every few hundred meters.

For a project like this, it means that the time taken with the ground control survey amounts to at least as much time as flying the aerial survey component, particularly as some of the terrain was difficult to access due to creeks, flooding, locked fences and 2m+ tall overgrown pasture.

Challenge 3: Time Constraint

Our final significant challenge was time: we were up against the clock on this project! We had initially estimated that a full week was required for the data capture. In the end though, scheduling issues around site access meant we had to complete detailed and comprehensive surveys in just four days.  

In order to achieve this, we added an additional crew member to the project, and also attempted to combine the RGB and NDVI sensors onto a single drone aircraft, carrying both cameras on board for simultaneous capture of the two data sources.

Also, rather than using traditional ground control survey, we accelerated this component using Propeller Aeropoints, which are automated ground control targets with integrated high-accuracy GPS units inside. Using these we could simply place them on the ground and press a button to activate them, rather than spending time surveying each target with traditional survey gear.

Some of these efforts paid off, and the job was completed on the accelerated time schedule, but only just, as it didn’t all go quite perfectly to plan!

Data Capture

We began the data capture using our newly-modified dual camera RGB+NDVI drone. Normally all new technology at AUAV is thoroughly tested prior to being used on a project, but in this case the time constraints meant that this modification could only be put together in the days immediately prior to the job.

We soon discovered that although the aircraft flew well and the RGB imagery was great, the NDVI camera was failing to trigger in this configuration. We spent some time trying to diagnose and rectify the problem, but soon made the call to fall back on our well-tested methods and technology. Therefore we proceeded to fly the site with two drones, using one of our custom-built aircraft for the RGB camera, and a senseFly eBee for the NDVI. The latter work well but only cover a small area per flight, and so this decision meant that not only were we flying two drones to get the job done, but the total number of flights required would triple. Our work was now very much cut out for us!

Thankfully the weather was favourable and the flying went well, with only the usual handful of technical glitches along the way, but nothing to significantly slow us down. And although we did see a few flying high overhead, there were no aggressive wedge-tail eagles attacking us, as they often do.

We soon established a good working pattern in which we would fly the two drones together at the same time (separated by 30m+ in altitude) which would result in our larger RGB mapping drone completing its work in half the time of the eBee. The eBee pilot would then continue flying while his colleague began to collect the previous ground control targets and leapfrog them ahead to the next area to be covered. By overlapping the tasks in this way, we minimised the amount of time taken by the requirement to fly each area twice.

The data capture took 4 very solid days, working as quickly as we could with only brief breaks. We arrived at first light each day and left with the sun on the horizon, so it was a very long 4 days for our hard-working crew.

Data Delivery

More than 200 gigabytes of image data was collected from the field, which then needed to be processed into the final 3D elevation and high-resolution orthographic map deliverables.

This amount of data would be challenging enough to process if it were all normal RGB colour imagery, but the multi-spectral NDVI imagery increased the difficulty in that it generates many small lower-resolution images rather than high-resolution images, meaning a lot more files for the software to correlate together, which then takes quite a bit longer to process. Additionally, we needed to process the RGB and NDVI imagery separately, in a way which ensured it all aligns perfectly together.

Top-down orthographic view of the high-resolution RGB and NDVI datasets in our online data portal, showing precise alignment of the two data layers (and use of the distance measurement tool)

The lower resolution of the NDVI camera also made the process of marking the ground control targets in the aerial imagery much more difficult. Below you can see how one of our ground targets typically looks in a high-resolution RGB image, and on the right a small cluster of pixels which needs to be interpreted as the target’s position in a low-resolution and colour-less image from the NDVI sensor:

A Ground Control Point (GCP), easily visible in the high-resolution RGB colour imagery, but difficult to accurately tag in the low-resolution, single-band multi-spectral imagery. Thankfully our software clearly reports if they are not correctly tagged.

Ordinarily we plan for a 1:1 relationship between time in the field gathering data, and time in the office processing it for final deliverables. In this instance though, the data processing took almost twice as long as normal due to the above complexities.

Lessons Learned

Although our dual-camera mapping drone didn’t quite work out for this project, the effort was not wasted, and we are already planning to use it on upcoming projects. The lesson here is of course one we already knew: where possible, always develop new technology well in advance of needing it, and test it thoroughly before using it in the field. In this case the time constraints didn’t allow for that to happen, but thankfully we still had the means to get the job done, albeit at a slower pace than we hoped, and requiring an additional crew member. Thankfully AUAV’s scale allows us to always approach projects with backup resources available, and although it is disappointing when we need them, it is always a relief to have multiple equipment options and staff on hand to guarantee a result.

Secondly, we continue to be reminded of optimism versus reality. During project planning we researched the area using Google Maps to assess how much ground we could realistically cover per day. This would have been quite achievable under normal rural property conditions, but in this case the inordinate number of fences and gates in the area as well as the flooding and many creek crossings required meant that our movements on the ground were severely hampered. The lesson here is to either be much more conservative in anticipating the time it will take, or if possible scout the location ahead of time, either in person or from the second-hand knowledge of locals or clients who have been to the site ahead of you.

Another lesson was regarding the Propeller Aeropoint ground targets. Although they are simple to place in the field, unless you’re careful, finding them again is the bigger challenge. At the end of the first day we had 3 which were incredibly difficult to find, despite us keeping accurate GPS records of where they were. We ended up locating them by using a drone to scout from the air. In rough terrain, you can be within a couple of meters and still not see them. The next day we adjusted our approach to flag their locations on nearby trees with bright pink survey flagging tape, and from then on they were easily located during the retrieval process.

Finally, again not a new lesson but one now ingrained to a deeper level: data processing of large mapping areas is *hard*. If it is too large to fit in the computer’s memory (i.e. larger than a few hundred hectares) then the data processing time goes up exponentially, and you need to break it down into chunks and ensure there is sufficient overlap between those chunks to guarantee alignment and accuracy of the results. We now work on a 1:1 factor of field work to office work for small projects, 1:2 for medium/large project like this one, and 1:3 for very large projects (3000+ha). It is important also that if a tight delivery timeline is required, the data processing will need to be distributed across multiple staff and multiple computers.

Many thanks to the University of New England for having us along on such an interesting project. Their work continues with ground sampling and a lot of careful data analysis, good luck guys! Also a huge thanks to Leon from AUAV, who once again went beyond the call of duty in flying up from Victoria at short notice to help get this project done.

26th April 2017

Andrew Chapman
Australian UAV, Director of Operations NSW
Mobile: 0488 130 900
Office: 1300 738 521


The final NDVI datasets, overlaid onto a satellite 3D terrain model of the wider area in AUAV’s secure online data portal